Estimating Toll Road Demand and Revenue (2007)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

This chapter documents the state of the practice in travel demand forecasting for toll revenues. It begins by describing travel demand models (the basis of the toll road traffic fore- casts), how they have evolved generally and specifically for toll road forecasting, and how the models relate to revenue forecasts. The specific problem of the performance of these models in toll road applications is illustrated by a comparison of projected and actual revenues from several facilities and a discussion of the factors that influence performance. Current and emerging practices in the treatment of these factors are then described, based on the literature search and survey. TRAVEL DEMAND FORECASTING MODELS, APPLICATIONS, AND EVOLUTION This section briefly reviews the practice of travel demand forecasting models and how these models and their applica- tions have evolved over the last several decades. The purpose is to provide a context for the ensuing discussion, at a level of detail and at a perspective that are appropriate to the dis- cussion of toll road traffic forecasts. The discussion is not intended to replicate the many existing texts on forecasting [to which the reader can refer for further details—see, for example, Meyer and Miller (12)]. Overview of Travel Demand Forecasting Models The demand for travel is a derived demand. People travel (and goods are shipped) as a function of human activities. These activities are commonly represented in a travel demand fore- casting model as demographic, socioeconomic, and land-use variables (e.g., population, employment, and jobs). Travel demand also is shaped by, and shapes, the trans- portation network. The “supply” of transportation services— the different modes, their relative costs (time-wise and money-wise, temporally and financially), and the relative ease of accessing one location versus another—determine how the demand uses the transportation network. Similarly, forecasts of demand define the required supply of trans- portation services (how many lanes of road at what capacity, where bus routes are needed, etc.). Many medium- and most large-size urban areas in the United States, and around the world, use a travel demand fore- casting model, albeit with various approaches and to varying 10 degrees of detail and sophistication. In the United States, MPOs use models to develop long-range transportation plans. Consistent with this plan, a transportation improvement pro- gram must identify a list of projects proposed over a 20-year (or longer) period. The program also must identify priorities for the next 3 years (and must account for all federally funded projects over that time) and must be updated every 2 years. In addition, as a basis for improving urban air quality, federal regulations require that long-range transportation plans be consistent with air quality objectives and targets (13). The so-called “four-step” modeling process represents the most commonly used formulation for travel demand fore- casting models (12). The process has been used for several decades in the United States and around the world. Figure 1 presents a generic outline of the main inputs, processes, and outputs of this travel demand modeling paradigm. The indi- vidual elements are described here (14). Inputs • Zone definition. The urban area is divided into small spatial analytical areas, similar in concept to census tracts. Generally, traffic zones are defined by homoge- neous land uses (residential neighborhoods, central business districts, industrial areas, etc.), major “traffic generators” (universities, hospitals, shopping centers, airports, etc.), or geographic boundaries (rivers, rail- ways, etc.). • Land-use inputs. These are defined for each traffic zone in terms of population, employment, floor space, etc. • Transportation network. This normally includes the major road and highway network (typically, all roads except local streets), as well as the public transport net- work (bus routes, subways, light rail, commuter rail, etc.). These are defined in terms of a link-node network. The network can be refined to differentiate between HOV lanes, bus lanes, truck routes, routes with restricted access, etc. The network also can account for tolls. Traffic zones are represented as “centroids” (a sin- gle point each on the map), and are linked to the main network by means of “centroid connectors.” • Observed travel characteristics. This is measured typi- cally by origin–destination (travel characteristics) sur- veys. These provide a quantitative portrait of travel characteristics in a city, typically on a weekday. CHAPTER THREE TOLL ROAD FORECASTING: STATE OF THE PRACTICE

11 Traditional origin–destination surveys are “revealed preference” surveys—that is, they observe how people actually behave. However, these have proved limited as predictors of conditions that do not exist currently in a specific city: particularly the use of a new transit tech- nology (notably, rail) where none currently exists, and the willingness to use a tolled highway where tolls are currently not in place. “Stated preference” surveys attempt to quantify and predict such behavior. Counts of vehicles and their occupants by type of vehicle, at various points through the road and transit networks, are also important inputs. Other inputs include link/intersection (node) travel times and speeds. All of these observed conditions are used to calibrate the model. Some urban areas are beginning to conduct “activity”- based surveys, which provide a more precise depiction of travel characteristics within the context of a household’s daily activities. In comparison, the origin–destination survey focuses on these travel characteristics alone. Process The four steps of the process comprise: • Trip generation—where the total numbers of trips that start and end in each zone are calculated as a function of the different land uses in each zone. The calculations take into account different trip purposes, which again are represented by land uses (e.g., the daily home-to- work commute is commonly represented by population or dwelling units at the home end and by the number of jobs at the work end). • Trip distribution—where the generated trip ends are distributed among all zones. The distribution is con- ducted as a function of the zonal land uses (e.g., home- to-work trips would not be distributed to zones where there is no employment) and the characteristics of the transportation network (i.e., a function of the relative accessibility of a zone, which is measured as a function of travel time–congestion and cost–transit fares, park- ing charges, road tolls, etc.). Different calculations are made for different trip purposes to take into account their different behaviors. The products of this step are expressed as matrices of trips for different purposes (e.g., stating that there are 100 trips for purpose “x” from zone i to zone j). • Modal split—where the distributed trips are allocated to the different available travel modes. Typically, the allocation is between automobiles and public trans- port; however, some models further differentiate among public transport modes (including park and ride), between HOV (i.e., automobiles in which there are two or more occupants) and SOV (i.e., automo- biles in which the only occupant is the driver), and Inputs: Zone Definition Inputs: Land Use / Socioeconomic Variables Inputs: Road and Transit Networks Inputs: Origin - Destination Travel Survey, Counts, etc. Trip Generation Trip Distribution Modal Split Trip Assignment Outputs: vehicles / hour and transit passengers / hour on networks; link speeds and travel times Simulated FIGURE 1 Outline of travel demand modeling process (traditional “Four Step” paradigm).

nonmotorized modes (pedestrians and bicyclists). Dif- ferent calculations may be made for different trip pur- poses, again to account for their different behavior; however, these are subsequently combined by mode for the next step. A common formulation is the logit function, which simulates the traveler’s utility according to out-of- pocket cost, door-to-door travel time, and other attri- butes of modal choice (such as trip distance, proximity of the transit stop to the workplace, in-vehicle comfort, the number of transfers required, and so on). Other, simpler formulations include diversion curves or fac- tors. Different formulations may be applied to different trip purposes. Typically, the resultant matrices for a given mode are combined for all purposes, resulting in a matrix of all automobile driver trips, all transit pas- senger trips, etc. The survey of practitioners indicated that the inclusion of modes varied. All respondents (i.e., those that had completed all three parts of the survey) indicated that passenger vehicles were modeled; how- ever, not all differentiated between SOV and HOV modes of travel. Most of these respondents included trucks and commercial vehicles as mode choices, and about half of these included transit in the model. The methods for mode choice modeling included logit (or similar) and other factors. Some of the respondents using assignment-only models calculated modal choice exogenously. • Trip assignment—where the trips for each mode are loaded onto, or assigned to, the respective transporta- tion network(s). This is a translation of demand, which is expressed as the number of trips by mode x (for all purposes combined) between zone i and zone j, into automobile traffic volumes on a given road link and rid- ership on a bus route, etc. Of interest to this synthesis is the treatment of auto- mobile driver trips [which are equivalent to automo- bile vehicle trips (i.e., there is only one driver per vehicle)]. There are several algorithms for assigning automobile vehicle trips. The “equilibrium assign- ment” is a common technique. This process allocates traffic to links so as to minimize the cost of the auto- mobile or transit traveler between his or her origin and destination; where “cost” is commonly defined as travel time (minutes) and, in some models, with a monetary cost expressed in terms of time (i.e., value of time). The latter allows for the impact of tolls or other pricing mechanisms on the driver’s choice of route. Equilibrium is achieved when, between the current and previous iterations, no driver (i.e., vehicle-trip) can improve his or her travel time by switching routes. In contrast, the so-called “all-or- nothing” assignment algorithm does not account for the build-up of volume or cost, and assumes that all link speeds (typically, the posted or free-flow speed) remain fixed without regard to the actual volume on each link. 12 Outputs Volumes by link and ridership numbers are the main outputs of the model, along with travel times and speeds across the transportation network by link. These outputs can be used in turn to identify costs, fuel consumption, and air pollutants, as well as revenues on a tolled facility. Comments on Modeling Process It is important to note that the aforementioned modeling process is not prescribed; that is, there is no one single or standard modeling process or universal method. Keeping in mind the different perspectives and uses of toll road demand forecasting, some comments are in order. • Commonly, MPOs and other transportation planning authorities focus on simulating peak-hour travel on the transportation network—that is, the time of day at which the transportation system carries its maximum volume. Normally, this occurs during the morning or afternoon commuter peak periods. However, the model may simulate different time periods, ranging from 24-h travel on a typical weekday to a peak period or a peak hour within that period. Factors may be used to derive peak-hour matrices from 24-h or peak-period matrices; however, in the absence of these factors or of direct modeling of the peak hour or period, the use of sophis- ticated algorithms (such as the equilibrium assignment technique) is effectively precluded. In contrast, toll road revenue forecasts typically require annual estimates of demand. Therefore, fore- casts from the aforementioned peak-hour models must be extrapolated. This requires the development of fac- tors for different time periods, which can include the peak period, daily, weekly, monthly/seasonally, and ultimately, annually. Factors may be developed according to observations of traffic volumes or trends (e.g., 24-h traffic counts, by hour) or other sources. However, the use of factors maintains the status quo, does not account for temporal changes (peak spread- ing) or mixes in the traffic composition, and may require special additional factors or assumptions to account for travel on weekends or holidays. Some models have addressed this by simulating several time “slices” during the day (e.g., the a.m. peak hour, a mid- day hour, and the p.m. peak hour). • Some models emphasize certain steps more than oth- ers, or the models may not include some steps or have combined others, or some parts of the process may be modeled exogenously. For example, the Quèbec (Canada) Ministry of Transportation’s urban models for Montrèal and other cities focus on the trip assign- ment step, using trip matrices that are derived directly from comprehensive high-sample, origin–destination surveys. Modal shares and demand forecasts are devel- oped exogenously to the assignment model, although

13 trip generation and trip distribution typically are not modeled. The point is that the treatment of a (nomi- nally) common modeling process varies among trans- portation planning authorities, which differences in turn necessarily are carried through in the treatment of toll demand forecasting. As a result, the comparability of toll demand forecasts and their performance may be limited. • Truck and commercial traffic is generally considered to be an important market segment for many toll roads. However, relatively few urban models simulate these trips explicitly. Some urban areas have developed truck models, which may or may not be integrated within the primary urban passenger travel model. A common treatment for including truck or commercial traffic is to factor the resultant automobile forecasts on each link according to the observed proportion of trucks or com- mercial vehicles in the observed traffic mix (according to traffic counts). Although this provides a simple tech- nique for capturing the “full” mix of traffic on a partic- ular facility, on its own it provides no way to account for tolling, other changes to the transportation system, or changes in demand. Moreover, the truck peak hours in many urban areas do not coincide with that of the dominant automobile peak hour. Evolution of Models The modeling process has evolved since the development of the four-step modeling process during the 1950s and 1960s. At that time, models were applied primarily to the planning of major transportation facilities (mainly highways) to accommodate rapid post-war urban growth. The four-step process is still the dominant formulation in urban travel demand models (12, p. 289). However, several concerns have encouraged the development of new techniques. • The ability of the process to address current planning needs, which have evolved from the planning of new highways to meet forecasted demand to better manag- ing that demand (e.g., through other modes as well as traffic management). • Inconsistencies among the four steps have been identi- fied with respect to their formulation, parameter values, costs, and variables (15). These inconsistencies have led to questions regarding their depiction of traveler behav- ior. In addition, the four-step process treats travel choices as independent choices, whereas in reality they are not mutually exclusive. For example, the decision to make a trip in the first place (generation) may be a func- tion in part of the availability of a particular mode (modal split). Some models have addressed this by com- bining steps [e.g., trip distribution and modal split (15) or the trip generation, distribution, and modal split (14)]. Other models have attempted to address this more simply by introducing feedback loops among the steps [i.e., which is not inherent to the four-step process (16)]—notably, the travel times that result from the trip assignment are fed back to trip distribution to provide a more realistic depiction of the “true” travel times between zones, with the distribution–modal split- assignment process then iterating several times until an equilibrium is reached. • Similarly, there is inadequate feedback between the travel demand forecasting model and its land use inputs. The implication is that the changed travel pat- terns can affect the distribution, magnitude, and type of development over time (e.g., an expressway extension that improves accessibility to a new suburb), which in turn affects the characteristics of travel demand. Efforts in different U.S. cities and elsewhere to develop inte- grated land use and transportation models have been documented (17). In addition, several techniques have been used to forecast these land use inputs (18): the importance for toll road demand forecasting is that there is no consistency in modeling technique; this time for key inputs. • Forecasting methods can be divided into two broad groups: macro-analytical methods, which are based on zonal averages, and micro-analytical methods, which are based on individuals and households. The four-step process is in the first group. Because of their low cost and technical simplicity, macro-level forecasts remain popular; however, it is precisely these two reasons that lead to questionable and inaccurate results (18). In con- trast, micro-level forecasts can predict impacts with more detail and accuracy. More generally, the development of micro-level forecasting capabilities also addresses the behavioral inconsistencies identified previously (simultaneous choices, lack of feedback, etc.), through the use of activity-based models. This approach treats travel as being derived from the demand for personal activities, so that travel decisions become part of an individual’s broader activity-scheduling process. In turn, activities are modeled, rather than only trips. The basic travel unit is a tour, which is defined as “the sequence of trip segments that start at home and end at home” (19). This allows for a more consistent and inclusive treatment of the individual’s decisions (when, where, why, and how to travel); links these decisions for all of an individ- ual’s trips over the course of the day, allows the deci- sions to be analyzed in the context of the decisions of other members of the households, and, allows for con- sideration of lifestyles (e.g., “commuting” by Internet) (12). The resultant chain of decisions means that higher-level decisions are fully informed about lower- level decisions (i.e., decisions are “nested”) (20). Emerging methods also allow the simulation of an individual’s activities dynamically, meaning that this micro-level treatment eliminates the need for zonal aggregations, allows the heterogeneous (travel) char- acteristics of the population to be analyzed, and has the

