The appraisal of transport projects involves using transport models to forecast the impacts of the transport project some decades into the future. These forecasts are instrumental in determining:
The transport model forecasts also inform calculations of social and environmental impacts.
In order to understand these issues, transport model forecasts are made for the opening year of the transport project, which may be 10-15 years in future, allowing for the time involved in the planning processes and in the scheme design and construction phases. But as the transport scheme will have an economic and/or engineering life of 30 years or more, and must be demonstrated to be able to carry the forecast travel demands, forecasts are likely to be made for later years in the life of the scheme (eg 15 years after opening).
Transport model forecasts of 20-50 years into the future are self-evidently subject to uncertainties and it is common to analyse these uncertainties. Most governments require some consideration of risk and uncertainty, but the scope varies widely.
I do not propose to address the arguments for and against rigorous risk assessment here, but to illustrate some methods of estimating the uncertainty in transport model forecasts.
The UK government guidelines on transport modelling and appraisal are given in Webtag, and this includes forecasting and uncertainty. Originally Unit 3.15.5 dealt with this topic but this is now superseded by Unit M4.
In both, the first requirement is to produce an "uncertainty log" which summarises all known assumptions and uncertainties, covering the forecast inputs and the model parameters and specification. Associated with this is the recommendation that sensitivity tests are used to identify those modelling aspects to which the forecasts are most sensitive.
Whatever we choose to call it, common to any attempt to understand forecasting uncertainty is something like an uncertainty log which identifies the key sources of forecasting uncertainty, and which is established using knowledge, judgement and sensitivity tests.
There has been much research on the topic of uncertainty and I was involved in one of the early studies by the UK government which reported in 1980 (entitled Highway Appraisal under Uncertainty). A paper describing the work can be found in Transportation, Volume 9 (1980) pp 249-267, entitled Uncertainty in the Context of Highway Appraisal.
The literature identifies three categories of error: forecast input error (IE), model specification error and model estimation error (because these two sources are difficult to separate I classify them both as model errors, ME).
In the study, we used a model developed for a then current highway project. The process followed was to:
In the Monte Carlo process, we ran the traffic model many times, in each run revising the forecast inputs and the model specifications (the coefficients) using sampled values for each source of error from their respective distributions.
The sources of error identified and quantified were:
We quantified the expected range of error for a given input or model coefficient using judgement, research and/or external views.
Finally, given estimates of the size of the uncertainties, the error distributions were generally a matter of judgement.
Having to run a complex model repeatedly in a Monte Carlo simulation of errors is not usually practical in a study, many models taking a considerable time to run. Consequently, a follow-on research project looked at whether this effort could be reduced using experimental design techniques.
In the following I illustrate some of the issues and solutions with examples drawn from transport project studies.
In any practical approach, something of the nature of the uncertainty log is the startpoint. The log should cover only the important sources, and sensitivity tests and judgement should be used to eliminate from consideration minor sources of uncertainty.
For the study of Speedrail (a high speed rail line between Sydney and Canberra) in 1999 the major forecasting uncertainties identified were:
For the surface access mode share forecasts in connection with planning the development of Stansted Airport in 2006/8, the major sources of forecast input errors in the air passenger forecasts included (this is not a complete list):
On the East Coast High Speed Rail Study (between Melbourne and Brisbane) which reported in 2014, the major sources of forecasting uncertainty were:
There are a number of issues to address in practical appraisals of uncertainty:
Concerning point 1, sensitivity tests are excellent for addressing the impacts of specific, important issues. In the studies of Speedrail and of East Coast High Speed Rail, sensitivity tests were used to investigate the impacts of changes in the characteristics of the competing transport modes (including competitive reactions to the high speed rail services). They are also useful if there are a few very significant uncertainties whose impacts need to be highlighted. In the case of East Coast High Speed Rail the pace of population and economic growth of the eastern states of Australia was addressed in sensitivity tests.
