Chapter 13
Forecasting the demand
for transport
Contributed by Geoff Riddington
Learning Outcomes:
On reading this chapter, you will learn about:
᭿ Alternative approaches to generating a forecast of demand for existing, new or improved
services
᭿ Issues surrounding asking people how they or the public would react to new or improved
transport services and the problems that will occur
᭿ Methods for identifying trends and projecting demand for existing services when no
major changes are expected
᭿ Methods for identifying and projecting seasonal change
᭿ Methods for forecasting demand when significant change is expected in the economic and
social environment
᭿ Methods for forecasting the impact of new or improved services in a competitive
environment.
INTRODUCTION
In order to assess if the provision of a new or improved transport service makes economic sense we
need to have some idea of how the public will respond, both immediately and, because transport
investment has a very long life, in the far distant future. In Chapter 3 the theory underlying the
demand for transport services was discussed. In Chapter 14 you will learn how we take our
demand forecasts and assess if they represent a good use of our money. In this chapter we examine
the practical issue of generating a forecast.
It should be emphasised from the outset that forecasting is not about applying a mathematical
formula to a set of data but rather it is about collecting information from all relevant sources and
analysing it in a consistent structured fashion. Remembering that transport is a derived demand,
it is only when nothing significant is happening or expected in the external environment that
the estimation of a mathematical trend or a pure data relationship is sensible. Thus, for example,
when climate change and oil prices are having an impact throughout the world, projecting demand
for air transport based only on air traffic data from 1970 to 2000 is not sensible. Huge and
309
environmentally damaging investment in airports based on such trends would appear to be equally
debatable.
Integrating information and data about the external environment mathematically can be
extremely complex and, indeed, may be impossible. In some cases a set of human brains will
produce better forecasts. However, the information in the brain has to be structured and organised
or we can be guaranteed biased and inaccurate forecasts. Trend Analysis and unstructured groups
vie with each other to produce the very worst outcomes. The cost of transport developments
means that poor forecasts will generate significant reductions in the welfare of the public. However
because of the regulated monopoly position of many/most transport suppliers, this will
not be reflected in lower profits, only higher overall costs. Thus we, the general public, have a
vested interest in ensuring that demand forecasting is taken seriously and the effort expended is
proportionate to the costs to us all of getting it wrong.
Finally it should be noted that there has been extensive research on how best to forecast, much
of which is, unfortunately, regularly ignored. Those wanting to carry out forecasting are recommended
to consult Principles of Forecasting: A Handbook for Forecasters and Practitioners (ed: J. Scott
Armstrong, 2001) which summarises much of the research and offers practical guidance. For the
more general reader Long-Range Forecasting: From Crystal Ball to Computer by J. Scott Armstrong
(1985) is strongly recommended for both entertainment and enlightenment.
GENERAL APPROACHES
There are, broadly speaking, three approaches to forecasting demand:
1 Qualitative: Surveys and Sampling
2 Time Series Analysis
3 Econometric Techniques.
Figure 13.1 provides a decision chart for this first choice of which method may be the most
appropriate to use for a given situation.
Figure 13.1 Choosing between methods
310
FORECASTING THE DEMAND FOR TRANSPORT
From this choice diagram there are some methods that will normally be associated with specific
types of problems. Time series methods will normally only be relevant in the short term (when
little externally will have changed). They are particularly strong at identifying and forecasting
seasonal effects.
Econometric methods need a lot of good quality data that can only come from public sources.
Some information will not be available for reasons of commercial confidentiality. There will also
be some difficulty in forecasting some of the key factors,e.g.the price of oil in 25 years’ time.Often
we need to develop a range of possible futures for these determining factors (termed scenarios)
and generate a separate forecast for each scenario.
Where there is no equivalent to the proposal, e.g. demand for travel to the moon, and no past
data, then the intentions of potential customers or the opinions of a range of experts have to be
sought. Intention Surveys will, in general, only be appropriate in the very short term. Expert
opinions will be required for new products and the very long term that involves significant technological
development. We now look at these Qualitative Methods and their potential problems.
QUALITATIVE METHODS
Qualitative Forecasting Methods are based on surveys of either potential customers or ‘experts’,
such as key suppliers. Intentions information can provide the basis of excellent forecasts. As an
example forward orders and commitments, such as deposits on holidays or numbers already
booked on a service, provide an excellent basis for making forecasts up to one year ahead. Similarly
asking existing customers of their intention to repeat the experience can provide excellent results.
The major problem with intentions is identifying who to ask. In some cases the target group
may be relatively small and contained. A new bus route to a commercial estate for example will
predominantly attract those who work or visit the estate and this group will define what is known
as the sample frame. For an accurate representation of what people say we only need ask some
of the group, what is known as a representative sample. It is important to ensure that no groups are
over-sampled and none under-sampled, so it is necessary to ensure that questioning take place
throughout the area and throughout the day. If we know the numbers in the various groups that
make up the sample frame, e.g. the percentage of females over 60, then we can ensure that each
group is properly weighted. Note that the numbers in each stratum do not have to be in proportion
(although it can make calculations easier if they are) but there does have to be a significant number
in the group sampled.
With new facilities and services, in the short term estimates tend to be too high. This is
primarily because people over-estimate their capacity to change and always believe extending their
options beneficial. A new bus service or cycle track may seem a wonderful idea when it is
suggested but when the car is standing in the drive and the rain is pouring down, a commitment to
use the bus or cycle to work will often be ignored. In the long term estimates may be too low
because the population itself changes. Individuals without cars will find the estate more attractive
as a place to work or shop and workers on the estate will choose accommodation along the bus
route. Forecasts from intentions thus need to be adjusted up or down to account for known biases.
In some cases the population concerned is almost impossible to identify or is simply too large to
sample properly. For example suppose we are interested in forecasting the number of people who
311
FORECASTING THE DEMAND FOR TRANSPORT
might want to use a weekly service from a local airport to an airport in southern Italy. Potentially
anyone in either local area, which will inevitably run into millions, might use the service and will
need to be sampled. Suppose the number in both catchment areas is 6 million and we need over
6,000 customers (0.1 per cent) to make the service viable. Even if in fact 9,000 are interested we
need a sample of 2,500 people to confirm that we exceed 6,000. If the actual is 20 per cent over
the minimum the sample size needed is 12,000. This sampling has to be undertaken in both Italy
and the UK adding further to the cost. Given the potential biases already identified and the
relatively small cost of short-term transfers of buses, planes and even ferry boats, most firms prefer
to simply try out possible routes for a limited period particularly if they can obtain local financial
support from the origin and destination areas and they have some indication from ‘gravity’ models
(outlined later in the chapter) that they might be viable.