potential to generate “emergent behavior” (i.e., behav- ior is not explicitly “hard-wired” into the model, based on its calibration to conditions at a particular point in time) (12). • Methods to model time-of-day choice are emerging. This refers to the relative lack of consideration of tem- poral considerations in demand modeling—that is, the traveler’s choices are related to choices regarding the time of day in which the trip is made. Time-of-day choice can be expressed in terms of the time “slices” that are modeled; the days that are modeled (e.g., weekday versus weekend or holiday); peak spreading (i.e., the allocation of trips between the peak hour or half-hour and the peak “shoulders,” as the expansion of the duration of the peak period over time); and time-of-day choice modeling [i.e., the explicit model- ing of the time at which the traveler starts his or her trip in order to arrive at a destination within a desired “envelope” (e.g., between 8:45 and 9:00 a.m. every morning)]. The aforementioned need to develop improved fac- tors for expanding peak-hour volumes to yield annual revenues is one manifestation of the importance of time-of-day modeling to toll road demand estimation. Also important is its potential to depict more accurately a traveler’s response to congestion: rather than switch routes (to an uncongested toll road), the driver may advance or delay the start time of his or her trip to a less congested time of day (or simply not travel). Where time-of-day choice is considered in practice, the most common consideration has been peak spreading. One peak spreading model accounted for congestion when determining the proportion of a.m. peak-period vehicle traffic. It was hypothesized that “the total congestion for a trip is a primary reason for peak spreading rather than the congestion of, possi- bly, one link,” thus establishing the need to account for congestion throughout the network in addition to trip purpose and trip distance. In other words, the advantages offered by a toll facility must be consid- ered in the context of its impacts on overall network congestion. Although the model predicted the flatten- ing of the a.m. peak period as congestion and trip length increase, it assumed that a constant duration of the peak period (in this case, 3 h). That is, the pro- portion of daily travel that occurs in the 3-h peak period was assumed to remain stable over time (which may not be appropriate in all cities or for toll roads, especially as the duration of peak period grows or off-peak traffic volumes increase). The report con- cluded that both trip purpose and trip distance, in addition to congestion, were important parameters in a model that predicts peak spreading (21). A more recent analysis concluded that the use of dynamic traffic assignment (such as equilibrium assign- ment) models as a means to predict the impact of new infrastructure should account for departure time choice 14 in addition to route choice. The resultant model took into account the need for travelers to arrive at their des- tination, for particular trip purposes (e.g., going to work, to classes, or to an appointment), at a particular time, which in turn determines their departure time. Each traveler has a preferred departure and arrival time, any deviation from which (owing to congestion) causes disutility. The dynamic traffic assignment models time choice simultaneously with route choice (22). Another emerging development is the use of “equilibrium scheduling theory,” which simulates departure time choice modeling in the context of an equilibrium net- work model. This approach models the build-up and decay of travel times during the peak periods, taking into account the disutility of arriving before or after a preferred arrival time window (23). • Network micro-simulation models have come into use as tools to simulate the dynamics of traffic along cor- ridors and networks. Whereas the travel demand fore- casting models simulate average speeds for an hour’s slice of traffic, these models use micro-simulation tech- niques to represent traffic flows microscopically through a network as a series of individual vehicles and tracks each vehicle’s progress at a finite resolution, which is typically one second or less. This logic permits consid- erable flexibility in representing spatial variations in traffic conditions over time and allows for the analysis of such traffic phenomena as shockwaves, gap accep- tance, and weaving. The importance of network micro- simulation models to toll road demand forecasting lies in the emergence of managed lanes as tolled options in several cities; specifically, in their ability to simulate the dynamics of individual lanes and the diversion of drivers between lanes. Network micro-simulation models provide a more detailed approach by taking into account the transient effects on speed and acceleration as the vehicle travels on a road network. By modeling vehicle kinematics (instantaneous speed and acceleration) on a road net- work, more reliable estimates of vehicle energy con- sumption and emissions result. Typically, micro-simulation traffic models are appli- cable only to sub-networks of larger urban areas. This is a result, in part, of the required level of detail necessary of model inputs and the subsequent strain these require- ments place on computing capabilities. The dynamics of individual vehicles are defined in terms of the number of departures from each origin–destination pair, the deter- mination of vehicle speed based on car-following logic, and requirements for lane changing. The speed of the vehicle along that first link, as well as any subsequent links, is updated at discrete time intervals typically between 0.1 and 1.0 s. Each update reflects the distance headway between the vehicle in question and the vehi- cle immediately preceding this vehicle; whereas the exact speed for any given distance headway is based on a link-specific, car-following relationship. Beyond the

15 speed restrictions, which arise from the above car- following logic, a vehicle’s progress can also be delayed at traffic signals, ramp meters, queues, and/or other bottlenecks. The effects can be time-varying and may vary both spatially and temporally, which permits the replication of shock waves within the model. When a vehicle travels down a particular link, it may make dis- cretionary lane changes to maximize travel freedom (speed). Conversely, mandatory lane changes may be required owing to the prevailing network geometry and routing behavior (24). In sum, these developments provide opportunities to improve the overall state of the practice in travel demand forecasting, as well as that of toll road demand forecasting. However, many of these developments are emerging: there is only a very small number of practical applications of activity-based models in the United States (20) and few ap- plications of time-of-day choice modeling. The TRANSIMS initiative of the federally funded TMIP also can be expected to affect transportation planning practice (13). Conversely, network micro-simulation models are well-established in transportation planning practice, with several recent man- aged lane applications. Methods for Modeling Toll Road Demand No state-of-the-art consensus exists among transportation researchers and practitioners regarding the best methods for achieving traffic and revenue forecasts (25). This mir- rors the general application of models in transportation planning practice. Methods being used today can still be categorized primarily by incremental or synthetic analysis, both of which the transportation planning community has been using. The choice of analytical method varies, based on the method that is used to develop origin–destination trip tables for a given time period, trip purpose, and travel market segment (25). A review of the state of the practice for value pricing projects in several U.S. cities identified the following five categories of modeling procedures (20). Although the review primarily addressed forecasts for managed lanes, the catego- rization is applicable more generally to toll road demand forecasts. 1. Modeled as part of an activity-based model—The state of the art in demand modeling allows for the inclusion of pricing into the decision hierarchy. A combination of revealed and stated preference surveys could be used as the basis, with the stated preference data allow- ing for the modeling of choices that do not yet exist. Only Portland, Oregon, a pioneer in the development of activity-based models, has applied this type of model to the subject (i.e., to an analysis of value pric- ing). The practical use of activity-based models in transportation planning is only now emerging and rep- resents a significant effort (20). Only one respondent to the survey of practitioners indicated the use of an activity-based model, which was applied to a toll bridge. 2. Modeled within the modal split component of a four- step model—Automobile trips on a tolled or non- tolled road are considered as distinct modal choices, with separate modal split functions for work (or work- related) and non-work trip purposes (given the corre- sponding differences in values of time). The advantage of this approach is that out-of-pocket costs can be modeled explicitly, because travelers’ utilities are “directly affected by the value of tolls and so are the respective modal shares”—that is, the approach ensures “robustness” in the results. The approach also can be expanded to trip distribution modeling, because the impedance incorporates the impact of tolls more explicitly. The ability to incorporate stated preference data into revealed preference data, as a means to account for nonexistent facilities, again was noted (20). Phoenix, Arizona, and Sacramento, California, were cited as examples of urban areas that have used this approach. The Phoenix [Maricopa Association of Governments (MAG)] model distinguishes between the SOV trip and the HOV trip. Because MAG allows vehicles with only two occupants to use its HOV lanes, the tolled/non-tolled choice is included only in the util- ity function of the SOV trips. The function includes a travel time savings term that is equivalent to the dif- ference between tolled and non-tolled travel time. MAG’s trip distribution model was being updated to account for these impedances (20). The analysis for the Minneapolis–St. Paul man- aged lane system (MnPASS) incorporated tolled SOV trips as a modal choice into the regional model. Values of time were developed for two trip purposes (home-based work and other trips), both classified by three automobile availability categories (number of vehicles per household, which had been found to be a determinant of value of time) and whether the trip was destined to the central business district (again, found to be a determinant). The basic values were adapted to this model from previous local studies or from experience elsewhere, given that the time frame available for the analysis precluded the collection of new data. The revised modal split function was used to screen alternative network configurations. For the purpose of the analysis, the resultant revised imped- ances were not implemented into the trip distribution component, to allow the alternatives to be compared on a common basis. In other words, although behav- iorally the tolls (i.e., the revised impedances) would affect trip distribution (given the appropriate feed- back loops), it was felt that the impacts would be small when compared with the ability to compare alternatives (26).

The MnPASS study used value-of-time data from an evaluation of the impacts of the Riverside Freeway (SR-91) tolled express lanes in Orange County, California. Based on observations from 3 years of operation (after the express lanes opened in 1995) and from traveler surveys, the evaluation found that gen- der (female) was a strong determinant of the use of the facility, with other factors [high income, middle age, higher education, and commuting to work (i.e., work or work-related)] also being indirect factors (indirect in that they determined the willingness to purchase an electronic transponder, without which drivers were unlikely to use the express lanes). Logit choice mod- els were developed according to these factors (27). The SR-91 evaluation identified several important determinants of managed lanes—gender, income, age, education, and trip purpose. Although the mix of deter- minants and their values might vary by location, there is (for example) an observed correlation between trav- elers’ income and the likelihood of using toll roads, with higher-income travelers more likely users than lower-income travelers (20). The importance of considering time-of-day impacts was underscored by a recent study of the impacts of the Port Authority of New York and New Jersey’s time- of-day pricing scheme, which it introduced in March 2001. The study found that 7% of passenger trips and 20% of truck trips changed behavior because of the new pricing scheme. The percent share of peak shoul- der trips for both trucks and automobiles also increased during weekdays (28). The modeling of toll demand as part of modal split requires that the generalized cost impedances (i.e., impedances that account for monetary values—such as tolls—as well as travel times) are fed back from trip assignment to trip distribution and modal split. The process iterates until a stable equilibrium is achieved; that is, when there are no significant differences in the impedances between two iterations (29). 3. Modeled within the trip assignment component of a model—This approach applies a diversion of trips within the trip assignment; that is, after (or in the absence of) demand modeling. It assumes that trip dis- tribution and modal shares (not differentiating between tolled and non-tolled automobile trips) remain unchanged in the absence of feedback loops. There are two general methods for modeling traf- fic diversion in trip assignment: The first translates the monetary toll into a time equivalent through the use of values of time. The equivalent times are then incorporated into the model’s volume-delay func- tions, which—using the equilibrium assignment technique—are used in turn to allocate trips among different paths according to travel time, capacity, and congestion. Queuing and service time at toll plazas similarly can be incorporated into the func- tion. (In essence, the tolls and plaza times are added 16 as “penalties” to the modeling of actual travel time.) Values of time can be derived for different trip pur- poses, income levels, etc. (25). The second method uses diversion curves as the basis for toll forecasts. This commonly takes the form of a logit function, which calculates the propensity to use a tolled facility (the facility’s share of traffic) as a function of the relative cost or travel time between the tolled and non-tolled route (i.e., for each origin– destination path that could use the tolled facility). The slope on the S-shaped diversion curves represents the elasticity of demand with respect to the relative cost or travel time using the tolled road. The elasticity of demand is related inversely to the value of time or willingness to pay. The shape of the curve can be determined in two ways: using observed data (in which case the value of time is implicit) or from a sta- tistically estimated logit function based on revealed and/or stated preference survey data. The curves can be fitted according to different trip purposes and vehi- cle occupancies. The diversion curve is applied to the relevant trip table (for a given purpose, income group, automobile occupancy, time period, etc.) to derive tolled and non-tolled trip tables. These then are assigned to the network to yield both updated imped- ances (and the process is repeated until an equilibrium is reached) and, ultimately, estimates of revenues (25). The primary benefit of using diversion models to estimate toll road demand is that they can be applied to an existing four-step model, without having to recalibrate it. However, the shape of the curve, and the data upon which it is based, generally are held as con- fidential or proprietary and so are not available to other users (29). A variation is the use of a dual minimum path (equi- librium) assignment, which develops two sets of paths for each origin–destination pair: one using the tolled facility (where applicable) and one without the tolled facility. A proportion of the total trips between each zonal pair is assigned to each network path, according to the relative respective total costs, which can include vehicle operating costs as well as travel time costs and the costs of tolls (2). An example of the second method (diversion) is provided by a traffic and revenue forecasting study for a proposed toll highway near Austin, Texas. A logit model was developed for several trip purposes, based on a stated preference survey. The utility func- tions for the work-related trip purposes were found to be sensitive to traveler income. The tolling diver- sion logit model was incorporated into the trip assignment component of an updated regional travel demand model. The model took into account differ- ent payment options (cash, cash plus electronic, and electronic only). The development of the logit model also accounted for toll road bias (the negative propensity to use a tolled road) and an electronic toll

