View metadata, citation and similar papers at core.ac.

uk brought to you by CORE

provided by Aaltodoc Publication Archive

Demand Forecasting in a Railway Revenue Management

System

Economics
Master's thesis
Valtteri Helve
2015

Department of Economics
Aalto University
School of Business

Powered by TCPDF (www.tcpdf.org)

 Aalto University, P.O. BOX 11000, 00076 AALTO
www.aalto.fi
Abstract of master’s thesis

Author Valtteri Helve

Title of thesis Demand Forecasting in a Railway Revenue Management System
Degree Master’s Degree (Economics and Business Administration)
Degree programme Economics
Thesis advisor Matti Liski
Year of approval 2015 Number of pages 82 Language English

Abstract
The research in Revenue Management has tightly focused on airline markets and
somewhat neglected other similar markets. The purpose of this thesis is to offer an
extensive overview on RM in the railway context. The backgrounds and concepts of
RM are discussed and the applicability of RM in railway markets is evaluated. The
differences between railways and airlines are also explored. I am especially focused
on the demand forecasting process and different methods that can be used to forecast
uncertain demand for a specific train. I also discuss how demand forecasting relates
to other RM components, such as capacity allocation.

Relevant RM theories and demand forecasting methods are compiled based on the
existing literature. Because of the limited availability of real demand data, I use hy-
pothetical demand data to illustrate how different forecasting methods can be applied
and how the performance of each method can be evaluated. I also compile an illus-
trative capacity allocation example using EMSR –model.

I conclude that the applicability of RM in railway markets is evident. I find four sig-
nificant differences between railways and airlines that are relevant to RM. Railways
tend to have more complex networks, less price differentiation, shorter booking lead
times, and less competitive markets. Illustrative demand forecasting examples indi-
cate that the evaluation of different forecasting methods is essential, since the perfor-
mance of different methods might vary substantially, depending on the available data
and the time horizon of desired forecast. Capacity allocation examples suggest that it
is particularly important that demand forecasts would provide the accurate predic-
tions of total demand and demand distribution between fare classes. However, it
should be taken into account that the findings of illustrative examples cannot be gen-
eralized, since the hypothetical data was used in the analysis. Thus further examina-
tion with real demand data should be required. Additionally, the issues of constrained
data and network effect are omitted from the analysis.

Keywords revenue management, willingness-to-pay, price differentiation, capacity

allocation, demand forecasting, railways
Table of Contents

1 Introduction 1
2 Revenue Management 4
2.1 Background and concepts 5
2.1.1 Origin in the airline industry 5
2.1.2 Basic concepts 7
2.1.3 General conditions for using RM techniques 10
2.1.2 Background of the development 12
2.2 Theoretical development of capacity allocation models 13
2.2.1 Littlewood’s model 14
2.2.2 EMSR –models 14
2.2.3 Network control 17
2.2.4 Virtual Nesting 18
2.2.5 Bid-price method 19
2.2.6 Mathematical programming 21
2.3 Key research areas 22
2.4 RM system flow 25

3 RM in the railway industry 26

3.1 Differences between airlines and railways 28
3.2 RM applications in railways 30
3.2.1 SNCF – Socrates (France) 30
3.2.2 DBAG – PEP (Germany) 31
3.2.3 VR Group (Finland) 33

4 Demand forecasting 35
4.1 Level of forecasting 37
4.2 Historical data 39
4.3 Constrained data 39
4.4 Forecasting methods 42
4.4.1 Moving averages 45
4.4.2 Exponential smoothing 45
4.4.3 Pickup models 47
4.4.4 Regression 48
4.5 Forecast performance 49
4.6 Network forecasting 51

I
5 Illustrative examples 53
5.1 Data 53
5.2 Forecasting 56
5.2.1 Scenario 1 56
5.2.2 Scenario 2 59
5.3 Capacity allocation 61
5.4 Conclusions of examples 65

6 Conclusions 67
References 70

Appendices 75

II
List of figures
Figure 1 – Simple differential pricing model (Belobaba, 2009). 8
Figure 2 – Probability distribution of requests (Belobaba, 1987). 15
Figure 3 – Probability of selling Sth seat (left) and EMSR (right) (Belobaba, 1987). 16
Figure 4 – Airline’s hub-and-spoke network (upper) and railway network (lower). 17
Figure 5 – Simple network with three fare classes. 18
Figure 6 – Virtual classes. 19
Figure 7 – Bid price as a function of remaining capacity (Talluri & van Ryzin, 2004). 20
Figure 8 – Bid-price booking limits (Talluri & van Ryzin, 2004). 20
Figure 9 – RM process flow (Talluri & van Ryzin, 2004). 25
Figure 10 – Two dimensional railway network. 28
Figure 11 – The components of Socrates RM system. (Ben-Khedher et al., 1998). 31
Figure 12 – Development of VR’s RM and pricing. 34
Figure 13 – Forecasting process. 36
Figure 14 – Constrained booking data caused by the booking limit (Guo et al., 2012). 40
Figure 15 – Expectation of constrained demand (Zeni, 2001). 41
Figure 16 – Booking profiles (Zeni, 2001). 44
Figure 17 – Forecasting notations (remodeled from Makridakis et al., 1998). 44
Figure 18 – A simple railway network. 51
Figure 19 – Historical booking profiles. 55
Figure 20 – Scenario 1: Forecast performance. 58
Figure 21 – Scenario 2: Forecast performance. 60
Figure 22 – Protected capacity from lower fare classes (Case 1 & 2). 63
Figure 23 – Booking limits (Case 1 & 2). 63
Figure 24 – Protected capacity from lower fare classes (Case 3 & 4). 64
Figure 25 – Booking limits (Case 3 & 4). 65

List of tables
Table 1 – An illustrative booking matrix (Zeni, 2001). 42
Table 2 – Hypothetical booking history matrix (Wickham, 1995). 54
Table 3 – Division of booking matrix. 55
Table 4 – Descriptive statistics of historical bookings. 56
Table 5 – Selected forecasting methods and specifics. 56
Table 6 – Scenario 1: Data used in forecasts of Week 3 final bookings. 57
Table 7 – Scenario 1: Forecasts and errors for Week 3. 57
Table 8 – Scenario 1: Forecast performance. 58
Table 9 – Scenario 2: Data used in forecasts of Week 3 final bookings. 59
Table 10 – Scenario 2: Forecasts and errors for Week 3. 59
Table 11 – Scenario 2: Forecast performance. 60
Table 12 – Case 1 details. 61
Table 13 – Case 2 details. 61
Table 14 – Case 3 details. 62
Table 15 – Case 4 details. 62

III
1 Introduction
Railways play important role in transportation infrastructure in Europe. The market environments
in European railway markets are slowly changing in consequence of deregulatory actions by the
European Commission (European Commission, 1991) and increased competition by the transpor-
tation substitutes, e.g. intercity buses. Because of these changes, railway operators are forced to
improve the efficiency of their operation models. A business practice called Revenue Management
has significant part in these efficiency improvements.