Forecasting by opinions happens in every organisation at least once a year in the development
of budgets. Politically every manager seeks to maximise the income to their divisions without
exposing them to unrealistic sales targets. The sales targets will normally be based on the opinions
of those ‘on the ground’ on the likely demand, particularly relative to existing demand. In practice
the political imperative of meeting targets leads to diversion of marketing effort often involving
price reductions. The outstanding record of low-cost airlines in flying close to capacity is not an
indication of excellent forecasting by managers but of a pricing system that combines bookings data
and pricing to ensure the plane is filled.
Forecasting using only expert opinion should normally be restricted to novel situations where
the alternatives of asking potential customers or using a quantitative model based on similar
situations are not available. Even then it is essential that experts from outside transport are
involved, indeed it has been said that the only expert that should be used is an expert in forecasting.
There are three major problems in seeking expert opinions, anchoring bias, group think and status
deferral. Because ‘experts’ tend to be elderly they give weight to early experiences which may not
be appropriate in times of rapid change. Thus anchoring bias often results in underestimates of the
rate of change possible. Group think tends to have the opposite effect. In a group, individuals do
not have responsibility for the outcome and are tempted towards more risky projects and higher
forecasts than warranted. This is particularly apparent if the senior manager in the group is
‘entrepreneurial’ and a risk taker. In these circumstances ambitious individuals will invariably defer
to the higher status individual since opposing, even if perfectly correct, tends to have negative
career effects.
To overcome these problems four rules should be adhered to:
1 The group should be facilitated (chaired) by an external
2 Experts from a variety of disciplines should be used
3 All opinions should be given equal weight
4 Reviews of previous forecasts should always be undertaken.
One popular approach known as the Delphi technique (named from the ancient Greek forecaster
The Oracle of Delphi) involves a group of individuals from different disciplines (and locations)
who are physically separated. The experts individually present their forecasts and the arguments
for them to all the others. They then revise their forecasts in the light of the forecasts of others and
present them again. This carries out round after round until consensus is reached. The job of the
312
FORECASTING THE DEMAND FOR TRANSPORT
facilitator is to identify, for the participants, the major points of dispute and, if necessary, force
agreement on these as a pre-requisite to agreement on the forecast. In the early years the
individuals were in separate rooms at the same site but current information technology provides a
perfect platform for the technique.
Although expert opinion has a limited role as the sole forecasting method it is important
in reviewing and modifying forecasts from the data-based approaches that are now discussed. It
should again be emphasised that forecasting is about combining information from all sources in a
structured fashion; experts will have knowledge that is not contained within the data.
TIME SERIES ANALYSIS
Introduction
On occasions we want to forecast demand for a transport service that has a long recorded history.
Riddington (1999) for example was interested in forecasting the numbers using charter flights to
skiing resorts in Europe, data which has been recorded by the Civil Aviation Authority since 1981.
The Department for Transport records and forecasts the numbers using the various modes,
road, rail, air and sea. Over time certain patterns emerge. The most obvious and important is
fluctuations over the seasons.
A further example would be on short sea crossings, where we may expect higher demand in the
summer than during the winter. Figure 13.2 shows the total number of passengers travelling on
ferries to and from the UK each year and Figure 13.3 shows the number each quarter.
As expected, there is a marked seasonal pattern with passenger numbers rising steeply in the
second and third quarters. There is also however a clear downward trend.
In time series analysis, from such cyclical patterns we seek to identify the three elements:
1 The Trend
2 Seasonal or Cyclical factors
3 The unusual (sometimes termed the stochastic factor or noise).
We then assume that these will continue into the future. As discussed earlier this may be reasonable,
particularly in the short term. However, in this example, given changes in air fares, holdups
Figure 13.2 Annual and quarterly short sea passengers, inward and outwards, 1998 to Q2 2008
313
FORECASTING THE DEMAND FOR TRANSPORT
and delays at airports and faster ferries, it would be a very brave forecaster who would assume the
downward trend illustrated will continue. A very well-known verse attributed to the famous
economist Roy Harrod runs:
A trend is a trend is a trend,
the question is will it bend?
Will it change course
through some unforeseen force
and come to a premature end?
If the forecaster is interested in forecasting on a daily or weekly basis then a number of sophisticated
methods such as ARIMA, ARARMA, Trigg-Leach and Brown’s Double Smoothing have been
developed. These involve weighted moving averages of the data to obtain an estimate of the
underlying level and weighted moving averages of variations from that level. After a number of
competitions between them, which method is ‘the best’ is still subject to much debate. Interested
readers are referred to Makridakis and Hibon (2000) for further information.
Trend curve analysis
Most transport economists are primarily concerned with the medium to long term and one
advocated approach is trend curve analysis. This approach involves trying to fit a line as closely as
possible to the annual (or deseasonalised) series. The procedure of ‘fitting’ is known as Regression.
Those unfamiliar with regression are referred to any introductory book on statistics or web sources
such as Wikipedia, which gives a reasonable introductory overview. In essence regression replaces
fitting lines by eye with fitting a line that minimises the square of the distance from the line to the
point. Figure 13.4 shows the line of best straight line fit associated with the annual sea passenger
series.
The formula for a straight line is Y = α + βX and the distance between the line and the actual
number is known as the stochastic term, the error or the residual term and is symbolised by ε.
Thus the formula for the actual number is Y = α + βX + ε, where, in this case, Y is the sea
Figure 13.3 Annual and quarterly short sea passengers, inward and outwards, 1998 to Q2 2008
314
FORECASTING THE DEMAND FOR TRANSPORT
passengers and X is the year. To obtain values for α, β and ε we employ statistical software such as
the Data Analysis Add-In in Excel. Table 13.1 shows the data and Figure 13.5 the output that is
generated.
The coefficients give the values of α (2134399282) and β (−1052246.72). To obtain a forecast
of passengers in 2020 we simply substitute 2020 for X in the formula Y = α + βX, which gives
2134399282 − 1052246.72*2020 = 8,860,905. This is considerably (and implausibly) lower than
the current level of 23,667,651.
The straight line implies that a fixed number of passengers (1,052,247) will be lost every year.
In reality growth and decline is normally a fixed percentage rather than a fixed quantity. This is
represented by an exponential curve Y = α * eβX
rather than a straight line. The exponential term e
is rather like π in geometry, and has a fixed value of 2.718282. It is important because the equation
results in a fixed growth rate given by β. Depending upon assumptions made about the stochastic
term, the values of α and β can be estimated either by linear regression of loge(Y) against X or by
iteration. In this case the growth rate was found to be −3.8 per cent and the 2020 forecast
13,871,777 considerably higher than the linear forecast.
The question arises which forecast is ‘better’. Remembering that forecasting is about
Figure 13.4 Annual sea passengers, underlying trend
Table 13.1 Sea passengers per year
Y (PAX) X (YEAR)
......................................................