17 collection bias (the increased likelihood of using a tolled facility, owing to the convenience associated with electronic toll collection). Both terms largely offset each other, with the toll road bias found to be common in regions that had no prior experience with tolling (30). It should be noted that the assignment impedances were not fed back to the trip distribution and modal split models; that is, the trip origins and destinations were assumed not to change under traf- fic diversion. The survey of practitioners indicated that both methods were used. 4. Modeled as a post-processor—This approach can be used either within the framework of a four-step model or exogenously using the output of the four-step model. Washington, D.C., and San Diego, California, provided examples of the former, in which assigned volumes are diverted (i.e., after trip assignment) from general purpose lanes to managed lanes according to the excess capacity available in the latter. An example of the latter is provided in Minneapolis–St. Paul, in which the outputs of the regional model were input to the FHWA’s Surface Transportation Efficiency Analysis Model to calculate costs and tolls as part of a pricing study. The procedures are operationally simple to implement; however, they are not sensitive to changes in traveler behavior (20). 5. Model as a sketch planning method—These are quick response tools that are used for project evaluation. Examples include the FHWA’s Spreadsheet Model for Induced Travel Estimation, which estimates induced traffic (as a result of faster facility travel speeds, traf- fic diverted from other facilities, destinations, or modes) as a function of elasticities of demand with respect to travel time, with price and demand equili- brated as part of the procedure. A modified version is the Spreadsheet Model for Induced Travel Estimation- Managed Lane, which uses a pivot-point logit model to estimate changes in travel demand according to changes in travel time and tolls as well as improved transit service. The “model is relatively simple to implement and can be considered a reasonable tool for the initial screening of alternatives or in situations where results of formal travel models are not avail- able.” The FHWA’s Sketch Planning for Road Use Charge Evaluation model also uses a pivot-point mode choice model to estimate changes in mode (i.e., man- aged lanes) and the associated revenues, costs, and travel time delays (20). The Texas Transportation Institute has developed a spreadsheet-based Toll Viability Screening Tool. The spreadsheet provides a way to assess the economic via- bility of a proposed tolled facility in advance of the need for a more detailed traffic and revenue forecast. It does so by assessing the potential variability of the initial demand, by subjecting various input parameters to a triangular distribution function (similar to a normal distribution)—that is, to a distribution that measures the likelihood of their occurrence. As input, the tool requires daily traffic volumes, toll rates, and assumed diversion. Results (revenues) are expressed as net present values. The tool also supports a risk analysis of the results, taking into account the distri- bution, which allows sensitivity tests of the inputs, which are the most important (31). Some survey respondents reported using a spread- sheet model to estimate travel demand for short-term (i.e., annual) revenue forecasts of an established toll facility, such as a bridge or tunnel for which travel demand was stable and dependable historical data were available. A review of literature indicated that this was a common practice. For example, Florida DOT’s Annual Report on its Enterprise Toll Opera- tions stated that for older, established toll facilities its forecasts were developed based on actual traffic and revenue performance, with adjustments for popula- tion growth and anticipated future events (such as new infrastructure) (32). The survey of practitioners did not demonstrate any con- sistent or dominant treatments of the types of choices that were modeled or how they were modeled. This appears con- sistent with the practice of travel demand modeling in general. This is supported by research that summarized the treatment of pricing in seven models at major cities across the United States, which found that no two of the seven models treated the subject in exactly the same manner (33). It is important to note that there is no fixed or standard process for determining which type of model must be used. Rather, the object is to ensure that the method meets the need. This was corroborated by the survey, which indicated that a range of model types was used. For example, one respondent to the survey, a state DOT, noted that it used two different models for two concurrent studies. One study was determin- ing the impacts of an impending change in the toll rate on an existing bridge. The second study was for a proposed HOT lane. The first study used a semi-modeling approach with implied elasticity, using different values of time for the toll bridge users. The respondent found this method to be effec- tive, given the known travel characteristics of the existing toll bridge users. The second method was fully model-based, using a combination of a traditional four-step travel model and a network micro-simulation model. This combination was chosen to capture the significant impact of small changes in traffic volumes on the HOT lane compared with the regu- lar lanes. Finally, it should be noted that practitioners appear to have responded to the specific needs of forecasting for toll roads. For example, the survey of practitioners indicated that various combinations of time periods were modeled; most of the cited travel demand models covered the week- day off-peak periods in addition to the (more typical) peak periods. In certain cases, nighttime and weekend periods

were also modeled to reflect the type of facility (i.e., the weekend model for areas of high recreational use). A small number of respondents incorporated time-of-day choice into their models or used peak spreading models. More commonly, practitioners relied on factors from other sources (e.g., from traffic counts or from the local MPO) to address time choice, whereas others did not include time choice at all. Evolution of Decision-Making Environment Concerns regarding the reliability, accuracy, and credibility of travel demand forecasts are not new. A 1989 U.S.DOT study compared projected and actual ridership and costs for 10 heavy- and light-rail transit projects in 9 U.S. cities. The study found that the actual ridership for each of the 10 proj- ects was significantly below the projections, whereas the actual costs were higher than the projected costs in 9 of the 10 projects. The projected ridership (i.e., benefits) and costs were used as the basis of investment decisions and of appli- cations for federal government funding (34). Although the study addressed only forecasts for transit, it is relevant to toll road demand and revenue forecasts because it received widespread attention in the transportation com- munity and also because it anticipated many of the issues that have since been identified in toll road forecasts. Hence, it provides an important context for the discussion of how the decision-making environment has evolved. For example, the study explained the need for the community to understand the accuracy of the ridership and cost forecasts in three ways: the transit projects represented the largest investment ever in public works in each of the nine cities, local officials in other cities where rail projects were contemplated would also rely on similar projections to make their own decisions, and local officials typically used similar analytical processes for other public investments. To this end, the study found that these “mistakes” in ridership estimates could not be explained by differences between projected and actual values of the deter- minants of ridership: land use inputs (which differed little from the actual), network configurations, assumed feeder bus configurations, or downtown parking prices (which tended to be lower than those modeled). Each of the 10 projects was selected among several alternatives. The study noted that although the accuracy of the forecasts for the rejected alter- natives could not be evaluated, for almost all of the projects the divergence between the projected and actual ridership and costs of the selected alternative was greater than the entire range of the ridership and costs of all the alternatives that were compared (which made it “extremely unlikely that a rail project would have prevailed in the presence of more reliable forecasts”). Rather, the study attributed the differ- ences to the structure and nature of federal transit grant and fund programs (effectively favoring high-capital transit investments), which provided little incentive to local deci- sion makers “to seek accurate information in evaluating 18 alternatives.” The result was a “bias” or “optimism” for rail transit (35). These differences (and those in other areas of public pol- icy) demonstrate a “serious ethical problem” in the use of fore- casts, with occurrences noted in which modelers had been directed by their superiors (including local elected officials) to “revise” their ridership forecasts upwards, to “gain federal [financial] support for the projects whether or not they could be fully justified on technical grounds. Forecasts are presented to the public as instruments for deciding whether or not a proj- ect is to be undertaken; but they are actually instruments for getting public funds committed to a favored project.” The goal of “exaggerated forecast[s] of demand and the cost underesti- mates” may be to “[get] the project built rather than honestly evaluating its social benefits” (34). Forecasts have not become more accurate over time. In a multinational statistical analysis of 183 road projects (tolled and non-tolled) completed between 1969 and 1998, the “forecasts [appeared] to become more inaccurate toward the end of the 30-year period studied” (36). More recent fore- casts were found to be more comprehensive than older stud- ies; however, “this greater depth has not yet appeared to improve the accuracy of the forecasts.” Newer forecasts did appear to respond to earlier concern; for example, by incor- porating better methods to forecast ramp-up volumes. How- ever, “whether this increased scrutiny has actually led to more accurate forecasts remains to be seen” (3). The role of private provision of public services (such as privately owned tolled roads) continues to evolve. Although not specifically directed at the reliability of traffic and rev- enue forecasts, an article about “intellectual dishonesty” in the ongoing debate may provide some context. For example, the toll revenues for a (hypothetical) bridge that is operated by a private company must cover its capital, operating and maintenance costs, as well as depreciation, which reflects the eventual need for rehabilitation or reconstruction as a result of wear and tear. Its toll rates must be set sufficiently high to cover these costs. Because the company accounted for annual depreciation costs when it issued its debt to construct the bridge (which presumably has been paid off by the time major reconstruction is required), a one-time debt allows the construction of a bridge that “can presumably last forever.” In contrast, public authorities in the United States are not required to account for depreciation, which means—for the same tolled bridge—its toll rates could be much lower. How- ever, it must issue new debt when the bridge is reconstructed (i.e., the public authority inevitably must account for depre- ciation, but does so in terms of a “perpetual debt”). This means that the public authority’s bridge seems “less expen- sive,” because its lower toll rates ultimately have transferred the debt from its actual users to future generations (37). The relevance to this synthesis is that the public’s expectations and inappropriate understanding of the real costs of public services may affect the choice of toll rates and, in turn,

19 traveler behavior (which may not be captured properly by the forecasting models or the data upon which they are based). A 1989 court case in the San Francisco Bay area claimed that state and regional planning authorities had not suffi- ciently met their obligations to reduce air pollution in their transportation plans. Much of the resultant findings focused on the adequacy and use of the regional travel demand fore- casting model in predicting air quality impacts. In particular, it took a much more “literal” interpretation of model fore- casts than had planners historically (i.e., given the planners’ understanding of the models’ limitations owing to errors in calibration, data input, or validation). A subsequent TRB study found that the “analytical methods in use are inade- quate for addressing regulatory requirements” (such as air quality conformity analysis) (13). The relevance to this syn- thesis is that the concerns about model inaccuracies and per- formance that this court case identified, which preceded the TMIP and which, in part, the TMIP was intended to address, mirrors and anticipates similar concerns regarding the per- formance of toll demand and revenue forecasts. RELATIONSHIP BETWEEN DEMAND AND REVENUE FORECASTS Revenue forecasts are dependent on travel demand forecasts and the assumptions on which the travel forecasts were based. Critical assumptions include local growth policies, the magnitude and distribution of future land uses, the intensity of development, projected economic growth, changes in traf- fic patterns, drivers’ willingness to pay tolls, and new com- peting roads in the transportation network. The level of uncertainty in revenue forecasts is proportional to the level of uncertainty in travel demand forecasts. Revenue forecasts are also dependent on the tolling tech- nology, toll rate structure and schedule, and the stratification of the toll road users (i.e., according to payment classes). Tolling schemes could include discounts for electronic tolling or multipass users, higher tolls for heavy vehicles, or variable tolls based on time of day or section of toll road used. Increases in toll rates can also affect the demand, especially as some authorities have elected to increase toll rates more sharply than projected to quickly generate revenues in the short term (when the projected demand had not materialized). As noted, the travel demand forecasts are commonly developed for a weekday peak hour or peak period for sev- eral modeled horizon years. Conversion factors are then applied to generate daily and yearly traffic volumes. Revenue is estimated by multiplying the forecast volumes by the toll amount, taking into account different toll rates for vehicle type, potential toll evasion, discounts, and other facility- specific factors. With each assumption, a degree of error is introduced into the revenue forecast. Another layer of com- plexity is added when a schedule of predetermined toll rate increases is applied to the traffic forecasts. In travel demand forecasting, the future year forecasts (20- to 30-year horizon) are more important and critical for long-term planning decisions. However, for revenue fore- casts, the initial years of operation are crucial in terms of assessing and managing financial risk. This is because the risk for default is typically at its highest during this period, which is also referred to as the ramp-up period (9). During the ramp-up period, traffic volumes may be significantly lower than forecasted as drivers slowly become aware of the toll facility and its potential for saving time and/or conve- nience, or if population or employment growth along the facility corridor (i.e., the potential market) is also less than forecasted. PERFORMANCE OF TOLL ROAD DEMAND AND REVENUE FORECASTS Sources of Information This section compares the projected and actual revenues for several facilities. However, to understand and interpret the comparison, it is important first to understand the sources upon which the information was based. The projected and actual revenues were derived from dif- ferent sources. Projections are commonly provided by the original traffic and revenue studies for the individual facility. The study is typically conducted several years before the facility’s opening date, as the basis for securing funding for the planned facility. The actual traffic and revenue studies proved difficult to obtain for three reasons: • An accessible single or universal source or database of these traffic and revenue studies does not exist. Mem- bers of the financial community, such as bond rating agencies, do have access to a database of financial offerings, which include traffic and revenue studies; however, access is available only by subscription. Moreover, the database is not exhaustive. • With some exceptions, facility owners generally were not willing to provide their traffic and revenue reports, which they considered proprietary or confidential. • Some authorities have updated their traffic and revenue forecasts, in the face of poor performance (projected versus actual) and given the availability of observed traffic and revenues. The new forecasts replace the orig- inal study (meaning also that newer models or forecast- ing methods may be used, as well as newer data)—that is, a series of forecasts may be available for a given facility. In general, the new forecast produces much closer results in the subsequent years. Accordingly, the authorities use the updated study as a comparison with actual revenues, which in turn often demonstrates a much better performance than the original traffic and revenue study would indicate.