RM has its origin in the airline industry. The extensive development of RM started in the end of
1970s after the deregulation of airline markets in United States. In consequence of the deregulation,
competition increased in the industry and compelled airlines to find new procedures in order to
survive in the changed market environment. (Talluri & van Ryzin, 2004) The answer to airlines’
problems was RM, which purpose was to enable airlines to manage their capacity as efficiently and
profitably as possible (Kimes, 1989). McGill and van Ryzin (1999) summarize that the main ob-
jective of RM is to maximize firm’s revenues. In order to maximize revenues, RM aims to answer
multiple demand management decisions, such as decisions related to pricing and customer seg-
mentations, by controlling prices and inventories. (Talluri & van Ryzin, 2004) The main concepts
behind these controls are related to differences in customers’ willingness-to-pay (WTP). Economic
concepts of product differentiation and price discrimination can be considered as the main deter-
minants of RM. (Belobaba, 1987) By way of these concepts firms are able to separate different
customer segments based on customers’ WTP and then use pricing and capacity allocation to max-
imize revenues from these customers.

Even if RM has its roots in the airline industry and the research has also strongly focused on that
specific industry, RM has its opportunities in the other industries as well. Kimes (1989) state that
RM is a tool for all capacity constrained service firms and that different RM techniques are appro-
priate when: 1) capacity is relatively fixed, 2) customers can be segmented based on their prefer-
ences, 3) product inventory is perishable, 4) products are sold in advance, 5) demand is uncertain
and fluctuating over time, and 6) marginal costs per customer are low.

But although all previous conditions would hold, how could a firm determine the most efficient
capacity allocation and pricing decisions? Like all business decisions, also RM decisions should

1
be based on accurate forecasts. Thus forecasting plays extremely important role in RM. McGill and
van Ryzin (1999) emphasize that all RM decisions are related to different forecasts. Particularly
important are accurate forecasts of customer demand, since they form the basis for capacity allo-
cation and pricing decisions. In addition to that forecasting is eminently important it is evidently
extremely difficult. Demands for each product change constantly due to exogenous factors but also
because of firm’s own actions. These changes can be uncertain, which makes it complicated to
forecast future demand. (Zeni, 2001)

As the most of the researchers studying RM and RM demand forecasting have focused on the
airline industry, the objective of this thesis is to investigate how RM and especially demand fore-
casting techniques could be applied in the railway industry. Compared to airline RM, railway RM
lacks academic research and publications. Armstrong and Meissner (2010) are one of the few who
consentrate solely on the railway RM. They give an extensive overview of railway RM and its
models.

The motivation of why the focus of this thesis is on railway RM is related to three issues. Firstly,
the impacts of RM on the airline industry have been significant. It is reported that RM applications
have increased airlines’ revenues by 4 to 6 % (Belobaba, 1987). Secondly, airline and railway
markets appear to be quite similar in many perspectives and this begs the question of whether
airline RM applications could be used in a railway context? Thirdly, there has occurred a lot of
development in railway RM in the recent years. Finnish national railway operator VR has made
serious changes in its pricing strategies. In 2013 VR adapted more demand based pricing system
as a part of wider RM investments (VR Group, 2014). Additionally, the focus of this thesis is on
demand forecasting, since it is presented to have such a significant role in RM (see e.g. McGill &
van Ryzin, 1999).

In this thesis I will attempt to give answers to four research questions:

1. What is Revenue Management?

2. How is RM applicable to railway markets?
3. How to compile accurate demand forecasts for a specific train?
4. How demand forecasting is related to capacity allocation?

The first two questions are concerned in Sections 2 and 3 based on existing literature. Based on the
literature I conclude that the opportunities of RM are evident in railway markets. Four significant

2
differences between railway and airline markets relating to the applicability of RM are recognized.
Railways tend to include more complex networks, less price differentiation, shorter booking lead
times, and less competitive markets. Each of these factors should be taken into account in the plan-
ning and implementation of RM applications.

In illustrative examples in Section 5, I show how different forecasting methods could be used to
forecast demand for a specific train. In order to show how capacity allocations are related to de-
mand forecasting, capacity allocation for hypothetical train has also been compiled. Because both
of the examples are based on the hypothetical data, any generalized conclusions cannot be drawn.
However, examples indicate that the evaluation of different forecasting methods is essential, since
the performance of different methods might differ substantially, depending on available data and
the specifics of desired forecasts. Supporting how important is the selection of suitable method, it
is presented that 10 % increase in forecast accuracy can increase revenues by even 3 % (Lee, 1990).

Examples also propose that the selection of appropriate forecasting method is impacted whether
long-term or short-term forecasts are needed. Capacity allocation example suggests that two de-
mand factors relating to demand forecasting are especially important. Firstly, demand forecasts
should provide information on the distribution of total demand between fare classes. Secondly, it
is essential to have accurate forecast of total demand itself.

The reminder of this paper is organized as follows. Section 2 gives an overview of RM. The basic
backgrounds and concepts are covered and some relevant capacity allocation theories are pre-
sented. Section 3 deals with the relationship between RM and the railway industry. Differences
between railways and airlines are discussed and a few examples of RM applications in railways are
presented. In Section 4, the emphasis is on demand forecasting and different forecasting methods
that have been used in RM. Section 5 contains illustrative examples of how demand forecasting
methods and capacity allocation can be applied. Section 6 concludes the findings.

3
2 Revenue Management
All sellers of products or services face multiple essential questions related to selling and pricing.
Which selling channel to use? How to set prices for those channels? How to segment customers?
How to differentiate products to these segments? How to set prices for different segments? How to
allocate the inventory to customer segments? How to change the price over time? All these ques-
tions relate closely to RM.

Kimes (1989) describes that RM is “a method which can help a firm sell the right inventory unit
to the right type of customer, at the right time, and for the right price”. She says that the objective
of RM is to answer these questions so that the revenue would be maximized. RM problem is then
determining how much to sell at what price, to which market segment and when. Cross (1998)
gives similar definition in economist’s terms: “a means of allocative efficiency that maximizes
economic wealth through dynamically forecasting the self-seeking activities of each individual
consumer.”

Talluri and van Ryzin (2004) present the same idea in a slightly different way. According to them,
RM focuses on demand management decisions and methodologies and procedures to implement
them. It consists of the methods of managing firm’s interactions with the market, with an objective
to maximize revenues. Additionally, Talluri and van Ryzin divide these demand management de-
cisions further into three decisions; structural, price, and quantity decisions.

Structural decisions relate to questions like: Which selling format to use? Which segmentation and
differentiation techniques apply? And what kinds of terms to include to purchases? Decisions like
these are normally quite strategic and are performed infrequently. Structural decisions can also
have effects on how a firm can react to price and quantity decisions. For example price advertising
campaigns can limit firm’s ability to adjust price decisions in short-term. (Talluri & van Ryzin,
2004)

While structural decisions are seen strategic, Talluri and van Ryzin (2004) separate two more tac-
tical decisions areas. Price decisions are related to pricing; pricing different customer groups and
pricing over time. Whereas quantity decisions try to answer questions related to the capacity con-
trols and product rationing; capacity allocation to different customer groups and whether to accept

4
or reject buying request of product in a particular price. When RM uses capacity allocation deci-
sions as the main tactical tool for managing demand decisions, it is said to be quantity-based RM
and while using prices based on these decisions, it is called price-based RM. These two distinct
strategies of RM are actually tightly connected. The quantity-based RM balances the number of
units offered available at each price, while these prices are determined by the price-based RM
decisions. So both strategies affect each other and they cannot be disentangled. Additionally, both
of these strategies are based on the differences in customers’ WTP. (Belobaba, 1987) This dichot-
omy and which of the strategies is primarily used, impacts on the theories and practices of RM.