33225861 1998
31381282 1999
28516814 2000
27753461 2001
28726426 2002
26523377 2003
25798730 2004
23693468 2005
23465161 2006
23667651 2007
Source: DfT Statistics
315
FORECASTING THE DEMAND FOR TRANSPORT
combining knowledge then if there are strong reasons to assume fixed growth (bending trends)
and/or that the results are more plausible, then that is the appropriate forecast. However, there are
also some indications from the statistics that the ‘fit’ of the growth rate model is better. A measure
of the fit is known as the Adjusted RSquared which for the linear is shown (on the third line of
the output table) as 0.910391568. This suggests that we can explain 91 per cent of the variation
in the data series by the model with 9 per cent being due to the stochastic noise, i.e. remains
unexplained by the model. The adjusted RSquared for the growth rate model is 0.92326468, i.e.
the exponential model explains more. It should be emphasised however that a good fit is not the
same as a good forecast. Great care should be taken when selecting models using fit statistics
because they are dependent upon a number of assumptions about the stochastic term that may well
not be true. It should also be noted that, in theory, test statistics from linearised models are not
directly comparable with those from simple linear models. That said it is a helpful additional factor
in deciding whether a linear or exponential model is better.
Because fit and forecast performance can be radically different, if the data set is long enough, it
is good practice to exclude the last few data points when estimating the model. Forecasts from the
estimated model are then compared with the actuals (technically known as ex post analysis). There
is some debate in the relevant literature as to the appropriate measure of the forecast. The most
common is the Root Mean Square Error of the one step ahead forecast.
Figure 13.5 Excel regression output
316
FORECASTING THE DEMAND FOR TRANSPORT
Estimating seasonal fluctuation by seasonal dummies
When planning capacity, if there is marked seasonal fluctuation, then a seasonal forecast is
required. Again there are a large number of methods involving weighted averages of the seasonal
differences (or ratios) some of which are quite complex. A simple approach utilising regression
involves the use of dummy variables. A dummy variable is a variable that takes a value of 1 when
the phenomenon is present and zero otherwise. It was originally developed to take account of
periods of war in economic models but is equally applicable to the season. In the seasonal case a
dummy variable for Q1 has a value of 1 when it is Q1 and zero otherwise. We can thus extend our
example in order to incorporate not only the effect of time (the trend) but also seasonal effects
into our forecasts. Table 13.2. shows the first two and a half years of the data set up in Excel and
Figure 13.6 the results of the linear multiple regression of sales against the year and the seasonal
dummies.
It is important to note that when you have four quarterly dummies you cannot also have a
constant, as effectively the quarterly dummies are the constants for each season. The coefficient of
X indicates the decline per quarter and the coefficients of the dummies are equivalent to the
constant term for that quarter. Figure 13.7 shows how closely the model fits the data; some 97 per
cent of the variation is explained.
To forecast we take the coefficient of the relevant quarter and add the coefficient of X * the
time (i.e. 2020.25 for the second quarter in 2020).
The linear model estimated above implies a constant addition/subtraction to the mean for each
quarter. However, most of the time, the seasonal effect is most likely to be a proportion of the
underlying mean. As with the growth model a ratio seasonal factor can be estimated by regressing
the log of the sales on to the time and the dummies. In fact this log-linear regression produced a
remarkable fit and an adjusted R Squared just slightly higher of 97.3 per cent compared to 96.9
per cent for the linear model. Incorporating seasonal effects into our model therefore has improved
the overall fit, and hence may be expected to produce better forecasts.
Table 13.2 Quarterly sea passengers with seasonal dummies – first two and a half years
Y (PAX) X (YEAR) Q1 Q2 Q3 Q4 LogY
....................................................................................................................................
5151807 1998.00 1 0 0 0 15.45486
9304349 1998.25 0 1 0 0 16.04599
11801764 1998.50 0 0 1 0 16.28376
6967941 1998.75 0 0 0 1 15.75683
5036822 1999.00 1 0 0 0 15.43229
9452280 1999.25 0 1 0 0 16.06177
10996127 1999.50 0 0 1 0 16.21305
5896053 1999.75 0 0 0 1 15.58979
4179625 2000.00 1 0 0 0 15.24573
8464104 2000.25 0 1 0 0 15.95134
.. .. .. .. .. .. ..
.. .. .. .. .. .. ..
.. .. .. .. .. .. ..
.. .. .. .. .. .. ..
317
FORECASTING THE DEMAND FOR TRANSPORT
Unlike the trend, the factors that determine different levels throughout a season do not change;
Christmas always occurs on December 25th, the summer holidays in July and August etc. Problems
do occur when dealing with monthly data due to situations when there are 5 weekends or when
Easter occurs in March rather than April but, certainly at quarterly level, estimates of seasonality
are accurate and robust. This is, however, not the case with trends. Whatever has caused the
decline in sea passengers may quite conceivably reverse or become more important. Investigating
Figure 13.6 Output from regression of sea passengers on time and seasonal dummies
Figure 13.7 Actual and estimated sea passengers by quarter
318
FORECASTING THE DEMAND FOR TRANSPORT
what has determined change and using that information to forecast the future is the role of
econometrics, which we cover next.
ECONOMETRIC METHODS
Introduction
In Chapter 3 we looked at the factors that broadly determine the demand for transport; income,
price, quality of service, price and quality of alternatives, journey time and population size. The
object of econometric methods is to try and identify the precise importance of each of these
(the usual measure being the elasticity) in order that we can determine the effect on demand of
changing these in the future. Unfortunately, as we shall see, this is not necessarily a straightforward
task.
The modelling process involves 6 stages
1 Understanding the Problem: Identifying all the key factors and making preliminary estimates of
the size and direction of the effect. This is covered in Chapter 2.
2 Obtaining the Data: Data on ALL the factors has to be acquired. It is no use having price data
for your company but not for your competitors. Data is of three broad types:
a) Time Series, where the data, going back a significant number of years, relates to a single
location
b) Cross Section, where the data has been obtained from a series of locations (or
individuals) at a specific point in time
c) Panel Data, where data has been obtained from the same series of locations over a
period of years.
Each type of data has strengths and problems specific to it which we will discuss in the next
sections but, generally speaking, the major problem is acquiring good quality data at all.
Statistics issued by the government or its agencies are invariably the first place to look,
e.g. Transport Statistics, National Statistics, The Census and the National Travel Survey. The
Civil Aviation Authority also produces some superb data series. All these are now available
on the Web.
It is extremely difficult however to obtain good data on two key factors, Price and
Quality. Consider trying to ascertain the impact of a more expensive but faster train from
Scotland to England. This needs information on the changing quality of the experience,
journey times, frequencies, changes in airport check-in time, queues at security, overcrowding,
traffic jams etc.