Other authorities prepare simplified projections of annual revenues. These are based on an extrapolation of the previous year’s (or years’) revenues, using growth factors that were developed from observed growth trends (e.g., in traffic volumes) without recourse to a travel demand forecasting model. Information on the “actual” revenues generally was more readily available. Annual toll revenue statistics generally were accessible from annual reports or directly on the owner’s website. However, there is considerable variation as to the amount of yearly data that each owner provides. For example, some owners report only the most recent year, whereas oth- ers provide information for several years. Most authorities reported only the three most recent years, with only a few pro- viding information for up to 10 (or more) years. That is, infor- mation for older facilities (i.e., pre-2000) was not readily accessible. The different sources, and the difficulty in procuring the different pieces of information, also suggest that the compa- rability of the projected and actual revenues for a given year may be limited. The definition of a “year” may vary between the projection and the actual (e.g., the definition of the fiscal year may reflect that of the owner rather than of the facility; and some facilities may have begun operation part way through the owner’s fiscal year). In summary, the comparison was derived from four types of sources: • Comparisons of actual and projected revenues, pre- pared by various bond rating agencies. • Financial offering statements for individual facilities, which include the traffic and revenue projections for the facility. These statements are circulated within the financial community by subscription to a central com- mercial service. • Financial statements or reports, prepared by individual authorities (owners). Generally, these were found on the respective authority’s website. In most cases, only the actual revenues were provided, although a small number of websites also compared these with the pro- jected revenues. Of the four types of sources, only this one is available to the general public. • Traffic and revenue forecasts, provided by individ- ual facility owners that responded to the survey of practitioners. It is important to note that several sources were used to com- pile the information for some facilities described in this syn- thesis. This is important for three reasons: First, as noted in the footnotes to Table 1, in some cases the reporting methods var- ied from year to year. Second, multiple sources of information, and different performance results, were sometimes provided for 20 a given facility and year. Third, although the actual perfor- mance information was provided, for some facilities the corre- sponding projected performance was not available. Comparison of Projected and Actual Revenues Table 1 summarizes the performances of 26 different toll highways throughout the United States. The table compares the actual revenue collected as a percentage of the revenue that was projected in traffic and revenue forecasts. The facil- ities are listed according to the year in which the facility opened (between 1986 and 2004). The results are presented, where available (or where applicable; some of the facilities opened too recently to have an established performance his- tory), for the first 5 years of operation. The table identifies the owner and the state in which the facility is located. Appendix D presents brief descriptions of the individual facilities. It should be noted that other facilities were also investi- gated; however, they were not included because of insuffi- cient data and information. Table 1 demonstrates considerable variation in perfor- mance, ranging from a low of 13.0% for the Osceola County Parkway in Year 1, to a high of 152.2% for the George Bush Expressway, also in Year 1. The table also shows that there is little consistency, as follows: • The results do not improve with newer facilities, which might have been expected given that the state of the practice in modeling generally is improving. The per- formance does not necessarily improve for a given authority [i.e., even as a history of models and forecasts is built up by (or for) a given authority, the perfor- mance does not necessarily improve as a new facility is planned]. • There is little consistency by year within a given facility, although the performance for some facilities improves when traffic and revenue forecasts are updated, based on actual in-operation performance. (The most recently opened facilities are too new to have recorded data for any but the initial year or two). • Most of the results demonstrate an underperformance (actual is lower than projected), albeit with some notable exceptions. However, the under/overperformance may vary within a given facility by year. • At least some of the results reflect updated fore- casts (although the existence of updates may not have been noted in the source material). This is corroborated by the survey of practitioners: In response to poor ini- tial performance, some respondents indicated that their model was recalibrated or the model networks were reconfigured; the demand forecasts or the revenue fore-

21 Authority/Facility Year of Opening Year 1 Year 2 Year 3 Year 4 Year 5 Florida’s Turnpike Enterprise/Sawgrass Expressway (6) 1986 17.8% 23.4% 32.0% 37.1% 38.4% North Texas Tollway Authority/Dallas North Tollway (6) 1986, 1987 73.9% 91.3% 94.7% 99.3% 99.0% Harris County Toll Road Authority (Texas)/Hardy (6) 1988 29.2% 27.7% 23.8% 22.8% 22.3% Harris County Toll Road Authority (Texas)/Sam Houston (6) 1988, 1990 64.9% 79.7% 81.0% 83.2% 78.0% Illinois State Toll Highway Authority/ Illinois North South Tollway (6) 1989 94.7% 104.3% 112.5% 116.9% 115.3% Orlando–Orange Expressway Authority/ Central Florida Greenway North Segment (6) 1989 96.8% 85.7% 81.4% 69.6% 77.1% Orlando-Orange Expressway Authority/ Central Florida Greenway South Segment (6) 1990 34.1% 36.2% 36.0% 50.0% NA Oklahoma Turnpike Authority/ John Kilpatrick (3) 1991 18.0% 26.4% 29.3% 31.4% 34.7% Oklahoma Turnpike Authority/ Creek (3) 1992 49.0% 55.0% 56.8% 59.2% 65.5% Mid-Bay Bridge Authority (Florida)/ Choctawhatchee Bay Bridge (38,39) 1993 79.8% 95.5% 108.9% 113.2% 116.7% Orlando-Orange Expressway Authority/ Central Florida Greenway Southern Connector (6) 1993 27.5% 36.6% NA NA NA State Road and Tollway Authority (Georgia)/GA 400 (3) 1993 117.0% 133.1% 139.8% 145.8% 141.8% Florida’s Turnpike Enterprise/ Veteran’s Expressway (3) 1994 50.1% 52.9% 62.5% 65.0% 56.8% Florida’s Turnpike Enterprise/ Seminole Expressway (3) 1994 45.6% 58.0% 70.7% 78.4% 70.1% Transportation Corridor Agencies (California)/Foothill North (3) 1995 86.5% 92.3% 99.3% NA 1 NA1 Osceola County (Florida)/Osceola County Parkway (3) 1995 13.0% 50.7% 38.5% 40.4% NA Toll Road Investment Partnership (Virginia)/Dulles Greenway (3) 1995 20.1% 24.9% 23.6% 25.8% 35.4% Transportation Corridor Agencies (California)/San Joaquin Hills (3) 1996 31.6% 47.5% 51.5% 52.9% 54.1% North Texas Tollway Authority/ George Bush Expressway (3) 1998 152.2% 91.8% NA NA NA Transportation Corridor Agencies (California)/Foothill Eastern (3) 1999 119.1% 79.0% 79.2% NA 1 NA1 E-470 Public Highway Authority (Colorado)/E-470 (3) 1999 61.8% 59.6% NA 95.4% 2 NA3 Florida’s Turnpike Enterprise/Polk (3) 1999 81.0% 67.5% NA NA NA Santa Rosa Bay Bridge Authority (Florida)/Garcon Point Bridge (42,43) 1999 32.6% 54.8% 50.5% 47.1% 48.7% Connector 2000 Association (South Carolina)/Greenville Connector (3) 2001 29.6% NA NA NA NA Pocahontas Parkway Association (Virginia)/Pocahontas Parkway (44,45) 2002 41.6% 4 40.4% 50.8% NA NA Northwest Parkway Public Highway Authority (Colorado)/Northwest Parkway (46,47) 2004 60.5% 56%5 NA NA NA Sources are cited in parentheses. Notes: Bold type reflects actual within 10% of projected. NA = traffic and revenue report not available or not provided. 1For these years, the Transportation Corridor Agencies combined the revenues (earnings) for the two facilitie s (Foothill North and Foothill Eastern). Accordingly, the individual performance for the two facilities cannot be calculated. 2Data reflect updated traffic and revenue study (40,41). 3Incomplete information (missing November and December). 4This is approximated owing to construction delays that only allowed the facility to be open for one-quarter of the expected full year. 5Projected performance for the 2005 fiscal year (48). TABLE 1 ACTUAL REVENUE AS PERCENTAGE OF PROJECTED RESULTS OF OPERATION

casts were revised. Other responses included revisions to the financial schedule, changes to the staging or tim- ing of the project, or the implementation of annual updates and peer reviews. On the other hand, several respondents noted that the forecasts were accepted and used as is (i.e., no impact). • Even with the availability of updated forecasts, only a small number of projections are within 10% of the actual revenues. These are indicated in bold type in the table. Comparison of Projected and Actual Traffic It should be noted that Table 1 and the preceding discussion compared projected and actual revenues, as opposed to traf- fic. However, a multi-national review of 183 tolled and non-tolled roads found significant inaccuracies in the traf- fic projections as well (36). Another study compared the traffic forecasts for 104 tolled facilities around the world. The comparison found considerable variability in the performance of the traffic forecasts for the first year (during ramp-up), ranging between 15% and 150% of actual performance. On average, the forecasts overestimated Year 1 traffic by 20%–30%. This “optimism bias” (error) was not reduced for subsequent years; rather, a mixed performance profile resulted. The mean pro- jected versus actual performance ranged between 0.77 and 0.80 over the first 5 years of operation. The comparison disaggregated the forecasts according to vehicle type, and found that the variability in traffic forecasts was “consistently higher” for trucks than for light vehicles (generally, private automobiles). This reflected the greater difficulties in predicting the trucking community’s response to tolls, given the variability in type and size of trucking operations. The significance is that trucks commonly pay higher tariffs than private vehicles, meaning that their con- tribution to revenue forecasts can be “significant,” out of pro- portion to their volumes. The comparison also noted the relative lack of tolled facil- ities that are more than 5 years old. This reflects “the innova- tive nature of the sector and that operational project-financed infrastructure concessions are a relatively recent phenome- non. A significant number of highway concessions globally still remained in design or under construction” (49). The difficulty in tracking the performance of the forecasts over time was also noted, given “the common practice of preparing revised or rebased forecasts for toll facilities whose predicted use departs significantly from expectations. In such instances, credit surveillance documentation may fail to report the original forecasts” (49). The analysis also compared four different forecasts for the same tolled facility. All of the forecasts represented “base- 22 case forecasts”; that is, they were modeling the same situa- tion but used different forecasting assumptions. The four forecasts varied between 26% (for the Year 5 forecast) and 255% (Year 35), with a steady increase in the interim. Very different projections of asset use result from relatively small divergence among the model input assumptions. . . . Traffic forecasts, particularly in the medium to longer term, can remain very sensitive to marginal parameter changes within the modeling framework, even though these parameter values are drawn from an entirely plausible range. In terms of assessing the reliability of future project cash flows, rigorous sensitivity testing clearly has a pivotal role to play in such cases (49). Explanation of Performance A second study assessed the performance of all but two of the toll facilities that are summarized in Table 1 (3). [The two exclusions were the Choctawhatchee Bay Bridge in Florida (Mid-Bay Bridge Authority) and the Northwest Parkway in Colorado (Northwest Parkway Public High- way Authority)]. The performance of each was assessed in the first 5 years of operation. All of these were start-up facilities. Whereas Table 1 considered the facilities chronologi- cally, to determine whether more recent forecasts presented any improvement in performance (as noted, no pattern was apparent), this assessment categorized the facilities accord- ing to several characteristics; location within the urban area, degree of integration with the existing road network, corri- dor income levels (i.e., the income levels of the drivers who would use the facility), time savings offered by the facility (i.e., the extent of congestion in the competing network and the availability of “competitive” non-tolled alternatives), value of time (e.g., the value of time would be highest in congested corridors traveled by high-income drivers), pro- jected traffic growth (also related to the reliability of the demographic and economic forecasts upon which the fore- casts were based), and the extent of development in the area served by the facility. The categorization resulted in four groups, although some overlap was noted: 1. High congestion, suburban areas; 2. Outlying roads of metropolitan areas; 3. Developed corridors, parallels of existing roads, and/ or faulty economic forecasts; and 4. Least developed areas. The general findings are summarized in Table 2. The table does not list the individual performances for each facility, because the values differ from those listed in Table 1. How- ever, the table demonstrates a decreasing performance accord- ing to the order of the four categories. In essence, improved performance resulted under the following conditions:

23 • Location within well-developed parts of a large metro- politan area, with established traffic patterns. • Location within high-income corridor, with resultant high values of time. • Well-connected to the road network. • Few or no reasonable choices for non-tolled alternative routes. • High savings in time offered by the facility. • Rapid driver acceptance of the new facility, with mod- erate ramp-up traffic growth and subsequently slower growth. • Moderate projected growth (i.e., appropriate account- ing for economic conditions; notably, through appro- priate consideration of the labor force, the aging of the population, and productivity). A third analysis of the poor performance of start-up roads identified four types of explanatory reasons (5): 1. Model input risk, which was exemplified by the use of regional travel demand models that had been developed for other purposes and that assumed land use and socioe- conomic forecasts that were appropriate for regional planning, but were not “sufficiently conservative” to support debt service; a “steady-state” forecast that does not account for “the very real likelihood” of economic fluctuations; weekend or truck traffic patterns that varied significantly from comparable experience; and differ- ences in actual values of time compared with estimates. The survey of practitioners found that few travel demand models had been created and calibrated specifically for a toll facility study. The survey revealed that only 15% of the models were devel- oped specifically for the subject toll facility study and another 31% had been calibrated for a previous toll facility study. The remainder of the survey respondents indicated that the selected model was based on an existing model that had been calibrated for other purposes. At the same time, survey respon- dents also reported that the use of a model from a previous toll or non-tolled study yielded the most accurate results. TABLE 2 PERFORMANCE BY CATEGORY Group Authority/Facility Characteristics Performance Explanation 1. High congestion, suburban Three facilities: • State Road and Tollway Authority (GA)/GA 400 • North Texas Tollway Authority/George Bush Expressway • Illinois State Toll Highway Authority/ Illinois North South Tollway • Well-developed urban/ suburban part of large metropolitan area • Higher corridor income • Substantial corridor traffic • High value of time • Good connections to facility • No competitive non- tolled alternatives • Modest projected traffic growth Approximated or exceeded projections • Moderate toll rates • Very rapid adjustment of traffic patterns following opening • Moderate traffic growth in first 2–3 years, then growing more slowly 2. Outlying Seven facilities: • Oklahoma Turnpike Authority/John Kilpatrick • Oklahoma Turnpike Authority/Creek • Florida’s Turnpike Enterprise/Veteran’s Expressway • Florida’s Turnpike Enterprise/Seminole Expressway • Florida’s Turnpike Enterprise/Polk • Transportation Corridor Agencies (CA)/Foothill North • Orlando–Orange Expressway Authority/ Central Florida Greenway North Segment • Less established traffic patterns • Less integral to the existing network • These were partial beltways • Usually serving above- average income areas, but with less- established development patterns • Further from employment centers • Moderate-to-high toll rates (although usage inelastic because drivers already accustomed to paying tolls) Mean ranged between 61% and 67% of forecasts, on average, with considerable variation • Substantial forecast revenue growth (35% average over first 4 years) • Forecast error appears to result from overestimation of initial base period usage (high ramp-up rates) (continued)

2. Ramp-up risk, with recent methods based on the use of other operating facilities as proxies, but with “spotty” results. 3. Event and political risk, for which were cited external factors such as the unforeseen construction or expan- sion of competing roads (San Joaquin Hills toll road), cancellation or postponement of expected expansions to the connecting network (Foothill Eastern), or the inhibition of expected development (which would have generated demand for the toll road) by a morato- rium on servicing (Garcon Point Bridge). The slow- down in air travel after the September 11, 2001, 24 terrorist attacks affected the forecasts for the E-470 toll road in Denver. Political pressures were cited as influencing fac- tors, given that transportation authorities exist in a political environment and this existence can depend on the support of elected officials (25). The challenge of evaluating projects that were generated initially for political reasons was noted. Business motivations were also seen as influencing factors, with politically connected business leaders seen as generating sup- port for toll projects that might not otherwise have been considered (25). Group Authority/Facility Characteristics Performance Explanation 3. Developed corridors Five facilities: • Harris County Toll Road Authority (TX)/Hardy • Harris County Toll Road Authority (TX)/Sam Houston • Transportation Corridor Agencies (CA)/Foothill Eastern • Transportation Corridor Agencies (CA)/San Joaquin Hills • Santa Rosa Bay Bridge Authority (FL)/Garcon Point Bridge • Corridors with more developed or already established traffic patterns • Usually constructed in large metropolitan areas or active tourist areas • “Solid” projected time savings • Moderate projected revenue growth Mean ranged between 51% and 60% of forecasts, on average, with considerable variation • Impacts of nearby non-tolled alternatives underestimated • Overestimated time savings • Overly optimistic economic forecasts • Failure to account for recessions • Overestimated corridor growth rates • High toll rates • Limited history of toll use in area • Unusual ramp-up problems • Expansion of competing non-tolled network 4. Least developed Eight facilities: • E-470 Public Highway Authority (CO)/E-470 • Toll Road Investment Partnership (VA)/Dulles Greenway • Osceola County (FL)/ Osceola County Parkway • Orlando–Orange Expressway Authority (FL)/Central Florida Greenway South Segment • Orlando–Orange Expressway Authority (FL)/SR-417 • Florida’s Turnpike Enterprise/Sawgrass Expressway • Pocahontas Parkway Association (VA)/ Pocahontas Parkway • Connector 2000 Association (SC)/ Greenville Connector • Specific traffic generator serving as project basis (e.g., airport) • Located in undeveloped area • Toll road expected to stimulate development • High revenue growth rates • Assumed periodic toll rate increases Mean ranged between 29% and 51% of forecasts, on average, with considerable variation • Insufficient existing traffic congestion • Overestimated time savings or value of time • High ramp-up growth rates, due to overestimated base period usage • High subsequent growth rates Source: Muller and Buono (3). TABLE 2 PERFORMANCE BY CATEGORY (Continued)

25 Another exogenous event was the development of competing routes or the failure to anticipate network improvements such as feeder roads or highway inter- changes. In some situations, noncompetition agreements have been developed that specify that other government agencies will not build competing facilities within a cer- tain protected geographic area. However, the agreements have not always been implemented. Survey respondents cited several exogenous factors that influenced the performance of the forecasts. All of these concerned the actual conditions under which the facility operated or was implemented. These factors included the actual operations and system reliability (e.g., actual congestion levels, operating speeds, and incidents); impact of the tolling technology on actual (recorded) traffic volumes (e.g., owing to unreadable license plates); violation rate; staging of the facility (or of other facilities); and changes in policy, mandate, legislation, ownership, etc. 4. Model error, which reflected the inherent variability in models regardless of how well the model was cali- brated and validated [i.e., the forecasts can never repli- cate the (eventual) actual traffic]. Although a model’s average error might be small, the average “may mask a problem, which when compounded within the model and over time, may severely skew results. This is an issue that is not discussed in an adequate level of detail in traffic and revenue reports.” The analysis further noted that the “simultaneous manifestations” of two or more of these problems contributed further to the poor model performance, with the forecasts “[amplifying] the negative variance between projected and actual traffic levels.” Finally, the survey of practitioners found that no single modeling factor influenced the performance of respondents’ forecasts. Respondents cited as factors the model structure; the process used to expand the modeled time periods to annual forecasts; the calibration process, coverage, and pre- cision; “control” over how the model outputs were used, ana- lyzed, or interpreted; the lack of transparency/opacity in the modeling and forecasting process; and the validity (i.e., appropriateness) of the model for financing purposes. TREATMENT OF SPECIFIC FACTORS AFFECTING FORECAST PERFORMANCE Drawing on the preceding discussion of the performance of the forecasts and the underlying reasons, this section exam- ines the treatment of specific factors that were identified as part of the scope of the synthesis, in the literature, and by practitioners. Demographic and Socioeconomic Inputs There are two relevant issues. The first concerns the use of long-range demographic and socioeconomic forecasts (so- called land use inputs to the model) that may reflect an MPO’s planning policy (i.e., as the source for these inputs) as opposed to market trends. Recent toll road demand and rev- enue forecasts have responded to these concerns by modify- ing these assumptions to account for input scenarios that were more conservative and that took into account historical trends and a more realistic assessment of likely future growth (5). An example is provided by a recent (2003) traffic and rev- enue forecast for the Transportation Corridor System (the Foothill/Eastern Transportation Corridor and the San Joaquin Hills Transportation Corridor), in which the local MPO land use forecasts were reviewed and refined in several ways: an update to the forecasts according to actual devel- opment that had occurred in the 5–6 years since the MPOs had prepared them; a review of job and household growth rates according to a variety of national, state, and regional third party sources; interviews with developers, realtors, and other related interests to identify issues that would affect future development and the regulatory environment in the study area; detailed field studies of 50 “focus areas” to iden- tify current and potential development capacity and con- straints to development; and the identification of candidate areas for redevelopment and infill development at higher (than originally forecasted) rates in the long term. Forecasts for different categories of employment were revised accord- ing to recent trends, and forecasts for residential develop- ment accounted for such variables as recent changes in prices. Overall, revised short- and long-term land use fore- casts were developed (50). The impact of the refined land use forecasts is not yet clear, given the recentness of the study. Although the actual revenue growth rates for 2003–2004 and 2004–2005 were greater than the projections (8.7% versus 4% and 9.7% versus 4%, respectively), a July 2004 increase in toll rates might have affected the results (51). The second issue is the lack of consideration of the impact of short-term economic fluctuations on travel demand. The impact of optimistic economic projections on traffic projec- tions was noted in several studies. The national recession of 1990–1991 affected the use of the first two segments of the Central Florida Greenway, which had opened in 1989 with first-year projections just slightly below actual, but with poorer results for the next two years (over the course of the recession). A “drag” from the recession was considered to have affected toll roads in Oklahoma City and Tulsa, Okla- homa, which opened just after the recession. Local economic impacts, such as the collapse in oil prices and the subsequent sharp regional economic downturn of 1986, left economic growth in the Houston area well below projections, with cor- responding impacts on the Hardy and Sam Houston toll road revenues. Even when regional economic activity was close to the original projections, the performance of some tolled facil- ities still fell short, because economic activity within the immediate corridor did not meet projections (e.g., the Saw- grass Expressway in Florida) or the expected build-out of res- idential areas was slower than expected (e.g., the Seminole