2.1 Background and concepts

2.1.1 Origin in the airline industry

Researchers are widely agreed on the origin of the business practice nowadays called RM. The first
capacity allocation applications were developed in 1950’s when the first overbooking applications
were implemented in the airline industry. These applications allowed airlines to overbook their
flight, so that the probability of full capacity usage would be maximized. (McGill, 1989) In the
early 1970s, some airlines started to offer restricted number of discounted tickets. For example,
BOAC (predecessor of British Airways) provided low fares to passengers that booked at least 21
days prior a flight departures. However, the number of discounted tickets was not based on pas-
senger booking behavior but more on some fixed rule, like protecting some percentage of capacity.
(McGill & van Ryzin, 1999) Wider development of modern RM applications increased after the
Air Deregulation Act in 1978 in the U.S. (see e.g. Talluri & van Ryzin, 2004; Kimes, 1989; Pak &
Piersma, 2002; Walczak, Boyd, & Cramer, 2012). In 1978, the U.S. Civil Aviation Board (CAB)
revoked restrictions of airline fares and schedules. Primarily this meant that airlines were free to
set and change their prices and operate routes without approval from the CAB. Before the act, the
CAB had regulated airline ticket prices based on standardized fares and profitability targets.
(Talluri & van Ryzin, 2004)

Talluri and van Ryzin (2004) give a descriptive summary of how the deregulation act changed the
industry. Ticket pricing became more complex, since airlines were able to set and change prices
without restraints. Because scheduling was also liberalized, major airlines started to develop hub-

5
and-spoke1 networks, which made possible to extend the service networks, which in turn increased
the complexity of pricing even more. In order to control more complex pricing structure, many
larger airlines started to invest more in the development of computerized reservation systems
(CRSs) and global distribution systems (GDS)2.

Another significant change was that the liberalization of pricing enabled airlines to compete with
prices. New types of airlines, low-cost and charter airlines, entered the market. Their strategy was
to decrease fares by constructing simpler point-to-point operations, lowering labor costs and cutting
down services on flights. They directed their services to customer segments with more price elastic
buying behaviors, such as leisure travelers, students, and other irregular travelers who would oth-
erwise use another type of transportation.

The change in pricing and competitive environment affected the performance of traditional airlines.
They needed to find new strategies to recapture lost price sensitive passengers. Talluri and van
Ryzin (2004) present that one of the first solutions for this problem was presented by Robert Cran-
dall, the vice president of marketing in American Airlines at that time, who understood that an
airline faces relatively fixed costs for operating a certain flight and thus a marginal cost of selling
an additional seat is near zero. Because of this fixed costs structure, an airline could in fact increase
revenues by selling surplus tickets to more price sensitive passengers in discounted prices. The
problem was to identify these surplus tickets and also ensure that less price sensitive passengers
would not switch to buy these discounted tickets. In 1978 American Airline started to offer fixed
number of discounted tickets with purchase restrictions for each flight. Purchase restrictions were
designed to prevent less price sensitive passengers from buying discounted tickets. Discounted
tickets had to be purchased in 30 days before the departure, those were nonrefundable and required
seven days minimum stay at the destination. This is an example of one of the first modern RM
applications. Talluri and van Ryzin (2004) highlight that the link between RM and the airline in-
dustry is rather unique, since there are few business practices, which origins are so strongly con-
nected to some specific industry. (McGill & van Ryzin, 1999)

1
A configuration of an airline’s network around one or more hubs, which work as a connection points in passengers’
itinerary from origin to final destination. (McGill & van Ryzin, 1999)
2
Computer and communication system that linked the CRSs of different airlines. (McGill & van Ryzin, 1999)

6
2.1.2 Basic concepts
As can be seen from the previous example of airline RM application, RM is highly related to the
differences in customers’ WTP and other product preferences. Both quantity and price decisions
are connected to the heterogeneity of customer preferences. The economic concepts behind this are
price discrimination and product differentiation, which are both tightly related to RM applications.
(Belobaba, 2009)

Belobaba (2009) emphasizes the importance of understanding how these terms differ. Price dis-
crimination refers to the practice where a firm can sell the same product or service at different
prices to different customers based on their individual WTP (Phlips, 1983). This is indeed ex-
tremely generalized definition of price discrimination and the studies of economics commonly sep-
arate three different degrees of price discrimination. First-degree price discrimination (also known
as perfect price discrimination) refers to firm’s ability to recognize the WTP of each customer and
then correspond the prices for each customer separately according to their WTP (Phlips, 1983).
This degree is highly theoretical, because it is basically impossible to recognize customer’s indi-
vidual WTP. Third-degree discrimination occurs if a firm is able to divide customers into segments
according to their WTP and then set the prices for each segment separately (Phlips, 1983). This is
probably the most common form of price discrimination and for instance student discounts repre-
sent this degree of discrimination (Varian, 1989). Second-degree price discrimination is connected
to the product differentiation. In second-degree discrimination, a firm can identify different cus-
tomer groups and differences in WTP between these groups, but it would need to induce self-
selection of customers. In order to achieve this self-selection, a firm can differentiate its products
so that a customer would choose the product, which corresponds his or her individual preferences
and WTP. (Alderighi, 2010)

The term product differentiation indicates firm’s ability to charge different prices for products that
contain small differences (Botimer & Belobaba, 1999). Thus the product differentiation is actually
a technique, which allows a firm to induce second-degree price discrimination. Most RM applica-
tions include characteristics of second- and third-degree price discrimination. Again an example
can be seen in the airline industry. The presence of product differentiation is evident, since airlines
offer differentiated tickets for different prices. An example of product differentiation (or second-
degree price discrimination) is classification between economy and business classes. (Alderighi,
2010) At the same time, airlines set different prices for tickets with similar characteristics, using
7
for example advance purchase restrictions. These price differences cannot be explained by product
differentiation, because the products (seats) are totally similar and differences in prices are only
based on customer’s WTP for a specific product, which indicates the presence of third-degree dis-
criminatory pricing. (Belobaba, 2009; Talluri & van Ryzin, 2004) Belobaba (2009) uses the term
differential pricing to indicate RM practices that use both price discrimination and product differ-
entiation principles.

Figure 1 – Simple differential pricing model (Belobaba, 2009).