Equally critical is the actual price paid both on the train and the competitive air, coach
and road services. Because the average price paid is a commercial secret we are forced back
to published ‘normal’ prices which may be substantially higher than actual price. In addition,
in the case of some airlines, the published price varies from hour to hour depending upon
the current loading. The ‘cost’ in both journey time and ticket price, of getting to the final
destination, must also be acquired.
319
FORECASTING THE DEMAND FOR TRANSPORT
A fundamental problem is that the omission of a key variable totally distorts estimation of
the impact of the other factors. Poor or substitute (proxy) data is better than no data but
inevitably decreases confidence in the outcomes.
3 Specifying the Model: It is tempting to assume that the factors are independent of each other,
that the factors are additive in their effects and the effect happens immediately and that all
the assumptions relating to the stochastic term are applicable. Unfortunately that is rarely
the case. The specification stage involves:
a) Selecting the functional form (linear, log-log, logistic etc)
b) Selecting the variables and in time series any lagged effects (e.g. with advertising)
c) Making reasonable assumptions about the stochastic term (e.g. that the bigger the
demand the bigger the stochastic term).
4 Estimating the specified model: Identifying values for the parameter estimates for each factor
to be used in the forecast α, β1, β2, β3, . . . that makes the predicted values lie as close as
possible to the actual values. This normally involves minimising the sum of squares of the
residuals.
5 Validating the Model: This involves examining the outcomes to ensure:
a) The values of the coefficients (the elasticities) have the right signs (income positive,
price negative and so on) and are of the expected size
b) The fit is statistically significant (The RSquared and F values in the output table)
c) The individual variables are statistically significant (the t value in the output table)
d) The residuals have no pattern over time or location.
If the model fails these tests then we need to return to the specification stage (or even the
understanding).
6 Simulation/Forecasting: Once we have the final model then it is a question of inserting
expected future values of the factors into the equations. Sometimes we can use forecasts
from other organisations but often we need to use suggested values obtained from expert
groups. Sometimes different sets of predictions are bundled up to form scenarios. Sometimes
we simply use trend extrapolation of the external factors.
This last section illustrates a theme of this chapter. The purpose of econometric modelling is both
to obtain information on the effects of the demand factors but also to integrate information from a
variety of sources into the forecast. We now look at some important specifications and problems
associated with them before illustrating how these come together in assessing demand for a new
route.
The gravity model
It would be extremely surprising if the demand for services (which includes the demand for
road space for cars and lorries) between two large cities such as London and Manchester was not
much, much larger than between two small towns such as Cirencester and Lincoln. It would also
320
FORECASTING THE DEMAND FOR TRANSPORT
be very surprising if the numbers travelling between two adjacent towns say Cambridge and
Norwich was not much larger than between Cambridge and a town of a similar size to Norwich in
Scotland, say Dundee. Hence the level of transport between two locations will be dependent
upon their respective population sizes and the distance between them. In more formal terms, our
understanding suggests that the flow of people between two points i and j, Fij, is a function of the
size of the origin Oi, the size of the destination Dj and the cost of travelling between them Cij. At
this stage it is important to note that the cost is directly related to the distance between i and j and
the time taken; a direct motorway will reduce the time and hence decrease what we term the
generalised cost.
To estimate our model in order to predict the numbers flowing down a particular route we
need data on flows between a number of towns and the generalised cost of travelling a route. This
is a typical example of cross sectional data.
The normal specification of the model is of the form:
Fij = αOβ
i Dγ
j Cδ
ij εij
where α,β,γ and δ are parameters to be estimated and ε is the stochastic term. In simple terms,
what these show is the relative impact each factor will have on the numbers travelling between
these two locations taking into account the differing units these may be measured in. Thus for
example the β is the impact of the size of the origin on total traffic flow. In a similar fashion, γ is
the impact of the size of the destination and so on. The gravity model gets its name from the
models of gravitational pull developed by Newton where the size of the bodies is multiplied
together with the distance between them.
It is also very easy to estimate the model by taking logs to give:
log Fij = A + β log Oi + γ log Dj + δ log Cij + Eij
where A is log α and E is log ε. If we assume that E is normally distributed then the normal
method of minimising the sum of the squares of the stochastic term (known as Ordinary Least
Squares) can be applied.
Riddington (2002) used a gravity specification to forecast the numbers of people who might
travel to a new supermarket in town H. The numbers of people in each of the 12 surrounding
towns and villages was known. The complication was the existence of 3 other supermarkets in
town D which currently attracted custom from the same towns and villages. The two key
determinants of the proportion of each settlement that would go to H rather than D was the
relative distance to H and the relative size of the shops; the more floor space/goods on offer, the
more attractive the destination. It was assumed that a fixed proportion of each settlement would
shop in either H or D in any week. A possible range of values for the impact of supermarket size
and distance were obtained from other gravity model studies. The mean of these values generated
a central forecast, the range of values and the possible range of outcomes. This work showed
conclusively that the estimates of diversion from D to H made by the proposers was exaggerated
and that, even at the most favourable range of values, the new supermarket would lead to serious
problems and small shop closures in the existing town centre.
The gravity model is central to large network planning transport models, particularly the four
321
FORECASTING THE DEMAND FOR TRANSPORT
stage Transport Planning Model, where stage one and two are to estimate the total number of
trips within a given area, and then the number of specific flows between two locations, before
subsequent stages then apportion these flows to modes and then routes. Gravity models are also
extensively applied in models of economic trade and thus used to predict trade flows. Nevertheless,
the focus is on populations and generalised costs and the specification is totally unable to
describe the massive growth in car and air journeys. For this we turn to traditional demand models
formulated over time.
Econometric demand models
Whilst population change is of critical concern in a local context and in developing countries,
in the developed world it has been relatively static and is not an explanation for the growth in
transport that has been a characteristic of the last century. As seen in Chapter 2, we have tended to
associate this growth with the growth of the economy and the associated growth in real personal
incomes.