Expressway toll road, also in Florida) (6). Practitioners have begun to consider the impact of short-term economic changes. The aforementioned Transportation Corridor System forecast took into account a “recession scenario,” which considered a “double dip” of below average job gains in the immediate term, followed by job losses for the next two years, then by a modest recovery and a recessionary dip in the seventh year. These inputs were used as part of a sensitivity test of the demand and revenue forecasts (50). Another observer com- mented that “supply-driven” land use forecasts (meaning forecasts that take into account factors such as growth in the labor force, demographics, and productivity) provided more stable results than did “demand-driven” inputs (such as fore- casts of population and jobs). The pending retirement of the “baby boom” generation was also seen to have an impact on the demand for travel. As an example, this observer cited the 1999 Foothill Eastern refinancing study, which preceded the aforementioned Transportation Corridor System study. This study accounted for a more stagnant labor pool after 2010, which in turn generated “far less” growth in the long-term traffic and revenue forecasts (3). A related issue concerns the ability to understand the travel characteristics of the users of a proposed facility. With reference to improving the performance of transit rid- ership forecasts, one observer proposed bringing the fore- casting horizon closer to the present, which “would reduce the range of developments that can cause projections to go awry, such as changes in the local economy or evolution of travel patterns in response to geographic redistributions of employment and population.” An “extreme variant” would be to predict ridership under current demographic and travel conditions, “which would isolate the increased ridership attributable to improved transit service from that owing to demographically induced growth in overall travel demand.” This “opening-day” ridership would be used as the basis for the evaluation of alternatives, rather than long-range fore- casts (35). Although the analytical horizon for toll road demand and revenue forecasts clearly cannot be shortened, a current-year or very-short-term toll demand forecast based on a hypothetical immediate opening of the facility would allow analysts and users to differentiate the demand that would result from the network improvement and that would result from assumed demographic or economic growth. Fur- ther analyses could test the impact of the facility, with and without tolls, again in the short-term, to isolate the impacts of tolls. In other words, although these short-term forecasts might have limited use in the development of absolute esti- mates of revenues, they would be valuable in grounding and interpreting the long-range forecasts (i.e., to provide a refer- ence against which to compare that proportion of forecasted long-term facility traffic that would use the facility whether or not the toll is in place or independent of assumed growth). Travel Characteristics The availability of appropriate data and the quality of these data were noted in the literature and in the survey as one of 26 the major sources of potential forecasting inaccuracies (25). These data include such variables as traffic counts, network characteristics, travel costs, land use, and employment. Inap- propriate base year data can result in model validation errors, which in turn affect all subsequent applications and forecasts. In practice, these data, which are the foundation of the fore- casts, were found to be subject to substantial numbers of mea- surement and processing errors (25). Similar problems were found with forecast inputs such as land use: The model may accurately reflect the assumed or calculated inputs; however, if the assumptions are erroneous, then the accuracy of the forecast will suffer (52). Current data collection practices are well-established. They include origin–destination surveys, trip diaries, activity- based surveys, stated preference surveys, traffic counts, travel time data, and speed surveys. Surveys of existing socioeco- nomic and transportation system characteristics are required for calibration. However, one observer noted that what was once a standard part of transportation planning is not usually undertaken to the same degree in contemporary planning studies (25). The Minnesota DOT model update noted earlier used two previous studies (one local and one from California) as sources of information, given the lack of data (26). Another observer noted the need for caution in the use of “imported” data to address gaps: “without great care and con- siderable experience, significant errors can be introduced into the modelling framework through inappropriate impor- tation of model parameters.” Other factors relating to data included the role of uncertainties and potential sources of error introduced by sampling [in surveys]. The appropriate categorization of travel markets in terms of their individual values of time and willingness to pay was also noted, with income levels and time sensitivity (i.e., trip purpose) being important determinants. The ability to save time was the most important determinant of whether or not a private auto- mobile driver chooses to use a toll road, whereas truck driv- ers also took into account the impact on vehicle operating costs (i.e., that a toll road’s “competitive advantage” for trucks must be measured both in terms of time savings and the ability to save on fuel costs and reduce vehicle wear and tear). Similarly, the importance of “who pays” also was noted, as was the difficulty in modeling this influence. Finally, assumptions regarding growth in vehicle ownership (also related to growth in Gross Domestic Product and income) were noted as influences on traffic demand in general (53). These findings generally were corroborated by the sur- vey of practitioners. Respondents cited the values used for value of time, willingness to pay, and other monetary val- ues as influences on the performance of the forecasts. Other influences included assumptions regarding land use fore- casts or future network configurations, with some respon- dents distinguishing public and political influences in these assumptions; availability, appropriateness, or sufficiency of the data, models, or analytical processes; environmental or economic development considerations; and economic cli-

27 mate. Overall, respondents recommended collecting more or better data to improve travel demand forecasting results. Value of Time and Willingness to Pay The treatment of the ability and willingness of potential users to pay was cited as a key performance factor both in the lit- erature and by practitioners. Values of time can be differen- tiated by purpose, mode, and/or vehicle class. Willingness to pay is a variation of value of time that accounts for how much travelers value different attributes of the toll facility, such as safety and reliability. The valuation of travel time is based on two underlying principles (54). The first principle states that time is valuable because people can associate it directly with results, such as making money or participating in a leisure activity—that is, the time spent in travel could be devoted instead to other activities. The second principle assumes that time can have an additional cost over and above that associated with the first principle; for example, travelers might find it undesir- able to have to walk, wait for transit, travel on a crowded bus, or drive in congested conditions. As a result, “the value of saving time may vary, depending on both the purpose of travel, which affects the possible alternative uses of time, and the conditions under which it occurs.” The measurement of the perceived value of a driver’s travel time yields the value of time. This influences a driver’s decision to use a toll road. Values of time vary from region to region, and what is assumed for one forecast may not be transferable to another forecast. The value of time is a func- tion of a driver’s purpose (where work trips are more valu- able than discretionary trips), income, and personality. The value of time is used to convert the monetary toll to time. This allows the monetary value to be incorporated into the model’s generalized cost function. As described earlier in Methods for Modeling Toll Road Demand, this is incorpo- rated into the calculation of route diversion (within or post- trip assignment), which in turn may be fed back to other parts of the modeling process. Two-thirds of the models reported in the survey of practitioners incorporated value of time, 10% used willingness to pay, and another 10% used both. One model incorporated travel time and travel cost in its mode choice utility equations. The choice and derivation of the values used for this deter- mination are the subject of considerable debate in the litera- ture. The U.S.DOT has developed a guidance document on the subject. Its purpose was to establish “consistent proce- dures” for use by the department in its evaluation of travel time changes that would result from transportation invest- ments or regulatory actions. The guidance stated clearly that locally derived data should be used to forecast demand on individual facilities (54). The guidance reviewed the factors that are associated with the value of travel time. For trips made during work or when the traveler could vary his or her work hours, the guidance noted that “the wage paid for the productive work that is sac- rificed to travel” could be used to represent value of time. The value of time for other (personal) purposes can be rep- resented by some fraction of the wage rate. Thus, the hourly income (before-tax wage rates, including fringe benefits) could be used as a “standard against which their estimated value of time is measured.” As well, higher income has been associated with higher values of time, meaning that toll roads that operate in higher-income areas should experience greater patronage (and support higher toll rates) (3). The guidance developed tables that expressed the values of time (and “plausible ranges”) as percentages of hourly incomes, categorized by local and intercity travel, business (work- related) or personal trip purposes, and mode (surface modes taken together and air travel). A separate category was devel- oped for truck drivers. A 2003 revision retained the method, but updated the actual hourly incomes (55). For toll road demand and revenue forecasts, the value of time generally has been assumed as a single value to represent an average characteristic for a given study area (25). How- ever, researchers have recognized that the use of an average value of time masks the heterogeneity among travelers, notwithstanding the existing categorizations of time values by purpose. Recent research examined the preferences of users of the SR-91 toll lanes (California) by analyzing different revealed and stated preference survey data sets. The research concluded that values of time and the value of reliability were high, although the values dropped for very long distances. However, these values contained considerable heterogeneity: to this end, the research examined the impacts of the time of day at which the trip was made, flexibility of arrival time, gen- der, age, household size, occupation, marital status, and edu- cation. The research highlighted differences in the data; for example, drivers with higher incomes were more responsive to the toll, according to the revealed preference data (i.e., according to how they actually behaved), but not according to the stated preference data. Finally, the data were used to develop a pricing policy model, which took into account the utilities calculated for the individual factors (as a means of illustrating the importance of accounting for heterogeneity in estimating the value of time) (56). Other researchers identified problems in the application or development of values of time, as compared with the findings of empirical studies of what travelers were actually willing to pay. These included (57): • Trip assignment models that simulated route choice (i.e., diversion under tolls) using travel time values that had been derived from empirical studies of mode choice. • Application of values of time to choices whose attri- butes were quite different from those that were used to calculate the value of time (e.g., comfort, convenience, and status). • Relationships between values of time and other influenc- ing variables (notably, income), which were assumed to

develop over time in ways that were inconsistent with other evidence. [Other empirical studies have suggested that the value of time increases over time, but not pro- portionately with income levels (57)]. • Values of time that were calculated from stated prefer- ence methods, which must be based on very-short-term (i.e., immediate) preference structures, but whose resul- tant values are applied to models that reflect behavior that takes some years to evolve. The researchers found that in cases where an average value of time represented a skewed distribution of values, there was a tendency to overestimate the revenue, and under- estimate the impact of a toll, because for a given mean value of time (i.e., for the value of time that was used in the demand and revenue forecast) there was a smaller number of indi- viduals who were prepared to pay the toll. To address this, the researchers recommended the establishment of a relevant set of purpose-specific, time value distributions; determining a way to address these distributions in forecasting demand; “growing” the values over time; accounting for the time values of automobile passengers (in addition to those of the driver); and establishing methods to convert disaggregated (heterogeneous) values of time into a single trip value that is appropriate to the specific project under consideration. The impact of tolls on shifts in driver behavior over time is beginning to be understood, as a history of toll facilities develops from which both static conditions and changes over time can be observed. In addition to addressing the relative paucity of data on the subject (given the relative newness of tolled facilities), historical data are important because driver behavior can change over time: this, in turn, affects the fore- casts that commonly retain base year rates, modeled relation- ships, and parameter values through all horizon years. This was demonstrated by recent research on the Lee County, Florida, variable pricing program, which found that shifts in traffic volumes varied over time, with the long-run (3–5 years) relative elasticity of demand being lower than that of the short run (1 year or less). The research also found that sev- eral demographic, socioeconomic, and travel behavior factors affected facility use over time. Drivers who were commuting, were full-time employees, had more persons in their house- holds, had a postgraduate degree, and were between 25 and 34 years old were found to be more likely to have increased their use of the facility. Retired drivers and those with lower incomes were less likely to use the facility (58). Two of the studies cited previously also commented on the use of stated preference surveys as the basis of valuating time (57 and C. Russell, personal communication, Sep. 20, 2005). Stated preference surveys attempt to measure the value of time by presenting hypothetical options to respondents to quantify how toll rates would affect driver behavior. They are widely used in toll demand and revenue forecasts. Stated pref- erence surveys for toll demand forecasting are generally in 28 three parts: (1) background information on a recent trip in the study corridor, (2) a set of stated preference experiments, and (3) demographic information. The background information provides revealed preference data about an actual trip, as well as baseline data to customize the stated preference scenarios. The variables of interest are determined and ordered into a series of scenarios that is presented to respondents as part of the experiments: the scenarios are designed so as to allow the subsequent estimation of the respondents’ relative prefer- ences for each of the tested variables. Diversion (multinomial logit) models and values of time in turn are calculated from these estimates. Travel time, toll cost, and income typically are included as attributes, with values of time calculated by trip purpose (work and non-work) and separately for automo- biles and trucks. Stated preference surveys in areas that do not have tolled facilities have tended to result in low values of time, because respondents express their “anti-toll road senti- ment.” In areas that have existing toll roads and severe peak- period congestion, respondents have tended to overestimate their values of time. In either case, the calculated values of time may have to be recalibrated to reflect actual conditions more reasonably. The availability of electronic toll collection (as opposed to cash collection) may also influence the value of time, in that electronic toll collection users may be less aware of and, therefore, less sensitive to the total toll paid on a trip (2), at least in the short term. Another practitioner has noted that the importance of understanding what (average) values of time derived from stated preference surveys repre- sent; namely, they are proxies for several attributes (comfort, safety, convenience, reliability, etc.) and that all the models derived from these data assume that the respondent has per- fect knowledge at the time of his or her decision making (C. Russell, personal communication, Sep. 20, 2005). The need for accurate ordering of preferences in the sce- narios has been noted by some researchers, who examined the applicability of stated preference surveys, and the methods used to estimate values of time from them, for toll demand analysis. The research surveyed a nationwide sample of respondents on their preferences of 13 alternatives that described the “essential elements of a commute,” including congested and uncongested travel times, travel cost (usually, a toll), and an indication of whether or not trucks were allowed on the road. The researchers then used the resultant preferences to calibrate different formulations of the diver- sion model. They found that the ordering of the preference scenarios was important and that the choice of model formu- lation gave “very different ratings” for the same data. Finally, the research found that the willingness to pay estimates were relatively low and did not vary much among drivers, which implied “that the average commuter does not appear willing to pay much to reduce automobile travel time.” The research noted that the distribution of willingness to pay was fairly lim- ited, meaning that few travelers (in the sample) were willing to pay considerably more than the average toll. The research concluded that “extreme caution should be used in estimating stated preferences based upon respondents’ ratings” (59).