Figure 1 characterizes the basic concept of differential pricing. WTP is defined by the demand
curve. Let’s assume that a firm has a perfect ability to segment consumers according to their WTP.
So the firm can use some price discrimination method to separate consumers into two groups and
then offers different prices for these groups. Offering P1 for high WTP consumers, the firm can
expect that Q1 consumers will purchase, because their WTP is equal or greater than P1. The method
used to segment consumers is assumed to prevent the high WTP consumers to move purchasing
P2 products and thus all high WTP consumers purchase in P1. Then offering P2 for low WTP
consumers, the expected number of purchases in this lower price is Q2 – Q1, since these consum-
ers’ WTP is less than P1 but greater than P2. (Belobaba, 2009) This simple example demonstrates
how the firm can increase its revenues by adopting differential pricing methods. The expected total

8
revenue under the price differentiation is 𝑃1 ∗ 𝑄1 + 𝑃2 ∗ (𝑄2 − 𝑄1), which is evidently larger
than the expected revenue under any single price (e.g. with P1, the revenue would be P1*Q1). On
the other hand, also consumers benefit from the differential pricing. Low WTP consumers purchas-
ing P2 obviously benefit, as under single price e.g. P1, the price would be higher than their WTP
and they would not consume at all. Also P1 consumers benefit if the single price without price
differentiation would be higher than P1. (Belobaba, 2009)

This type of price differentiation is not feasible in all situations. There are few conditions that have
to be considered. Firstly, the perishability of the product relates to the applicability of price differ-
entiation. The economic concept called Coase conjecture states that if a firm sells durable good
and if consumers are able to predict price changes and are patient enough, the firm loses its power
to price differentiate and end up charging price that equals to the competitive price. (Coase, 1972)
For example, if the product in Figure 1 is completely durable and consumers predict price changes,
P1 consumers would end up to wait until the price drops to P2 and would make their purchases
then. Secondly, there should be heterogeneity in consumers’ preferences. If all consumers are ex-
actly identical the price differentiation will lose its power to generate additional revenues. Third
condition presumes that re-sales should be prevented. This can be achieved by prohibiting re-sales
with contractual manners (an airline ticket includes passenger name) or because of natural reasons
(health care services cannot be transferred to another person). The last condition relates to firms’
pricing power. Under perfect competition, a firm cannot affect the prices and thus it is not capable
of implementing price differentiation. (Talluri & van Ryzin, 2004)

Albeit differential pricing can be used to increase revenues, it is not enough to maximize revenues
alone. Both high and low WTP consumers tend to favor same products and thus under the capacity
limitations it would be preferable to sell more high priced products. Use of differential pricing
methods (e.g. advance purchase restrictions) directs the low WTP consumers to buy before the high
WTP consumers. So there is a risk that low WTP consumers will displace the high WTP consumers
and decrease the positive revenue effect of differential pricing. (Belobaba, 2009) Belobaba (1987)
points out that capacity allocation methods should be incorporated into the differential pricing in
order to maximize expected revenues. A firm should set some limits to the offered numbers of
products in each price class, so that revenues can be drawn also from the potential later-buying
high WTP consumers. Section 2.2 presents theoretical models of how these limits are set.

9
Both of these concepts, differential pricing and capacity allocation, are involved in RM (Belobaba,
1987; Kimes, 1989). In the ideal environment, a firm could use these concepts to maximize total
revenues by setting an optimal set of different prices and allocating the capacity in an optimal way
according to the demand of each price category (Belobaba, 1987).

2.1.3 General conditions for using RM techniques

Considering the descriptions of RM and the background concepts, such like differential pricing and
capacity allocation, there are several conditions for RM techniques to be appropriate. Kimes (1989)
compiles the six general conditions for using RM techniques: 1) fixed capacity, 2) customer heter-
ogeneity, 3) perishable inventory, 4) product sold in advance, 5) uncertain and fluctuating demand,
and 6) low marginal costs. However, Talluri and van Ryzin (2004) suggest that these conditions
are not completely restrictive, but would increase the effectiveness of RM.

1. Fixed capacity
Kimes (1989) states that the first condition relates to the capacity constraint. RM techniques are
relevant only if a firm cannot (at least in the short-term) adjust its capacity to correspond current
demand. Weatherford and Bodily (1992) amplify that totally fixed capacity is not necessary, but it
is essential that adding additional capacity would cause high incremental costs. For example, an
airline decides in advance what sort of an aircraft it will use in a particular flight. After the selection
of airplane type has been done, the capacity can be considered fixed. The idea is that the airplane
cannot be changed even if all seats on the flight are sold out, without significant change costs. This
condition links to the sixth condition. (Kimes, 1989)

2. Customer heterogeneity
A firm implementing RM practices should be able to divide its customers into different types ac-
cording to their preferences. These preferences can relate to WTP, preference for the different
products or the differences in purchase behavior over time. (Talluri & van Ryzin, 2004) In the
airline industry widely used manner is to segment customers using different ticket restrictions, for
example advance purchase restrictions and restrictions related to cancellation and change policies.
(Kimes, 1989). Time of purchase restrictions are particularly used in the airline industry and they
are founded on the assumption that less price-sensitive passengers book their flights closer to the
departure date (Weatherford & Bodily, 1992).

10
3. Perishable inventory
The perishability of inventory is a main factor that distinguishes service firms from many manu-
facturing firms. A relevant assumption for applying RM techniques is that the inventory perishes
after a specific time. This means that the output in one period cannot be used to fulfill the demand
in future periods (Talluri & van Ryzin, 2004). Weatherford and Bodily (1992) emphasize that if
the inventory could be stored, RM techniques would not be needed, but inventory management
approaches alone would solve the problems. For example, after an airplane departs, unsold seats
cannot be inventoried and thus they represent wasted inventory. By minimizing the inventory spoil-
age, a firm can operate more efficiently. (Kimes, 1989) This is a common characteristic for all
service providers; seats for a theater, a sport events, a restaurant or any other transportation mode
(Weatherford & Bodily, 1992).

4. Product sold in advance

As mentioned, RM systems often use the time of purchase restrictions to differentiate products. A
reservation system allows a firm to sell inventories in finite time horizon before the actual use. This
will create some security because a firm knows that some of the capacity will be used in the future.
On the other hand, it creates uncertainty, because a firm has to decide whether to accept early
buying requests, which often means trading between certain low prices and uncertain high prices.
(Kimes, 1989)

5. Uncertain and fluctuating demand

Demand for a service or product is not necessarily constant over time. Demand can vary seasonally,
weekly, daily etc. and these variations occur with some level of uncertainty. The more fluctuating
and uncertain the demand is, the more difficult the demand management becomes. Different RM
techniques can be used to estimate some of the demand fluctuation and uncertainty, which would
help firms to allocate capacity more efficiently and increase the revenue. (Kimes, 1989).

6. Low marginal costs

This condition links to the first condition, fixed capacity. For a capacity constrained firm the mar-
ginal capacity change costs are high. However for a RM system to be efficient, the marginal cost
of providing additional unit of available capacity has to be relatively low. (Kimes, 1989;
Weatherford & Bodily, 1992) Again the airline industry provides a good example; when an airline
chooses the type of an airplane for particular flight, this does not only fix the capacity but also the

11
total cost of service, because the total cost does not depend on how many passengers actually are
on a flight. (Talluri & van Ryzin, 2004)

In addition to these six conditions Talluri and van Ryzin (2004) add that a firm applying RM also
needs different information systems to collect and store data related to demand and customer be-
havior. In Section 3, I evaluate how these six conditions are fulfilled in the railway industry.

2.1.2 Background of the development

In principle, the decision problems related to RM are not new at all. Far before the deregulation in
the airline industry and the first simple capacity allocation methods, sellers have made decisions
related to RM. Talluri and van Ryzin (2004) highlight that the demand management decisions
themselves did not change after the deregulation act in 1978, but the real innovation happened in
the methods of how these decisions are made.