Our understanding based on the theory of Chapter 3 suggests that, at the national/regional
level, the demand for a particular mode (road, rail, air, sea) will be determined by National
Income, comparative price, comparative journey times, frequency and comparative quality. Data
for income over a number of years can be obtained easily. Journey time information for the same
years needs to be identified with allowances for travel to public transport and times to enter and
exit the terminal (including the time to wait for the service). As discussed earlier, obtaining good
quality price data for a number of years is problematic and normally ad hoc series based on what
can be obtained over the years has to suffice. Finally there is the issue of comfort. The modern car
is significantly more comfortable and reliable than its real price (i.e. net of inflation) equivalent
of twenty years ago whilst the low-cost airline is not as luxurious as its historical forebears. We
normally either:
1 Simply assume quality changes are part of the stochastic term (not normally appropriate)
2 Allow for these changes by introducing proxy variables that are linked to quality
3 Develop what are known as hedonic price indices.
Broadly speaking econometric demand models have two functional specifications: linear and multiplicative
(log-log). There is some evidence that consumer responses to economic changes are
constant, e.g. that a 10 per cent rise in incomes will lead to a 7 per cent increase in demand,
whatever the current level of income. This is a situation of constant elasticity and is characteristic
only of a multiplicative model. Thus, on these grounds alone (there are others), the following
specification might be used:
Log Qt = α + β1 log Yt + β2 Log Pt + β3 Log Jt + β4 Log F t + εt
Where Qt is number of passengers, Yt is income, Pt is relative price, Jt relative journey time and Ft
is relative frequency. Once again in simple terms the parameter estimates (β terms) are the
relative impacts on demand of each of the factors included, taking into account the units in which
each is measured.
322
FORECASTING THE DEMAND FOR TRANSPORT
Because of the difficulty with national data on Price, Journey time and Frequency, a truncated
form Log Qt = α + β1 log Yt + εt is often found (e.g. DfT traffic forecasts). The relationship between
demand and income is very strong and found throughout the developed world. It is also very easy
to make forecasts. For example if β1 (the income elasticity) in a model of car journeys has a value
of 1.4 then growth in GDP of 2 per cent will lead to a growth of 2.8 per cent in car journeys.
Of course forecasts of GDP for 20 years ahead are required in order to forecast transport demand
20 years ahead. In practice we normally present a range of possible growth rates for the economy
and consequently a range of possible forecasts.
It is undoubtedly true that this procedure simply replaces a trend curve projection for
passengers with a trend curve projection for the economy as a whole. However, other information
about economic factors may be available and one of the most important strengths of this approach
is the structured combination of information about the future. Unlike the time series approach
outlined earlier, we are not only reliant on a single past trend within the transport system, but also
the impacts of external factors upon that trend. Hence if a trend was to bend, there would be a far
higher possibility that our forecast will pick it up.
The use of only GDP in the specification rather than all the variables is much more critical (as is
the omission of lagged values in short period models). Suppose that over the same period the
relative price of motoring has been falling. If it had a price elasticity of −1 and prices fell at 1.5 per
cent per annum then the growth rate due to price change is 1.5 per cent. If the total growth is
2.8 per cent then only 1.3 per cent of that growth is attributable to economic income, not the
full 2.8 per cent. If the decline in relative price was reversed then growth due to price change
would be negative and the growth rate correspondingly lower. Thus it is absolutely essential that
we include all relevant variables or the results concerning individual elasticities can be totally
misleading. Failure to include the correct variables or the correct functional form is known as
misspecification.
Misspecification and the demand for ferry services
Earlier in this chapter we observed the number of ferry passengers in decline. Since foreign travel
is a luxury, our economics suggests it is likely to grow with income, hence it will have a positive
price elasticity. We might also suspect that cheaper more available air services and the advent of
the Channel Tunnel might have an effect on demand for ferries. Table 13.3 shows data collected on
passenger numbers by different modes from the Social Trends database.
Our first task is to see how well our basic understanding fits to the data. Graphing the data
is always a good start but in this case Figure 13.8 does not tell us much that we did not already
know.
Another useful tool is the correlation matrix. This shows how closely any two series are aligned.
A figure close to +1 shows a strong positive (complementary) relationship, close to −1 a strong
inverse (competitive) relationship. For example we would expect number of cars and sales of
petrol to have a high (>0.8) correlation but number of cars and bus passenger numbers to have a
strong negative correlation. (<−0.8). Tables 13.4 and 13.5 show the correlation matrices between
the numbers by mode and real GDP for data from 1981 and from 1995.
A significant difference between the two is the strong negative correlation between Sea Travel
and Income in recent years (−0.93). It may be that this reflects the idea that sea travel on holiday is
323
FORECASTING THE DEMAND FOR TRANSPORT
Figure 13.8 UK international passengers by mode
Table 13.3 UK international passengers by mode (million)
YEAR AIR SEA TUNNEL TOTAL REAL GDP
....................................................................................................................................
1981 11.4 7.7 0.0 19.0 56.2
1991 20.4 10.4 0.0 30.8 73.0
1995 28.1 10.0 3.2 41.3 79.5
1996 27.9 10.7 3.5 42.1 81.7
1997 30.3 11.5 4.1 46.0 84.3
1998 34.3 10.5 6.1 50.9 87.1
1999 37.5 10.4 5.9 53.9 89.7
2000 41.4 9.6 5.8 56.8 93.1
2001 43.0 9.7 5.6 58.3 95.3
2002 44.0 10.0 5.3 59.4 97.3
2003 47.1 9.2 5.1 61.4 100.0
2004 50.4 9.0 4.8 64.2 103.3
2005 53.6 8.1 4.7 66.4 105.2
2006 56.5 8.4 4.7 69.5 108.2
Table 13.4 Correlation matrix. Passengers by mode from 1981
AIR SEA TUNNEL TOTAL REAL GDP
....................................................................................................................................
AIR 1.0000
SEA −0.2784 1.0000
TUNNEL 0.7433 0.1755 1.0000
TOTAL 0.9886 −0.1499 0.8270 1.0000
REAL GDP 0.9900 −0.1485 0.7698 0.9933 1.0000
324
FORECASTING THE DEMAND FOR TRANSPORT
an inferior product with a large negative income elasticity, i.e. as we get richer we are ceteris
paribus less likely to take our cars on to a ferry to Europe. Conversely we may well believe that this
is a short-term effect brought on by rapidly declining air fares in recent years and that the longer
series with a very much weaker correlation (−0.148) is a better indication of what to expect in
the future. It is important to recognise that our understanding from economic theory of what
underlies change is crucial to modelling and forecasting.
To forecast demand for sea ferries we really require the price of ferry services, price of air
services and the price of tunnel services. As discussed earlier obtaining a ‘price’ for a single route
is extremely difficult, for a combination virtually impossible. Since we expect price and demand
to be quite strongly inversely related we can sometimes use ‘numbers’ as proxies for prices. In
addition the number of air passengers also reflects speculative increases in capacity. Thus it seems
quite reasonable to try to explain the number of sea passengers by numbers on the other modes
and income.
Figure 13.9 gives the Excel Output of the multiple regression of Number of Sea Passengers
on passengers using air services, passengers using the tunnel and GDP.