29 Finally, one study noted the “growing body of empirical evidence that travelers value reliability as an important factor in their tripmaking decision.” Reliability generally reflects the day-to-day variability in expected journey times, owing to nonrecurrent congestion such as incidents, weather, construc- tion, and so on. Reliability is considered important in variable- priced HOT applications, where tolls are adjusted according to traffic volumes, to maintain a specified level of service. Therefore, average travel times on HOT lanes may be only slightly reduced from those on non-toll lanes; however, the day-to-day fluctuations in travel time variability are reduced significantly. Reliability can be critical for travelers with fixed schedules (such as individuals with daycare pick-ups or those going to the airport), and is not necessarily correlated with the traveler’s general value of time. However, there remain few (if any) operational demand models that account for reliabil- ity in traveler values of time or that measure the value of reli- ability, because of a general lack of data: this reflects the few examples of operational toll roads using variable pricing from which empirical data can be drawn (29). Tolling “Culture” An international review of toll road traffic and revenue fore- casts demonstrated better performance for countries that had a “history” of toll roads, compared with those for which road tolling was new. In “countries with a history of tolling, con- sumers can be observed making choices about route selec- tion, effectively trading off the advantages against the costs of using tolled highway facilities. The consumer response can therefore be more readily understood by forecasters preparing predictions for new or extended facility use.” In contrast, in countries where tolling is new, there are no revealed preference data on consumer behavior, which “leaves forecasters more reliant on theoretical survey tech- niques and assumptions about how drivers may respond to tolls” (60). The existence of a tolling “culture” also affected the ramp-up forecasts, with there apparently being “little transferability of experience between projects (particularly those in different countries). Ramp-up tends to be project- specific” (53). Other influences on traveler choice included the conve- niences that toll roads offered or were seen to offer. Improved safety was cited as an influence that would attract drivers to a “perceived” safer highway, such as one with fewer trucks (61). Commuters also valued reliability, and travelers may be willing to pay for predictability of travel time, especially for time-sensitive activities. The type of toll collection was cited as a third influence on drivers’ route choice (61). Truck Forecasts The treatment of truck forecasts was the subject of an analysis that found that the variability associated with such forecasts was consistently higher than that for light vehicles. Although truck traffic typically comprises a relatively small portion of the traffic mix, they commonly pay two to five times and sometimes as high as 10 times the respective car tariff and so their contribution to total revenues can be significant. The choice by trucking firms to use toll roads was also found to depend on, among other factors, the size of the firm, with inde- pendent owner–operators (that dominate some trucking sec- tors) being “very sensitive” to tolls (49). Ramp-Up The ramp-up period reflects a toll facility’s traffic perfor- mance during its early years of operation. This period may be characterized by unusually high traffic growth. The end of the ramp-up period is marked by annual growth figures that have (or appear to have) stabilized and that are closer to traf- fic patterns that have been observed on other, similar facili- ties. The ramp-up period reflects the users’ unfamiliarity with a new highway and its benefits (“information lag”), as well as a community’s reluctance to pay tolls (if there is no prior tolling culture) or to pay high tolls (if there is a history). The performance of the facility during ramp-up is particularly important to the financial community, because the “proba- bility of default is typically at its highest during the early project years” (9). As the initial year forecasts in Table 1 indicated, the ramp- up performance has been problematic and inconsistent. One analyst noted that the consideration of ramp-up forecasts had three dimensions (9): • Scale of ramp-up (i.e., the magnitude of difference between actual and forecasted traffic). • Duration of ramp-up (from opening day to beyond 5 years). • Extent of catch-up [i.e., having experienced lower-than- projected usage at the time of the facility’s opening, to what extent could observed traffic volumes catch up with later year forecasts? The significance of catch-up volumes, according to one analyst, was that “projects with lower-than-forecasted traffic during the first year operations also tend to have lower-than-forecasted traf- fic in later years (referring to toll and non-tolled roads, as well as transit and inter-urban rail) (62)]. In other words, the conditions that define ramp-up were spe- cific to each project, given that the factors that influenced ramp-up also varied (e.g., signage and marketing, the tolling culture, and the availability of competing free routes). Another analyst noted that until recently ramp-up has largely been ignored. Subsequent efforts have considered ramp-up; however, these have tended to be based on the use of other facilities as proxies. However, as more facilities have opened, actual operations and influencing factors could

be observed. An inverse relationship between time savings and ramp-up has been observed, such that greater time sav- ings appeared to correlate to a shorter ramp-up (5). As an example, another analyst noted that the estimates of time sav- ings for the Hardy and Sam Houston toll roads in the Hous- ton, Texas, area were very different, with the Hardy saving less than 5 min per average trip and the Sam Houston saving more than 10 min. “At some level, this would indicate that the basic need for the Sam Houston was probably greater” (which was demonstrated by the significantly better perfor- mance of the Sam Houston toll road; see Table 1). Also, income levels along the Sam Houston corridor generally were higher than those in the Hardy corridor (meaning that the value of time would be higher for the former) (6). In the absence of extensive observations, however, one analyst proposed the use of “revenue-adjustment” factors for ramp-up forecasts. These varied according to the expected duration of the ramp-up period, the extent to which the ini- tial observed volumes would have to catch up (i.e., whether the initial volumes were less than projections, and by how much), and the source of the toll demand and revenue pro- jections (meaning that those commissioned by banks appeared to be more accurate than those of others, such as the sponsor; with the former group “typically referred to as ‘con- servative’ forecasts” and used as the base case in financial analyses). The factors were derived from a comparison of the performance of 32 tolled facilities around the world. The fac- tors ranged from a 10% reduction to first-year revenues for facilities with a two-year ramp-up and no catch-up required, according to a bank-commissioned forecast, to a 55% reduc- tion to first year revenues for facilities with an eight-year ramp-up and initial volumes 20% less than projected, accord- ing to forecasts commissioned by others (9). However, it should be noted that although they mask considerable varia- tion, the averages of the actual versus projected results listed in Table 1 improve over time, from 59% in the first year to 70% in the fifth year. An alternative treatment was to recognize the potential for an extended ramp-up by accounting, in the forecast, for the market served by the facility; the duration was assumed to be extended if the facility was “development dependent” and could also be shortened if it was in a built-up area. The start- ing point (base forecast) also could be reduced according to experience on other nearby facilities. This allowed the “inability to accurately predict a key factor [to be] balanced by very conservative assumptions” (5). Time Choice Modeling The need for models that more accurately capture the dif- ferences in travel patterns by time of day, day of week, and even season was identified by the financial community. The object is twofold: first, to account more explicitly for the temporal variation in composition of trip purposes, origins 30 and destinations, and vehicle types, including (in addition to peak-hour travel) the off-peak, midday, night, and weekend (5). This would replace the common use of factors to expand peak-hour models to daily and then annual traffic volumes for purposes of revenue forecasting. It also would replace the use of 24-h models, from which estimates of hourly traf- fic volumes must be derived (typically using factors) as the basis for forecasting diversion under tolls; some MPO mod- els, which are used as the basis for toll road demand and rev- enue forecasts, simulate only daily traffic; see, for example, “Tyler Loop 49—Level 2 Intermediate Traffic and Toll Revenue Study” (63). Research on the impacts of variable pricing in Lee County, Florida, found that shifts in trip start times changed over time; that is, long-run elasticities of demand were lower than short-run elasticities (58). The second object is to add and integrate time-of-day choice with mode and route choice. The Florida Turnpike Enterprise provides an example of how time-of-day choice was integrated with modal and route choice. It has relied on what it considered best practice in its toll-forecasting proce- dures for periodic updates of traffic and revenue forecasts, as well as for the planning, design, and economic feasibility assessment of proposed new facilities (33). A basic nested logit approach was used to describe travel behavior for mode, route, and time of day, with the following choices (33): • Mode: automobile, drive alone; automobile, two occu- pants; automobile three plus occupants; bus; and rail. • Route: tolled or non-tolled roads. • Time of day: desired travel time or time-shifted trip. Sixteen statistically estimated nested modal choice mod- els were developed from survey data. These models incor- porated four time periods and four trip purposes. The models also included specific decision tree hierarchies for transit and occupancy classes. Four specific time periods were modeled: a.m. peak, p.m. peak, midday, and nighttime. Risk The need to address risk and uncertainty more comprehen- sively was cited by the financial community and by researchers. At the same time, the survey of practitioners indicated that only a small number of respondents conducted a risk assessment, with most of the remainder verifying their results through judgment or reality checks and others not doing any verification. Many assumptions and variables must be interpreted and relied on to complete a traffic and revenue study. The ability to ensure exactness and accuracy in all of these is limited for representations of existing conditions as well as forecasts. A common treatment has been to address uncertainty through the simple use of conservative assumptions or ranges. How- ever, this was not always possible, as indicated by some

31 survey respondents, who noted the impact on the accuracy of forecasts of exogenous factors such as public or political inputs, land use, and network assumptions. For example, when uncertainties existed about whether a particular com- peting road would be constructed, it had been the practice to conservatively assume the competing road would be com- pleted if it was expected to have a negative impact on the toll facility. If it was not expected to affect the toll project, it was then assumed not to be in place (61). The literature indicated that sensitivity analyses on key variables was common practice, such as the area growth rate, value of travel time, planned toll rates, and other variables that were region-specific or that had shown a high degree of variability in the past. The SR-520 Toll Feasibility Study for the Washington State DOT provides an example. The pur- pose of the toll feasibility study was to determine the revenue potential and traffic impacts of tolling a replacement bridge on SR-520. To quantify uncertainty, two tolling objectives were modeled to “bookend” the upper and lower bound of the reasonable toll possibilities within the corridor (64). However, respondents to the survey of practitioners recog- nized that sensitivity analysis might not be sufficient, because just over half of the respondents stated the need to conduct more risk assessment in the forecast process. A risk analysis process can evaluate thousands of different scenar- ios to quantify the probability of a “range of potential out- comes” (65). In other words, it is important, first, to ensure that risk assessment is incorporated explicitly into the forecasting process and, second, to make the distinction between the assessment of risk and other indirect treatments of uncer- tainty (such as judgment or sensitivity analysis). The first is related, in part, to the inclusion of an appropriate and com- plete set of assumptions and inputs, and in part to ensuring that model makers and users of the model outputs all under- stand the implications of alternative choices or influences (whether qualitatively or quantitatively). The second requires a proper understanding of the roles of the different treatments of uncertainty. For example, some financial analysts have commented that sensitivity analysis does not adequately reveal the range of possible outcomes in a toll road forecast (65). Instead, a range of possible out- comes could be explored, based on Monte Carlo simulation and the probability not only of the variables acting as indi- vidual occurrences but in combination with each other based on their respective probability of occurrence (65). Another treatment is offered through “reference class forecasting,” which uses the experiences of past projects to help statisti- cally identify the probability of given inputs occurring at a particular value (36). Toll road demand and revenue forecasts have given little or no consideration to the possibility of a series of events occurring simultaneously (65)—for example, if economic growth recedes, oil prices spike, and a large development that was scheduled to be in place at the time of the opening of the toll road is cancelled. Traditional sensitivity analysis typi- cally took each of these assumptions and varied them one at a time; however, these assumptions often varied by arbitrary amounts. A further problem was that in reality, rarely, if ever, did these assumptions vary from actual outcomes one at a time (65). Financial analysts have noted that the simultaneous occur- rence of several vulnerabilities contributed to toll roads gen- erating lower-than-expected traffic levels and, accordingly, toll revenues (5). Although risk analysis has been incorpo- rated into some analyses, one analyst noted that many project uncertainties were external to the traffic model environment. Because these uncertainties may not be fully captured in the model probability analysis, the model outputs must also be interpreted within the framework of the risk analysis (60). The National Federation of Municipal Analysts (NFMA) is comprised mainly of research analysts, who are responsible for evaluating credit and other risks with respect to municipal securities. NFMA has worked with nonanalyst professionals in various sectors to develop recommended best practice guidelines for certain markets, including the toll road demand and revenue forecasts. In these guidelines, NFMA has attempted to account for the likelihood that there are many possible outcomes if future events do not follow the projected assumptions that are predicted in the model. Given the large number of input variables required in the modeling process, NFMA found that the results of forecasts can be significantly influenced by changes in these inputs (65). Simplistically, by applying the appropriate background data inputs to the toll forecast, a model could produce a traffic and revenue forecast that is most likely to occur, often called a base scenario. Whereas a single best statistical estimate may be desired by some, there are limitations to a single expected outcome. A proposed mitigation strategy is the assignment of a probability distribution to all inputs, or at least to those inputs identified as most influential to the process. Each individual variable, with its own probability distribution, can be fluctu- ated simultaneously. This capability supports a better approximation of reality then can be obtained, because in practice variables do not generally change one at a time but concurrently with varying rates of change (65). NFMA’s best practices guidelines developed a disclosure requirements list with respect to creating a range of possible outcomes for traffic and revenue studies. The list included the following: • Creation of a no-build traffic forecast (including truck and congestion analysis) for the study area, without the toll road. • Creation of a baseline traffic and revenue forecast (as per standard practice).