When there are a significant number of sources of uncertainty, the value of sensitivity tests alone reduces, the problem being that the many sources of uncertainty cannot be combined in a discrete number of simple sensitivity tests without generating very unlikely scenarios. In such circumstances, Monte Carlo techniques should be able to provide a more balanced assessment of the risks.
This is a matter for the context and jurisdiction. But, in principle, the estimate of forecasting uncertainty based only on input errors underestimates the risk, which may be serious in some contexts.
In estimating the size of the errors and their distributions, as shown by the earlier example, there is inevitably much judgement, which can be informed by:
The selection of error distributions is mainly based on judgement: for trip matrices gamma or some similar non-negative distribution would be appropriate, while rectangular distributions are useful where there is no strong reason to expect the error distribution to be centrally focused.
With large numbers of sources of error, as there inevitably are in complex transport model systems, the estimation of forecasting uncertainty can be intimidating. As I have earlier remarked, while the original research demonstrated that estimates of forecasting uncertainty could be produced, the Monte Carlo process involving repeated runs of complex transport model systems is not practical for most studies.
There are other issues too. With a large number of errors sources, the work required to determine the uncertainties associated with each will also not be acceptable. Further complicating the error combination task with large numbers of error sources is the possibility of correlations between the errors, particularly the case if large numbers of model coefficients are included in the uncertainty log.
What all this implies is the need to design the approach to the study of uncertainty in a way which can meet practical constraints but does not compromise its value. The discussion that follows is based on some real examples which illustrate ways in which this may be done.
On the projects in which I have been involved, there were two fundamental aspects of the approach to estimating uncertainty. The first was to develop a much simplified, more aggregate version of the model in which the key uncertainties could be represented, and which was capable of being used in a Monte Carlo framework. Associated with this, the second aspect was to focus on the overall errors of the model output, rather than on the individual errors associated with every element of the model specification (eg all the model coefficients).
As an example, in one project (A) we created a simplified model of the form:
Scheme patronage = base market * market growth * diversion * (1+induced travel)
The base market was aggregated into around 20 segments (by trip purpose, mode of transport and broad geographic sector), for which survey error calculations were made. For these segments also, calculations of the growth factor errors were made (based on uncertainties in the population and GDP forecasts and the elasticities of travel demand to income growth).
The segments were further aggregated for the diversion error analysis. Sensitivity tests of variations to the mode choice model coefficients were used to derive an overall estimate of the uncertainty of the diversion forecast for each aggregated segment. The range of uncertainty for the induced travel forecast was based on judgement.
The simplified model and the associated levels of uncertainty were incorporated in the "@Risk" package. Monte Carlo analyses were used to derive the overall range of uncertainty of the forecasts of patronage (and yield).
A variant of this approach was used in other studies (B and C), both concerned with forecasting the mode shares for new public transport facilities.
As before, the key uncertainties, a subset of the full range of uncertainties, were first identified (in the log) and sensitivity tests were used to identify the effects on the model forecasts of each individual source of uncertainty. These were then combined in a Monte Carlo simulation of the form:
Fk = F0 *(1+u1k) *(1+u2k) *(1+u3k) *(1+u4k) *(1+u5k) *(1+u6k) .......
F0 is the core model forecast
Fk is a revised forecast "k" allowing for uncertainties in each source
ui are the individual sources of uncertainty and uik are the sampled uncertainty perturbations in forecast "k" for each source "i".
This simple model is run many times sampling values of ui from their respective distributions of uncertainty, to give a distribution of the forecast Fk. In both studies the Monte Carlo approach was implemented within a spreadsheet.
In study C, it was further postulated that some sources could be correlated and the combined variance calculations were amended to allow for partial correlations between these sources.
Also in study C, the models were far too complex to consider estimating the uncertainty due to model errors by sensitivity testing individual coefficients. However, the forecasts of market share made by this model had been extensively validated against the market shares achieved by other, similar projects. Therefore, instead of individual model parameter tests, we used the results of the validation to estimate the likely overall range of uncertainty around the market share forecasts.
In all three studies, individual sensitivity tests were also used to illustrate the impacts of those combinations of uncertainties of most concern to the decision-makers.