The overall fit of the model is very good (95 per cent variance explained) but, more importantly,
the model coefficients seem to have the right sign. The income elasticity of demand was
defined in Chapter 2 as the rate of change in Quantity divided by the rate of change in Income, i.e
in symbols (∆Q/Q)/(∆Y/Y) or (∆Q/∆Y)/(Q/Y). In a linear model ∆Q/∆Y is given by the
slope of the line, i.e. the coefficient (in this case 0.429856). The elasticity will, however, vary over
different values of Q and Y so, by convention, it is usually calculated at the mean values of Q and Y.
The income elasticity of sea travel calculated at the means in this example is very high at just under
4. Surprisingly the truncated data gives almost identical results.
This modelling exercise is important because it highlights the way that increases in air
passengers, rather than the channel tunnel, has been the most important factor in slowing down
demand for ferry services. If airline growth is checked because of higher fuel prices and carbon
pricing then we would confidently expect significant growth in the ferry market well in excess of
the growth of GDP.
It is important to note that if data on prices were available it would be far better than using the
proxy variables of passenger numbers. In addition a better modelling strategy, which is illuminated
in the case study at the end of this chapter, might well be to model the total market and relate that
to GDP and model mode choice separately based on factors such as price and journey time.
The use of ordinary least squares is in theory limited to situations where the stochastic
term:
Table 13.5 Correlation matrix. Passengers by mode from 1996
AIR SEA TUNNEL TOTAL REAL GDP
....................................................................................................................................
AIR 1.0000
SEA −0.9471 1.0000
TUNNEL 0.1440 −0.0812 1.0000
TOTAL 0.9939 −0.9226 0.2437 1.0000
REAL GDP 0.9975 −0.9354 0.1044 0.9890 1.0000
325
FORECASTING THE DEMAND FOR TRANSPORT
1 Has Zero mean
2 Is normally distributed
3 Has constant variance
4 Is independent over time and space
5 Is uncorrelated with the independent variables.
As discussed earlier, there are also problems if the factors are not independent of each other.
Econometric packages offer a number of estimation methods to deal with these problems and these
are important if the main objective is to identify precisely the effect of one factor on another and
the data quality is good. However, when we come to forecasting, low-quality data and the problem
of obtaining good forecasts of the independents normally means the accuracy will always be
limited. In these circumstances research suggests that theoretical sophistication is not as important
as structured thought about the future.
MODELLING CHOICE
Background
It is often the case that we are more concerned with forecasting the share of existing traffic
than the growth of that traffic. Consider, for example, investing in a new toll motorway that
runs almost parallel to an overcrowded existing motorway. The key question is how many
Figure 13.9 Regression output from Excel
326
FORECASTING THE DEMAND FOR TRANSPORT
vehicles we might expect at various levels of toll. The first task is to generate a gravity type
model that incorporates tolls and reduced journey times within the generalised cost function.
This will then give us a first estimate at the total traffic on the route. The next task is to share
the traffic between the toll and the free motorway. This is not however the end of the game
because if the tolls are set too high, traffic will stick to the free motorway, journey times
overall will increase and traffic will drop on both. In these type of situations we undertake what
is known as iteration, i.e. we put the new forecast of traffic and journey time/average price into
the model, generate a new share, recalculate total demands and generalised costs, re-estimate
shares, re-estimate costs and totals, and so on. Modelling choices and shares is not, however,
straightforward.
Based on our economic theory we assume that we will not undertake a journey unless we
achieve some Utility (satisfaction). Against that we have to place the Cost, in terms of journey time
(J), price (P) and lack of quality (L). We represent the importance of, for example, each minute
or of an inch of leg-room by β1 and β2. We call the sum of the values generated from the factors
the systematic value which for a choice Q is given by Vq = α − β1*Jq + β2*Lq − Pq + ε.
Thus say service Q is chosen over service R, or service Q is chosen over service S, and so on,
then we assume that it is likely that the systematic value of Q is greater than the systematic value
of R, S, . . . . i.e. the difference between Vq and Vr, Vs, . . . is likely to be greater than 0 (or the
ratio >1).
Data and specification
Choice modelling data comes in two forms:
1 Individual Data gathered within one time period
2 Market Share data which can be cross section, time series or panel.
Increasingly, when contemplating quality changes, a survey is undertaken where customers are
presented with a number of alternatives and asked to choose between them. For example we might
be interested in the best mix of speed, comfort and price on a train route. Customers are
presented with a set of combinations of the three at different levels and asked to select their choice.
They will then be presented with another set of combinations, and their choice recorded and then
a third set and so on. This generates a mass of individual data and is known as a Choice
Experiment.
For market share data Pr(Q) symbolises the proportion making a choice (e.g. going by mode
Q). For individual data Pr(Q) symbolises the probability that the individual will make a choice (e.g.
go by mode Q). Note that we cannot measure this probability, only the result. We assume that the
choice with the highest chance of being selected will in practice be selected.
Assume for the moment there is only one alternative to mode Q, mode R. The larger the utility
of mode Q relative to R, the higher the value of Pr(Q). It would be possible to simply regress
Pr(Q) against the difference in the systematic values of Q and R, e.g. Pr(Q) = β(Vq − Vr) + ε. This
is illustrated by the straight line SS in Figure 13.10
Note that when the utilities of R and Q are equal, the predicted market share (probability) must
be 50 per cent. Note also that Pr(R) = 1−Pr(Q). Thus ever larger negative values of Vq − Vr
327
FORECASTING THE DEMAND FOR TRANSPORT
increase the probability that an individual will use mode R, whilst ever larger positive values
increase the probability that an individual will use mode Q.
There are two major problems with this linear specification:
1 It suggests that it is possible to have more than 100 per cent of the market
2 It contradicts the law of diminishing marginal utility.
In this context the law simply tells us that, even if mode Q is vastly superior to mode R, a few
people will still opt for R. Thus we would expect the line of best fit to look more like Figure 13.11
One mathematical function that gives a line of that shape is known as the Logistic and has the
form
Pr(Q) =
eβ(Vq − Vr)
1 + eβ(Vq − Vr)
where e is the exponential operator (a constant).
With this function when:
Vq = Vr, Pr(Q) = 0.5,
Vq − Vr is very large, Pr(Q) tends to one,
Vq − Vr is large and negative, Pr(Q) tends to zero.
At first sight this looks very difficult to deal with, but some simple algebra produces:
Loge ΂
Pr(Q)
1 − Pr(Q)΃= β(Vq − Vr) + ε
Figure 13.10 Change in market share versus difference in systematic values
328
FORECASTING THE DEMAND FOR TRANSPORT
In order to clarify this ‘simple’ algebra, if we let β(Vq − Vr) = x and Pr(Q) = y then
y = ex
/(1 + ex
). Since (1 + ex
)y = ex
, y = ex(1 − y) and y/(1 − y) = ex
. Taking logs gives
Loge(y/(1 − y)) = x.