Significant changes in the airline industry combined with advances in science and technology cre-
ated a ground for establishment of new approaches for decisions making. Scientific advances in
economics, statistics and operations research enabled better demand models and estimates, which
further helped to compute more optimal solutions to complex decision problems. At the same time
the development of information technologies provides the capabilities to automatize transactions,
gather and store a growing amount of data, quickly solve algorithms and finally execute and control
detailed demand management decisions. (Talluri & van Ryzin, 2004)

Talluri and van Ryzin (2004) present how the advances in science and the development of infor-
mation technologies affect the decision making. One consequence is that more complex and exten-
sive demand management, which would be impossible through manual procedures, became con-
ceivable. Science and technology also improved significantly the quality of these demand manage-
ment decisions. Talluri and van Ryzin (2004) summarize that “models and systems are better at
separating market signals from market noise, evaluating complex tradeoffs and optimizing and
producing consistent decisions. The application of science and technology to demand decisions
often produces an improvement in the quality of the decisions, resulting in a significant increase
in revenues”. Many empirical studies manage to show that the proper use of RM applications can
lead to significant increases in revenues. For example in the airline industry, numerous studies find

12
that RM applications can increase revenues by 4 to 6 %. (see e.g. Belobaba, 1987; Smith, Leimkuh-
ler, & Darrow, 1992)

2.2 Theoretical development of capacity allocation models

In this section I describe in more detail some of the relevant RM theories and try to give some
perspective how RM has developed since the first theoretical applications in 1970’s. McGill and
van Ryzin (1999) list assumptions that the earliest capacity allocation models require. 1) Sequential
booking classes, which means that all booking classes are apart and do not interlace with each
other. 2) Low-before-high price booking arrival pattern, meaning that low priced tickets will be
sold before the high priced tickets. 3) There should be statistical independence of demand between
booking classes, i.e. any information on the actual demand on one price class cannot be derived
from the actual demand on another price class. 4) No cancellations or no-shows; this assumption
simply claims that it is not necessary to place any attention to overbookings. 5) The assumption,
no batch bookings, justifies to look at one booking request at a time. 6) The last assumption of the
earliest theories requires that only single-leg is considered at a time.

Additionally, demand for each fare class is assumed to be independent of the different controls set
to other classes (Talluri & van Ryzin, 2004b). This indicates that the likelihood of receiving a
booking request does not depend on the other substitutive products that are available at the time of
the booking request. In the other words, each consumer is demanding only one type of fare class at
a time.

All these assumptions create quite restrictive and unrealistic environment. However, the develop-
ment of theories enabled to relax some of these assumptions so that the new models would work
better in more realistic setting (McGill & van Ryzin, 1999). While more developed models are
more realistic, they are also more complex. The form of the assumptions demonstrates how tightly
first theories are linked to the airline industry. In this section I use the terminology relating to
airlines’ RM. Nevertheless these models can be applied in the other industries as well, especially
in the other transportation industries.

13
2.2.1 Littlewood’s model
Littlewood’s model (Littlewood, 1972) is considered the first theoretical model of quantity-based
RM (McGill & van Ryzin, 1999; Pak & Piersma, 2002; Walczak et al., 2012). The model proposes
a simple solution for airline’s seat inventory control problem, with a single-leg and two fare classes
and it requires that all six previously stated assumptions hold (McGill & van Ryzin, 1999). Little-
wood (1972) suggests that the low fare class should be closed when the expected revenue of selling
the same seat at the higher fare exceeds the certain revenue of selling another low fare seat. Thus
the constraint on accepting low fare class is:

𝑟 ≥ (1 − 𝑃) ∙ 𝑅 (2.1)
𝑟
(1 − 𝑃) ≤ (2.2)
𝑅

Where, 𝑟 is the price of low fare class and 𝑅 is the price of high fare class. 𝑃 represents the proba-
bility that the acceptance of a low fare passenger will result in the following rejection of a high fare
passenger. It is calculated from the demand distribution of high fare passengers, which is com-
monly assumed to follow normal distribution. Thus the low fare class should be accepted as long
as (1 − 𝑃) reaches the ratio of the low fare and high fare classes. (Littlewood, 1972) Although
Littlewood’s model is designed for airlines’ capacity allocation, it would be applied also in the
other industries (Talluri & van Ryzin, 2004).

2.2.2 EMSR –models

Whereas Littlewood’s model concerns the seat inventory problem with a single-leg and two fare
classes, Belobaba (1987) extends his model to include multiple fare classes in a nested reservation
system3 and presents the term EMSR – expected marginal seat revenue. Like Littlewood’s model,
also the EMSR model was at first designed for airline’s capacity allocation problem. The EMSR
model is a heuristic revenue maximization model that determines the booking limits for each fare
classes for a future flight by using the historical demand data and average fares. Belobaba (1987)
takes a probabilistic approach to capacity allocation problem. Since the demand varies stochasti-
cally over time, the future demand can be represented by a probability density function. He assumes

3
Nested reservation system refers to booking system, in which units that are available to one particular price class are
also available to all higher price classes but not the reverse. Hence the booking limit (L) defines the number of bookings
accepted in that class and all lower classes and contrary protection level (C – L) can be defined for all higher classes.

14
that the total demand for a flight follows Gaussian (normal) distribution and defines 𝑝𝑖 (𝑟𝑖 ) to be
the probability density function for the total number of requests for fare class i. The number of
seats allocated to a certain fare class is represented by si and so the cumulative probability that all
requests for a fare class will be accepted is:

𝑃𝑖 (𝑠𝑖 ) = 𝑃(𝑟𝑖 ≤ 𝑠𝑖 ) (2.3)

𝑃(𝑟𝑖 > 𝑠𝑖 ) = 1 − 𝑃𝑖 (𝑠𝑖 ) = 𝑃𝑖 (𝑠𝑖 ) (2.4)

Where 𝑃𝑖 (𝑠𝑖 ) is the probability of receiving more than si request for fare class i. Figure 2 clarifies
the relationship between 𝑃𝑖 (𝑟𝑖 ) and 𝑃𝑖 (𝑠𝑖 ).

Figure 2 – Probability distribution of requests (Belobaba, 1987).

15
The EMSR of the sith seat in fare class i can be then formulated by multiplying the average histor-
ical fare level in class i, fi, by the probability of selling at least si seats:

𝐸𝑀𝑆𝑅𝑖 (𝑠𝑖 ) = 𝑓𝑖 ∙ 𝑃𝑖 (𝑠𝑖 ) (2.5)

Figure 3 – Probability of selling Sth seat (left) and EMSR (right) (Belobaba, 1987).