• Sensitivity analysis while simultaneously varying toll road inputs simultaneously (the following list is a guide only, but it should be used as the minimum standard): – Population growth, – Employment growth, – Personal income growth, – Toll elasticity by consumers, and – Acceleration of planned transportation network. • Debt service analysis with toll road project sensitivity analysis. A risk analysis using the Monte Carlo technique was applied to several factors that were used to develop traffic and revenue forecasts for a proposed toll road linking Hong Kong with nearby industrial areas in southern China. The risk analysis found that whereas variations in the forecasts of population had insignificant impacts on the traffic and rev- enue forecasts, the impacts of variability in the trip genera- tion were very large, with standard deviations of the forecasts being of the order of double (or more) than the base forecasts. Variations in the estimated diversion rates to the facility and in the toll rate also were found to be significant. From these findings, the authors pointed out that the impacts of variabil- ity and uncertainty in these factors can influence the traffic and revenue forecasts, and noted that “if the effects of varied key assumptions and scenario options are not examined, the real optimal rate of return could be missed” (66). One risk analysis consultant noted that the determination of risk should not focus on a single outcome but should explore a range of possible outcomes (D. Bruce, personal communication, March 4, 2005). This process first deter- mines the degree of risk in each input variable by developing probability distributions for all variables. Risk analysis is carried out by allowing all the underlying variable estimates to vary simultaneously, which can be done using simulation techniques such as the Monte Carlo technique. The risk and uncertainty in the underlying input variable is then translated into a probabilistic, risk-adjusted forecast of output variables such as traffic levels, toll rate, revenue, and debt service cov- erage. Finally, the variables that drive risk, that is to say the variables that have the greatest influence on the forecast, are identified. Bias The existence of “optimism bias” in transportation projects has been noted by some observers. “It is in the planning of such new efforts [referring to projects in a city ‘for the first time,’ where ‘none existed before’] that the bias toward opti- mism and strategic misrepresentation are likely to be largest” (36). Another analysis noted the influence of bias on the per- formance of forecasts, with “systematic optimism bias” cited as a “distinguishing feature of toll road forecasts.” The analy- sis recommended that “base case forecasts should be adjusted to take account of any suspected optimism bias.” 32 Bias was differentiated from “general error” in modeling, with the performance of traffic forecasts for tolled and non- tolled roads generally being “very similar.” The update also suggested the need to consider distributions in error (67). On the other hand, in a project for which multiple bids are assessed competitively, the forecasts that are used as the basis of financing, and which are available to the financial community, are those of the winner: the forecasts for the unsuccessful bids generally will project lower revenues (D. Johnston, personal communication, Aug. 17, 2005). The literature review uncovered no formal methods or guidance to address optimism bias in toll road demand and revenue forecasts. However, the British Department for Transport recently developed a guidance document on pro- cedures to address optimism bias in transportation planning (68). Although this document addressed the cost side (for both roads and public transport), its approach could inform any future discussion of optimism bias in demand forecast- ing. The guidance document noted that transportation proj- ects always must be considered “risky,” as a result of long planning horizons and complex interfaces. It was theorized that optimism bias resulted from a combination of the struc- ture of the decision-making process and how the decision makers were involved in the process. Optimism bias was not an unknown or imaginary phenomenon; rather, it was a log- ical product of the participants involved, their interests, the framework for conditions for funding, and the resulting incentive structure they encounter. The causes of optimism bias were grouped into four cat- egories: technical, psychological, economic, and political. Technical causes included the long-range nature of the plan- ning horizons and that often the project scope and ambition level can change during the development or implementation of the project. Psychological causes were explained by a bias in the mental make-up of the project promoters and forecasters who can all have reasons to be overly optimistic in the approval stage—for example, engineers want to see things built and local transportation officials are keen to see projects realized. Economic causes were demonstrated by the argument that if the project went forward and was imple- mented, then more work was created for the industry; and if the participants were involved directly or indirectly, there could be an affecting influence. Finally, political causes were seen as influencing the perceived optimism bias in terms of interests, power, and the prevailing institutional setting that surrounded decision making on transportation projects. Project appraisers were seen to have demonstrated a ten- dency to be overly optimistic. To address this bias, apprais- ers should make explicit, empirically based adjustments to the estimates of a project’s costs, benefits, and duration. These adjustments should be based on data from similar past projects and adjusted for the unique characteristics of the project.

33 Optimism bias uplifts were introduced as methods of combating optimism bias in the decision-making process. They were established as a function of the level of risk that the British Department for Transport was willing to accept regarding cost overruns in transportation projects. The gen- eral principle was that the lower the level of acceptable risk, the higher the required uplift. The optimism bias uplifts should be applied to estimated budgets at the time of the deci- sion to build. To minimize optimism bias, preliminary findings indi- cated that formal and informal rules aimed at changing the established culture should be applied. The following four benefits to applying optimism bias uplifts to budgets were identified: • Emphasis on establishing realistic budgeting as an ideal while eliminating the practice of overoptimistic bud- geting as a routine. • Introduction of fiscal incentives against cost overruns. • Formalized requirements for high-quality cost and risk assessment at the business case stage. • Introduction of an independent appraisal process. Model Validation Model validation and model calibration are not the same: Calibration demonstrates how the model (and its individual components) replicates observed historical data, whereas validation proposes to demonstrate how “reasonably” the model’s functional forms and parameters predict actual observed behavior. It was noted earlier (Explanation of Per- formance) that the inherent model error in toll road demand forecasts was not addressed, and often masked, by existing model validation that demonstrated base year forecasts “near 100% to actual traffic” (5). A thorough evaluation of toll road demand and revenue forecasts was found to com- prise three steps: technical quality and merits of the fore- casting process, interpretation and professional judgment on forecast results, and risk analysis and sensitivity analysis on key input variables (25). A “major criticism of transportation demand models is the general lack of concern for, and effort put into, the validation phase of [model development].” Time, budget, and data con- straints in “typical” practice contribute to this lack of concern; however, “the improvement in predictive capabilities of transportation demand models and in the credibility of these models with decision makers, rests to a large extent on the analyst’s ability to validate the procedures used.” Three gen- eral approaches to model validation are described here (12): • Reasonableness checks of parameters and coefficients— for example, checking whether or not a particular value is within an expected range or has the correct sign. The objective is threefold: to ensure that the model does not violate theoretical expectations and, if it does, to iden- tify the source of the error (the model or the expecta- tion); to ensure that the model does not exhibit any “pathological tendencies”; and to ensure that it is inter- nally consistent (meaning that its outputs do not violate any assumptions used to generate them). Reasonableness checks of the model outputs also can help the user determine whether the forecasts are reason- able. For example, the projected market for a proposed tolled facility should be verified to ensure that it is rea- sonable. Readily available techniques, such as a select link assignment, could be used. [This technique isolates which travelers (as measured by trip origins and destina- tions) are projected to use a particular facility. The tech- nique then could be used to determine how this subset of travelers would behave under changes in network configuration—for example, with or without tolls, or with or without the facility in place]. However, one con- sultant noted that many traffic and revenue studies do not always include this (simple) verification, or a demon- stration of the time advantages offered by the proposed facility (C. Russell, personal communication, 2005). • A rigorous test of a model’s predictive capabilities is provided by using the model to predict demand for a time period other than that used for the model cali- bration; this assumes the availability of (at least) a second set of observed data. The use of the model to predict some historical condition can be instructive; for example, in identifying the need to better account for unforeseen influences (such as conditions of eco- nomic stagnation). • When data for multiple time periods are not available, an alternate (but more restricted) test involves the random splitting of the “one-period” data into two sets. One set is used for calibrating the model, which is then used to predict the second set’s demand. This allows the model’s predictive capability to be validated against an indepen- dent set of data, although the validation is limited by the lack of temporal difference between the two sets. The temporal, budgetary, and data constraints posed to model validation were cited by another traffic and revenue consultant as fundamental reasons for inaccuracies in toll road demand and revenue forecasts. In particular, basic data were often old, incomplete, or unreliable because of limits to sample size; in turn, this implied that the models that were developed from these data were subject to “substantial error” (D. Johnston, personal communication, Aug. 17, 2005). Peer Reviews Peer reviews are processes in which external experts (i.e., in modeling) can provide technical guidance and advice to the proponent’s team during the course of the development of the data, models, and forecasts. Peer reviews have been a part of the toll demand and revenue forecasting process in the past,

but on a very limited basis and at a very small scale (5). Accordingly, the value of the peer review process has been somewhat limited to date, because the process must be con- tinuous to show any improvement or advancement of tech- niques. To benefit from this process, detailed reports must be prepared and independent meetings conducted with bond rat- ing agencies to discuss the extent of the review and the results achieved (5). The requirements vary for peer reviews for models in U.S. transportation planning practice. The approaches of FTA, FHWA, and TMIP are described here. • FTA has specific requirements and standard review procedures for forecasting as part of its New Starts discretionary grant program. These review procedures were implemented to enable FTA to evaluate and rate grant applications on an unchanging, unbiased level. Projects seeking New Starts funding must first pass a locally driven, multi-modal corridor planning process, which has three key phases; alternatives analysis, pre- liminary engineering, and final design. The list of evaluation criteria includes operating efficiencies (e.g., operating cost per passenger mile) and cost- effectiveness (e.g., the incremental cost per hour of transportation system user benefits) (69). • FHWA allocates funding on a formula basis; meaning that states and MPOs do not have to compete on a project-by-project basis for funding. As a result, the same degree of standardization is not required of forecasting procedures as in FTA’s New Starts program. However, FHWA does provide technical assistance to state DOTs and to MPOs in an attempt to ensure that the travel demand forecasts used are credible and based on proper planning practice. Three types of assistance are provided (B. Spears, personal communication, Aug. 18, 2005): – A checklist of questions that help FHWA, state DOT, and MPO staff understand what is required of the model. FHWA staff is encouraged to ask these ques- tions during the triennial certification reviews that are conducted with those MPOs with populations of more than 200,000. – The creation of a peer review team, whose responsi- bility is to review the forecasting process and make recommendations for improvement. These teams typically include from four to eight travel modelers from other agencies, MPOs, state DOTs, academia, or consulting firms, who meet with local planners for one or two days of review. – The revision of travel forecasts used in specific proj- ect documents, such as transportation plans, confor- mity determinations, and environmental studies. If a peer review is convened, then several pieces of information and documentation must be made available to the team, including an inventory of the current state of transportation in the area, key plan- 34 ning assumptions used in developing the forecasts, and descriptions of the methods used to develop fore- casts of future travel demand (70). • In addition to its core methodological research, TMIP reviewed ways to improve existing modeling processes. To this end, a peer review panel recommended several types of improvements to the practice of travel demand modeling, including the peer review process. The rele- vant recommendations are summarized here (71): – Land use should be integrated into multiple stages of the travel demand model, and the integration of land development patterns into the models is paramount. – Understanding and incorporation of freight-based activities into travel demand modeling, because it can require specialized surveys, and their effect on traffic. – Migration to activity-based or tour-based modeling to be conducted only with ample funding, resources, and proper documentation; otherwise, it is recom- mended to improve the existing trip-based models. – Agency creation of coordinated data collection strategies and standardization guidelines for regional modeling. – Improvement of data quality by supplementing existing data sources with specialized add-on sur- veys. This can help to complete data sets that are incomplete as a result of low survey participation or inadequate funding. – Consistency checks undertaken throughout the mod- eling process. – Model should be designed to be flexible enough for a variety of toll and HOV modeling policies to be evaluated. – Micro-simulation modeling can be used for more detailed modeling of areas that have unusual charac- teristics or that are highly diverse. – Consideration of time-of-day variables, despite cur- rent modeling difficulties. – Agency pooling of resources and sharing of their experiences through best practice publications. – High-quality documentation of the components of and assumptions in agency and region travel demand models. One example of the successful use of a peer review, as part of an overall process, was provided by a toll authority that responded to the survey of practitioners. This authority has used its travel demand model for several studies, including feasibility, policy, investment-grade forecast, design, review or audit, and state environment analysis studies. For the analysis of the toll facility feasibility study, the existing model was updated and enhanced. The success of the forecasting has been attributed to the regular refinement of the MPO model (upon which its model is based) over a 10-year period. The demand and revenue forecasting process has a built-in critical peer review process. The toll authority has taken a proactive approach in updating and enhancing a specific aspect of the model on an annual basis; revalidating the model annually

35 using traffic counts, origin–destination surveys, speed and/or travel time surveys, and land use inputs and network charac- teristics. The toll authority noted that it has recently received a rating upgrade on its revenue forecasts. There are no formal requirements for peer reviews in toll road traffic and revenue forecasts, although the bond rating community has called for more and improved reviews (5). Financial backers do conduct their own “stress tests” of the revenue and financial forecasts. In addition, as part of its project oversight and credit monitoring, The Transportation Infrastructure and Innovation Act, the federal credit assis- tance program for major surface transportation projects, requires a project’s senior debt to have the potential to achieve an investment-grade bond rating. It also requires the development of an ongoing oversight and credit monitoring plan for each project, which includes a risk analysis, and requires that traffic and revenue forecasts be updated and monitored (72). In comparison, the general practice in Europe is to con- duct three sets of forecasts: the grantors of the concession (the governments); the facility sponsors (proponents); and the financial backers (lenders, investors, and/or auditors) (C. Russell, personal communication, 2005). The govern- ments’ forecasts are normally considered to be overly opti- mistic, because they are used to develop long-range policies and plans. The proponents’ forecasts are usually the most extensive, although they do not always provide the best results because they are dominated by the model. The finan- cial forecasts (audits) are usually smaller efforts that are intended to review the proponents’ forecasts (although more substantive efforts may follow if the review identifies funda- mental problems). The audit relies on sensitivity tests, spreadsheet modeling, and stress testing: at the end of the process, the auditors also must assume responsibility for the forecast. The proponents pay for their forecasts as well as those of the financial backers. As a result, the proponents try to con- trol the latter’s audit. Some proponents now bring the audi- tors into the process early, before the lenders are appointed; this puts some pressure on the auditors, but also provides an opportunity for them to suggest improvements early in the process and—by the proponents who have put forward their case—allows issues generally to be understood. Conversely, in some places the use of two or more inde- pendent forecasts for proponents “has been dropped because (at double the cost) it confuses the audience,” with the result that—in a “tight bidding timeframe” much effort is wasted determining who is right, rather than exploring the key risk issues and refining one approach (D. Johnston, personal communication, 2005).