The first term is known as the LogOdds and is simple to calculate. For example if mode Q has
60 per cent of the market then R has 40 per cent and the logOdds is loge(0.6/0.4) = loge
(1.5) = 0.405. To estimate β where we have proportions data, we simply carry out the normal
regression of the LogOdds against the differences in the systematic values.
The problem with individual data is that we can only observe the choice made, i.e. it is either
100 per cent mode Q or 100 per cent mode R. This is illustrated in Figure 13.12.
Where the difference is very large we would expect the choice to be obvious. However, where
our model predicts that for example there is a 53 per cent chance that the individual will choose
Q then we will not be very surprised if they choose R. We use this idea to choose the parameters
β to ensure that the situations when the predicted likelihood is high, then that is the appropriate
choice and are less worried at values close to 0.5. This approach is known as Maximum Likelihood
Estimation and is available in large statistical packages such as SPSS, LIMDEP or STATA. It is
important to note that, because there is not much information conveyed by each individual about
the ‘norm’, very large amounts of individual data (typically in excess of 1,000 individuals) are
needed for reliable forecasts.
Forecasting shares
To summarise, the economic understanding suggests that choices are made on the basis of
differences between factors such as journey time and price. Data can be either from individuals on
their choices or from the market on the shares. The specified model has to take into account the
logical limits of proportions and the law of diminishing marginal utility. One common form is the
Figure 13.11 The Logistic curve with market share data
329
FORECASTING THE DEMAND FOR TRANSPORT
Logistic although other forms such as the cumulative normal, known as the probit, can be used.
Estimates of the parameter coefficients are found by regression of the LogOdds on the difference in
the factor levels. In the case of proportions data we can use Ordinary Least Squares (OLS)
regression (the usual approach) but for individual data we require specialist Maximum Likelihood
techniques. Once we have the model estimated then forecasting is simply, as before, a case of
substituting new values into the model.
Table 13.6 gives some of the data of mode shares between metropolitan areas (some of which
are on islands).
This data generated a model of the form:
Loge(Pr(Land)/Pr(Air)) = −0.04 *(Difference in Journey Times) −
0.03 *(Difference in Price)
Figure 13.12 Logistic curve and individual data
Table 13.6 Sample data on market shares and differences in systematic values
Complete Journey Time (Mins) Price (£) Share
Location Land Air Difference Train Air Difference Train Air
....................................................................................................................................
1 350 120 230 56 98 −42 0.0% 100.0%
2 280 280 0 98 98 0 50.0% 50.0%
3 35 120 −85 56 98 −42 99.1% 0.9%
4 90 80 10 88 78 10 33.2% 66.8%
5 620 134 486 25 231 −206 0.0% 100.0%
6 324 212 112 45 123 −78 10.5% 89.5%
7 350 220 130 56 98 −42 1.9% 98.1%
330
FORECASTING THE DEMAND FOR TRANSPORT
Or
Pr(Land) = exp(−0.04 *(Difference in Journey Times) − 0.03* (Difference in Price))/
(1 + exp(−0.04 *(Difference in Journey Times) − 0.03 *(Difference in Price))
Suppose current differences are 60 minutes faster by air and £100 more expensive but increased
security increases air journey times by 50 mins. The model forecasts that the current land market
share will increase from 64.6 per cent to 93.1 per cent.
Developments in choice modelling
The importance of the topic, coupled with major advances in computer power, have led to very
significant and important developments in choice modelling. The most obvious has been extension
to models with more than two choices. The simplest of these, the multinomial logistic, has a
significant limitation (known as the independence of irrelevant alternatives) and a model that
nests choices is often preferred. In the nested case a first choice might be public v private followed
by a choice between bus and train after public has been chosen. Estimates can vary significantly
depending upon the way the nests are constructed. Figure 13.13 shows two possible structures for
a comparison of road, rail, car and coach. Once again our understanding of the way people make
decisions is critical to the parameter estimates and ultimately the forecasts.
More recently developments in optimising algorithms have opened the possibility of multinomial
probits with and without subjective priors (e.g. estimates of the degree of independence
between the choices which are made by the modeller on the basis of their experience and
understanding). It must be emphasised that these developments are dependent upon many
thousands of observations of the choices made, and are most usefully found with choice experiment
data. Readers that want to find out more are recommended to read Ben-Akiva and Lerman
(1985).
Before concluding this chapter we look at a real case where most of the topics discussed come
together.
Case study 13.1 Forecasting demand for a new ferry service
– This case is based on Riddington (1996).
Background
Islay and Jura are two large adjacent island off the South West of Scotland.Although similar in
size (253 and 147 sq miles respectively) they are remarkably different in character. Islay is
Figure 13.13 Possible nesting structures for a four mode choice
331
FORECASTING THE DEMAND FOR TRANSPORT
relatively flat and fertile and sustains a substantial population of 6,500. Jura is a much wilder
island with a backbone of mountains known as the Paps. Despite its closeness to the mainland
the population is only 461, giving it one of the lowest population densities in the UK. There is
only one minor road running up the lower two thirds of the island.
Transport to Islay is based around 3 return services per day from Kennacraig on the
mainland to Port Askaig or Port Ellen on Islay, the crossing taking some 2 hours. A short ferry
from Port Askaig provides the link to Jura.
On Wednesdays the basic service is modified with the morning boat continuing on via
Colonsay to Oban and returning in the afternoon.This service aims to provide the possibility of a
day shopping trip to the mainland but clearly demonstrates the relative isolation of the islands,
7 hours on some of the roughest water in Scotland for four hours ashore. Not surprisingly
alternative routes and services have been suggested and this case is concerned with one of these,
a short sea crossing to Jura.
As can be seen from the map (Figure 13.14) the crossing from Keills to Lagg is only 5.8
miles and would take just over 20 minutes.An hourly service is thus feasible with just one vessel.
The vast majority of passengers would then proceed south through Jura to the second 15 minute
crossing to Islay. Total journey time to Islay, using the short sea crossing, would be roughly
halved.
In 1988 the government commissioned consultants to examine the economics of the proposed
service. The consultants used Trend Curve Analysis to project demand overall and an
Intention Survey of current customers to identify the proportion of passengers who would
transfer to the new service. Time Series Analysis was used to identify seasonal fluctuation to
identify the capacity that would be required.
This research forecast considerable demand for the new service. However, it was also clear
that the road system both on the mainland and on Jura would need to be significantly improved
at major cost. In addition it was claimed that the new link would be unsuitable for the heavy
lorry traffic from the distilleries and a direct service from Islay would still be required. To the
dismay of many local groups, the proposal was not continued because of the substantial road
investment required.
Figure 13.14 The Islands of Argyll in South West Scotland
332
FORECASTING THE DEMAND FOR TRANSPORT
The forecasts developed used by the consultants were the subject of some criticism and in
1996 were re-examined. In the 1996 case an Econometric Demand model was developed and
then a Choice Model to identify likely market shares.