Then in the general form of k nested fare classes offered, the optimal value of 𝑆𝑗𝑖 is:

𝐸𝑀𝑆𝑅𝑖 (𝑠𝑗𝑖 ) = 𝑓𝑗 (2.6)

Where i denotes a higher fare class and j is a lower fare class. Additionally, the optimal booking
limits on each fare class j can be determined by:

𝐵𝐿𝑗 = 𝐶 − ∑𝑗<𝑖 𝑠𝑗𝑖 (2.7)

This model considers reservation system with complete nesting of fare classes under the shared
inventory. (Belobaba, 1987)

Since EMSR is a heuristic model, it does not provide optimal capacity allocation unless under two
fare classes (McGill & van Ryzin, 1999). However, it is shown that the losses in revenues associ-
ated with non-optimal capacity allocation of EMSR are not significant, approximately 0.5 to 1.5
percent in the worst case (Brumelle & McGill, 1993). The EMSR model is quite simple to imple-
ment and easy to understand and thus it has been used generally in RM systems.

Later Belobaba developed EMSR model to avoid pooling effect, revised model is called EMSR-b
(the original model is known as EMSR-a) (Belobaba, 1992). Pooling effect, or statistical averaging
effect, is produced by aggregating demand across classes. The problem arises if the fare classes are

16
very close to each other. For example if all fare classes are the same, the EMSR-a model will set
too high protection levels for each class. Actually, all fare classes under identical revenues should
be aggregated. The EMSR-b model avoids this problem by aggregating demand rather than aggre-
gating protection levels. Explicitly, the demand is aggregated and treated as one class with the
weighted average revenue. (Talluri & van Ryzin, 2004)

2.2.3 Network control

After the deregulation in the airline industry, airlines started to construct hub-and-spoke networks.
This change in operation networks increased the number of passenger itineraries. The capacity
allocation became more complex and firms had to be able to manage capacities of numbers of
connecting flights over a network (Talluri & van Ryzin, 2004). The capacity allocation problem
changed from single-leg, A – B, problem to origin-destination (OD) or network problem, A – H –
B (Figure 4). Thus the route A – B actually includes the combination of A – H and H – C itineraries
and this affects the optimal capacity allocation. Littlewood’s model and EMSR –models are de-
signed for single-leg capacity allocation and therefore they are not suitable for network capacity
control (McGill & van Ryzin, 1999).

The presence of network effect is evident in other industries as well. In the hotel industry, the
problem is to control room capacity over sequential days when customers stay multiple nights.
(Talluri & van Ryzin, 2004) In the railway industry, the problem is managing the capacity between
OD with several legs. The rail track between OD generally includes numerous intermediate stations
(Figure 4) and thus the network effect is also evident in railway markets and it cannot be neglected.

Figure 4 – Airline’s hub-and-spoke network (upper) and railway network (lower).

17
The existence of network effect creates methodological but also implementation problems. As
mentioned, the capacity allocation methods for single-leg cannot be implemented efficiently in the
network context. Precise optimization in network context is actually impossible, because it creates
too large and multidimensional models. However, the development of network models has led to
variety of approximation methods. (van Ryzin & Vulcano, 2008) A few methods that are used to
network capacity allocation are presented next: Virtual nesting, bid-price method, and mathemati-
cal programming.

2.2.4 Virtual Nesting

Virtual nesting method is one of the first approximation methods related to network capacity allo-
cation and was initially developed by American Airlines (Smith et al., 1992). It is based on disin-
tegrating the network problem into a set of single-resource problems (single-leg in the airline con-
text) and thus it is referred to hybrid of network and single-resource controls. (Talluri & van Ryzin,
2004; van Ryzin & Vulcano, 2008)

The idea is to create single-resource controls, but while basic single-resource controls are based on
fare classes, the controls of virtual nesting models are based on virtual classes. These virtual classes
are constructed by connecting sets of itinerary-fare-class combinations that use a certain resource.
This step is called mapping or indexing. Network value, or net network revenue benefit, is used to
group these virtual classes. (Talluri & van Ryzin, 2004) To illustrate virtual nesting method, let’s
consider an example of OD route A –B with two legs, A – H and H – B, and three fare classes in
both legs (Figure 5).

Figure 5 – Simple network with three fare classes.

18
In order to simplify the example, fares are assumed to represent network value and they are used
to group virtual classes. Six fare classes are connected into three virtual classes based on their
values (Figure 6).

Figure 6 – Virtual classes.

These virtual classes are then used to compute booking limits for each resource using a single-
resource optimization, for example EMSR –model (Williamson, 1992). An OD booking request is
accepted if all the virtual classes are available for that OD itinerary and contrary request is rejected
if even one virtual class is closed (van Ryzin & Vulcano, 2008).

In addition to the fact that virtual nesting model produces only approximation of optimal network
capacity allocation, Talluri and van Ryzin (2004) highlight a few noteworthy drawbacks of virtual
nesting model. The use of virtual classes can cause some difficulties in data collection and fore-
casting. If the data is in the virtual class level, it might be difficult to compute real demand levels
for each resource. Virtual classes can also create practical difficulties, since interpretation of virtual
classes is significantly more complex than single-resource capacity allocation problems. Despite
these drawbacks, virtual nesting model manages to incorporate some of the network information
in RM system and affords quite effective compromise between single-resource and full network
controls.

2.2.5 Bid-price method

The first application of bid-price method has been introduced by Simpson (1989) and further ex-
amined by Williamson (1992). The difference between booking limits and bid-price controls is that
the bid-price controls are based on the marginal value of each unit of inventory. Under single-leg
problem, a booking request is accepted if its revenue is greater than the marginal value of that unit,

19
i.e. the bid price. (Williamson, 1992) A simple example of single-leg and three fare classes can be
used to describe the idea of bid-price system. (Talluri & van Ryzin, 2004)

Figure 7 – Bid price as a function of remaining capacity (Talluri & van Ryzin, 2004).

In Figure 7, the bid-price is presented as a function of the remaining capacity, while the total ca-
pacity being 100. When there are more than 60 units left, the bid price of an additional unit is less
than 50€ and thus all three fare classes (A, B and C) are available on sale. As the remaining capacity
drops below 60, but is greater than 20, the bid price increases over 50€ and thus only booking
requests for classes A and B are accepted. When the remaining capacity decreases under 20, the
bid price yet again increases, now over 100€, making only fare class A available on sale. Figure 8
represents the relationship between bid price and capacity allocation.

Figure 8 – Bid-price booking limits (Talluri & van Ryzin, 2004).

20
Although this example includes only singly-leg, bid-price method is also applicable in a network
system. In that case, bid prices are set for each resource separately and a product, including one or
more resources, is sold if its price exceeds the sum of the bid-prices related to its resources. (Talluri
& van Ryzin, 1998)

Williamson (1992) emphasizes that bid-price control is rather simple method of capacity allocation.
Only single bid price at any point of time is needed to capacity allocation, while other capacity
allocation models require booking limits for each fare class. However, in order to ensure the effi-
ciency of bid-price controls, bid prices should be updated after each sale. (Talluri & van Ryzin,
2004)

2.2.6 Mathematical programming

One of the first mathematical programming models is designed by Rothstein (1971). His model
uses dynamic mathematical programming to find a solution to the overbooking problem. Later,
Alstrup, Boas, Madsen and Vidal (1986) use mathematical programming to determine overbooking
policies for flights with two types of passengers. Also models for network system have been de-
veloped (Glover, Glover, Lorenzo, & McMillan, 1982). Albeit these mathematical programming
models give good solutions for multiple fare class allocations, they do not necessarily work well
in a nested fare class system (Kimes, 1989).