The model of overall demand
Caledonian MacBrayne, the ferry operator, have published detailed statistics on passenger and
vehicle numbers on each of their services for many years which provided an extremely useful
dataset to establish the effect of factors such as population,and distance from Central Scotland
(the main elements of the Gravity Model).In addition prices were obtained from past brochures
and incomes from national statistics. The final model specified was
Qit = aP 1
it Y 2
t N 3
i J 4
it it
Where:
Qit represents the passenger or vehicle numbers going to island i in year t
Pit represents the charge for passenger or vehicles going to island i in year t
Yt represents the income of passengers in year t
Ni represents the resident population of island i
Jit represents the total journey time from the central belt to island i in year t
is a constant, 1, 2, 3, 4 are demand elasticities and it represents factors specific to
island i in year t.
The separate models for passengers and vehicles were estimated using ordinary least squares
and gave the following elasticities:
Tourism is an important element in demand for ferries to the islands and demand for those
whereitwasmostimportant(Arran,MullandSkye)wasunderestimated.ConverselyinIslay/Jura
passenger numbers were substantially underestimated, in part because the more spectacular
areas of Jura are currently difficult to access, whilst vehicle numbers were over-estimated. This
peculiarity is the result of competition in earlier years with a private operator which has resulted
in lower vehicle rates than usual for this length of journey. In policy terms the model suggests
higher prices will generate higher revenue.
Assuming real prices and populations are relatively constant,a forecast of demand for Islay/
Jura was generated assuming 3 different growth rates and a reduction in journey time of an
hour. This gave an increase of 29.5 per cent over estimate (fit), 50.4 per cent over actual for
passengers and 95 per cent over estimate, 54 per cent over actual for vehicles. These high
impacts correspond surprisingly well with the initial report where demands were forecast to rise
between 33 per cent and 66 per cent.
Table 13.7 Passenger and vehicle elasticities
Price Income Population Journey Time
...........................................................................................................................................................
Passengers −0.58 1.30 0.68 −0.86
Vehicles −0.87 1.91 0.78 −0.41
333
FORECASTING THE DEMAND FOR TRANSPORT
The choice model
On four islands Arran, Bute, Mull and Skye there were two ferry services, the first a lengthy
direct service to the mainland and the second a much shorter, more frequent ferry service but
with a much longer drive to and from the central belt. The proportion using the short service to
each island over time was available in the data set as was the difference in prices, journey times
and frequencies between short and long.
Because of the limitations of public transport, passengers not in vehicles had to travel on the
long routes.There was,in effect,no choice for these passengers and hence the choice model was
inappropriate.
The basic specification for the vehicles was:
Loge ΂
Pr(QSit)
Pr(QLit)΃= 1 (PSit − PLit) + 2 (FSit − FLit) + 3 (JSit − JLit) + it
Where, for each island i in year t
Pr(QSit) is the proportion of vehicles using the Short Sea route
Pr(QLit) is the proportion of vehicles using the Long Sea route
PSit is the Price of using the Short Sea route
PLit is the Price of using the Long Sea route
FSit is the frequency of the Short Sea route
FLit is the frequency of the Long Sea route
JSit is the total journey time using the Short Sea route
JLit is the total journey time using the Long Sea route
1, 2, 3 are coefficients to be estimated and it is the stochastic term
Over 97 per cent of the variance was explained by the model with all coefficients highly
significant.
It was thought that the through price to Islay using the short crossing and the Port Askaig
ferry would only be slightly cheaper than the long route but would be substantially faster and
more frequent. Applying the new journey times and frequencies suggested that just over 80 per
cent of vehicles would switch to the short crossing; identical to the Intentions Survey result of
80 per cent.
Conclusion
The econometric model suggests that the numbers travelling to Islay/Jura would increase substantially
and that the vast majority of both old and new clients would switch to the new route
even if the existing route carried on as now. Of course losing 80 per cent of the business would
put the long route into crisis and, one imagines, inevitable closure. The implications of closure
on distillery traffic and on the road system in general resulted, as stated previously, in no action.
Undoubtedly however with forecasts such as these it will be reconsidered again in the next
decade.
334
FORECASTING THE DEMAND FOR TRANSPORT
CHAPTER SUMMARY AND REFLECTION
This chapter has shown how to set about getting demand forecasts for new and existing services,
be they roads, airports, trains or boats. The methods range from simply ‘asking customers’ to
sophisticated ‘choice experiments’. There are limitations in the methods described in this chapter
and a large number of more complex versions of the methods are available. However, just as fit does
not necessarily equate to forecast quality, complexity does not necessarily equate to accuracy.
Simple methods applied consistently and intelligently are likely, in the long run, to be more
accurate.
CHAPTER EXERCISES
This chapter could end with several number-crunching exercises on forecasting; however, that is all
they would be, exercises in number crunching. The basic theme of this chapter has been that we need to
THINK when forecasting future transport trends and future demand, because it is through such
thoughts that better forecasts are produced. Exercise 13.1 thus concerns a fictional case and asks you
to consider the use of different forecasting approaches and the various data requirements necessary to
produce an accurate forecast.
Exercise 13.1
A local transport authority is interested in developing Park and Ride facilities. In the short term this
could be a dedicated bus service from an out-of-town shopping estate to the city centre. In the longer
term there is the possibility of a new ‘parkway’ station,funded by the local authority.The station would
provide both access to main-line services from the city and a park and ride facility to the city centre.
The rail operator is worried about rush hour overcrowding and wants to institute a peak hour supplement
for travellers. The local authority has extensive road and bus data on a peak/off peak basis and
can access more data from outside the region. It has some data from other parkway rail schemes.
1 For the bus service an intentions survey has been suggested
a) Suggest an appropriate sample frame
b) Discuss the strata that might be used in a stratified random sample
2 Identify how we might forecast in five years time
a) Vehicle flows at 8.30am on a Tuesday morning in October
b) Vehicle flows at 3pm on a Saturday
c) Why the month might be an important factor
3 Explain why a gravity model would be of little use in this problem
4 You are interested in building a model of the number of people who might switch to the bus
335
FORECASTING THE DEMAND FOR TRANSPORT
a) Identify the factors that you would want to include in your model
b) Discuss the associated data requirement and where such data might be found (or
obtained)
c) Explain how a choice experiment might be used to assess the comparative impact of
the price of the bus service and of bus priority lanes
5 Discuss in general how you might set about forecasting usage of the new station
a) By those going ‘out’ from the city on mainline inter-city services
b) By those using it as a park and ride facility going into the city
c) The impact of peak price supplements
336
FORECASTING THE DEMAND FOR TRANSPORT