Kimes (1989) describes two simple mathematical programming models: linear programming and
probabilistic linear programming. The linear programming model is rather simple, its objective is
to maximize revenue with two constraints: 1) the sum of the total number of units allocated to each
fare class has to be less or equal to the total capacity, and 2) the number of units allocated to each
fare class has to be less or equal to the expected demand for that fare class.

𝑀𝐴𝑋 ∑𝑛𝑖=1 𝑟𝑖 𝑥𝑖 (2.8)

Constraint to:

∑𝑛𝑖=1 𝑥𝑖 ≤ 𝐶

𝑥𝑖 ≤ 𝑑𝑖 , 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑖

𝑥𝑖 , 𝑑𝑖 ≥ 0

21
Where i is a fare class, xi, the number of units sold in fare class i, di, the demand for fare class i, ri,
a given revenue from selling unit at fare i and C, the total capacity. The model assumes that the
demand is deterministic and thus it might create unrealistic solutions. (Kimes, 1989)

Another basic mathematical programming approach is the probabilistic linear programming. The
distinction between these two models is that the probabilistic linear programming model presents
demand in a probabilistic form.

𝑀𝐴𝑋 ∑𝑚 𝑛
𝑗=1 ∑𝑖=1 𝑝𝑖𝑗 𝑟𝑖𝑗 𝑥𝑖𝑗 (2.9)

Constraint to:

∑𝑚 𝑛 𝑚
𝑗=1 ∑𝑖=1 𝑥𝑖𝑗 ≤ ∑𝑗=1 𝐶𝑗

In this model i is fare class and j is inventory unit. The number of allocated units xij is described:

𝑥𝑖𝑗 = 0, 𝑖𝑓 𝑢𝑛𝑖𝑡 𝑗 𝑖𝑠 𝑛𝑜𝑡 𝑠𝑜𝑙𝑑 𝑎𝑡 𝑟𝑎𝑡𝑒 𝑖

𝑥𝑖𝑗 = 1, 𝑖𝑓 𝑢𝑛𝑖𝑡 𝑗 𝑖𝑠 𝑠𝑜𝑙𝑑 𝑎𝑡 𝑟𝑎𝑡𝑒 𝑖

The probability of selling unit j at fare i is presented by pij and the revenue from selling unit j at
fare i is rij. Again, C represents the total capacity. Like in linear programming, the problem of
nested fare structure is existed also in probabilistic form. (Kimes, 1989)

2.3 Key research areas

Research on RM is mainly focused on four main areas: forecasting, capacity allocation, overbook-
ing, and pricing. The overview of RM research by McGill and van Ryzin (1999) is used to shortly
describe these four main areas.

Forecasting
Forecasting is an important part of RM, since it forms the basis for the other RM components. For
example, demand and cancellation forecasts are used to determine capacity allocation, overbook-
ing, and pricing decisions. McGill and van Ryzin (1999) underline that conducting reliable fore-
casts is extremely difficult. Numerous factors relate to demand forecasts; demand volatility, sea-
sonality, sensitivity to pricing actions, and so on.

22
The forecasting research can be divided into three categories: macro-level, micro-level, and choice
modeling. Macro-level forecasts are large-scale, aggregated forecasts of e.g. total industry level
demand. Micro-level forecasting involves more disaggregated forecasts, for example forecasts of
passenger demand for specific flight or train. Choice modeling, alternatively, is interested in fore-
casting individual’s socioeconomic behavior. For example, forecasting individual’s choice be-
tween transportation modes relates to choice modeling. (Lee, 1990) McGill and van Ryzin (1999)
present also other factors that are related to forecasting. Firstly, it is essential to be able to derive
the demand distribution. Many empirical studies (see e.g. Belobaba, 1987) use normal distribution
as an approximation of demand distribution. Additionally, in dynamic models particularly, it is
important to have specifically determined booking patterns. The stochastic arrival is relatively used
method for constructing the distribution of total demand. Also one significant issue relating to
forecasting is the constrained data. The past demand data might be biased, because the capacity
restrictions may have limited the observed total demand. Therefore, forecasts can be downward
biased if constrained data is used to construct forecasts. (McGill & van Ryzin, 1999)

Later in Section 4, demand forecasting methods and related issues are considered more closely in
railway RM context.

Overbooking
According to McGill and van Ryzin (1999), overbooking problem is the oldest research area of
RM. The first overbooking models were constructed far before the Air Deregulation Act in 1978.
The focus of the first models was to control the probability of denied bookings. In the airline con-
text, overbooking estimations depend on the predictions of how many passengers will appear for
boarding at the time flight departures. Thus the overbooking research relates on the research of
forecasting of passengers cancellations and no-shows. The first overbooking models were static,
but later also dynamic models were developed. These models also take account of the dynamics of
cancellations and reservations. (McGill & van Ryzin, 1999)

Capacity allocation
As can be seen from Section 2.2, capacity allocation models are the main determinants of efficient
demand management. The research of these models evolved in the seventies, concentrating with
single-resource capacity allocation problem under several conditions (see e.g. Littlewood, 1972;

23
Belobaba, 1987). The following researches expanded the initial theories to include capacity allo-
cation under multiple resources, also called the network problem (see e.g. Simpson, 1989; Wil-
liamson, 1992). Since these network controls are significantly more complex, the most of these
theories provide only approximations of optimal capacity allocations (Talluri & van Ryzin, 2004).
Different mathematical programming and simulation approaches were also developed in the net-
work environment. (Pak & Piersma, 2002)

Additionally, more dynamic models were developed. The earliest models determine capacity allo-
cation controls at the start of the booking period, while dynamic models monitor the booking pro-
cess over time and adjust capacity allocation controls dynamically over the booking horizon. (Pak
& Piersma, 2002)

Pricing
Pricing is nowadays seen as a part of RM and it is clear why; price differentiation between custom-
ers is the main concept behind RM. Price is basically the most important factor of customer demand
behavior. (McGill & van Ryzin, 1999) Since price has impact on demand, the capacity allocation
decision cannot be separated from the pricing decision but these two decisions are linked together.
Therefore it is essential to RM applications to have joint pricing and capacity allocations systems.
(Gallego & van Ryzin, 1997)

There exists an extensive literature on different pricing models related to RM. Nevertheless, McGill
and van Ryzin (1999) state that the large part of the literature concerns with monopoly pricing or
price competition at industry level. For example, Borenstein and Rose (1994) study the relationship
between the level of competition and price dispersion in the airline industry. Dana (1999) as for,
constructs optimal pricing strategy for monopoly and oligopoly and shows that both include intra-
firm price dispersion. Dana’s model is highly related to price-based RM decisions. After the over-
view of McGill and van Ryzin (1999), several researches have been published relating to dynamic
and other pricing methods. (Bitran & Caldentey, 2003)

24
2.4 RM system flow
RM systems include several different components and areas of interest. So how are these compo-
nents combined and implemented in organizations? Talluri and van Ryzin (2004) represent a gen-
eral description of operations in RM system. RM system flow generally follows four steps: Data
collection, estimation and forecasting, optimization, and controls (Figure 9).

Data collection process includes collecting and storing the relevant historical data. The data should
contain information on demand, prices and other relating factors and this available data is then used
in the later operations. Next step is estimation and forecasting. Collected data is used to estimate
the parameters of the demand model. Forecasts are then compiled based on these parameters. Also
other forecasts, such as no-show and cancellation rates, can be compiled. Optimization processes
seek the optimal set of different controls (e.g. capacity allocation, prices, discounts and overbook-
ing limits) to apply until the next optimization. These controls are then used to monitor the sales of
the inventory. (Talluri & van Ryzin, 2004)

Typically RM system cycles through these steps over time. Different factors, like the volume of
data, the speed of a business cycle and the type of forecasts and optimization methods used, relate
to how frequently these steps are performed. For example, in the airline industry, RM system cycles
through these steps repeatedly and especially demand forecasts and re-optimizations should be
performed daily. (Talluri & van Ryzin, 2004)

Figure 9 – RM process flow (Talluri & van Ryzin, 2004).

25
3 RM in the railway industry
The market environment in European railway markets has been slowly changing after the reform-
ing directive by the European Commission in 1991 (European Commission, 1991). The directive
includes three cornerstone issues: 1) it separates railway infrastructure from operations so that they
are fully separated or that one operator governs both, but organizations and accounts are separated
within the operator, 2) it creates independent regulatory institutions for railways, and 3) it opens
the national railway markets for third parties. The objective of these issues is to increase competi-
tion and economic efficiency in European railway markets. (Friebel;Ivaldi;& Vibes, 2010) Albeit
the changes in market environment have been gradual and still after nearly 25 years many European
railway operators have monopoly status in national railways, the objective of improving efficiency
has increased the potential of RM in railway markets.

Research regarding RM has been intensive during the last 40 years. However, researches have
strongly focused on certain industries, especially on the airline industry. Also hotel and car rental
industries have gathered some interest, but academic research has almost entirely neglected the
railway industry. (Armstrong & Meissner, 2010) Ciancimino, Inzerillo, Lucidi and Palagi (1999)
suggest that this is perhaps due to the fact that the rail transportation does not have such a significant
role in the U.S., where flying is considered as the main mode of intercity transportation. Because
RM started developing in the U.S. after the Air Deregulation Act, it is understandable that the main
focus of the research has been on the airline industry.

Nevertheless, in Europe and Asia railways are widely used mode of transportation between cities
(Ciancimino et al., 1999). Especially in Central Europe and Japan high-speed trains play crucial
role in medium long intercity markets. Thus the most of the few researches considering RM in the
railway industry have been published in Europe or Asia. Armstrong and Meissner (2010) compile
an overview of railway RM and its models. They emphasize that the opportunities of RM in the
railway industry are similar to those found in the airline industry. Ciancimino et al (1999) develop
a capacity allocation model, which directly concentrates on the railway RM problem. They con-
sider a single-fare and multi-leg capacity allocation problem. Bharill and Rangaraj (2008) study
RM in the context of Indian Railways. Also the applicability of RM to high-speed railways has
been studied by a couple of researches (see e.g. Wang;Lan;& Zhang, 2012 and Chuang;Chu;& Niu,

26
2010). Kimes (1989) gives a generalized description of RM and underline that it can be used by
any capacity-constrained service firm.

In Section 2.1.3, six general conditions for using RM techniques are presented. Taking account of
these conditions, it seems to be evident that RM has great opportunities in the railway industry, as
Armstrong and Meissner (2010) present. The first condition, fixed capacity, seems to be appropri-
ate assumption in railway context for two reasons. First, like airlines, also railways tend to schedule
their timetables for long period and thus the capacity is fixed in short-term. Additionally, the length
of platforms on each route causes capacity limitations also in the long-term, because additional
cars cannot be added if the total length of the train exceeds the length of the shortest platform on
the route. The second condition, customer heterogeneity, seems to be fulfilled partially in the rail-
way industry. It is evident that railway passengers can be divided into different segments based on
their WTP, but the passenger mix is shown to vary significantly between routes. The high share of
business passengers is focused on particular routes, while on some routes the passenger base might
include exclusively leisure passengers. Thus the applicability of RM depends on the route specifics.
(Talluri & van Ryzin, 2004) The third condition, perishable inventory, is as obvious as it is in the
airline industry and the other service industries. After a train has departed, unsold seats cannot be
inventoried and therefore they represent a wasted inventory. The fourth condition, product sold in
advance, has become more relevant in the recent years, because of the development of reservation
systems but, still, this condition does not hold perfectly in the railway industry. Armstrong and
Meissner (2010) remind that many railway passengers tend to buy their tickets just before the de-
parture from the station. Finally, both fifth and sixth conditions, uncertain and fluctuating demand
and low marginal costs, can be seen to hold in the railway industry. Demand can fluctuate season-
ally, weekly or daily and part of these fluctuations can be uncertain. The condition of low marginal
costs holds because the cost of providing a certain train does not depend on the number of passen-
gers on that train, assuming that the condition of fixed capacity holds. All in all it can be generalized
that the common conditions for RM techniques hold reasonably well in the railway industry.

27
3.1 Differences between airlines and railways
Though the railway and airline industries can appear pretty similar from the perspective of RM, it
is crucial to also view the fundamental differences between these industries (Armstrong &
Meissner, 2010). This section presents four significant differences that relates to RM and have to
be considered if implementation of airline RM applications to the railway industry is planned.

More network oriented

The network problem in the railway industry is two dimensional and thus even more complex than
networks in the airline industry. As Talluri and van Ryzin (2004) note, the railway network can be
considered as a combination of hotels’ and airlines’ networks. Single OD route includes several
intermediate stations and is then constructed by several legs. This OD route network corresponds
to the length of stay network of hotels. While hotel revenue managers are interested in how many
nights customers are staying, railway revenue managers have to consider how many legs each pas-
senger is going to travel.

On the other hand if the whole railway network is considered instead of a single OD route, the
network resembles more airline’s hub-and-spoke network. The railway network contains many
connection points where passengers can switch trains making it like a hub-and-spoke network.
(Talluri & van Ryzin, 2004) Figure 10 illustrates two dimensionality of railway network.

Figure 10 – Two dimensional railway network.

28
Less price differentiation
Another difference is the level of fare class differentiation used in RM applications. In general,
railway operators tend to use less fare classes and advance purchase restrictions than airlines do.
Like mentioned earlier, the passenger mix tend to vary more between railway routes than it varies
between airline routes. Also the amount of business passengers can be non-existent on some rail-
way routes. Thus many tourist-oriented railway markets rely on product differentiation (e.g. 1st
class, 2nd class etc.) and identifiable customer groups (e.g. student discounts) instead of using price
differentiation in a form of multiple fare classes and advance purchase restrictions. (Talluri & van
Ryzin, 2004)

Shorter booking lead times

While people are used to booking and buying their airline tickets way before the departure, the
convention is still slightly different in railway markets. The number of so called walk up passengers
is significantly greater. This means that many passengers are likely to purchase their tickets from
the station just before the departure. (Armstrong & Meissner, 2010) This holds especially for
shorter routes and business passengers.

Less competitive markets

The fourth significant difference relates to the market environment. After the Air Deregulation Act
in the U.S. and similar deregulating actions in 1990s in Europe, the airline markets have become
highly competitive. On the contrary, despite similar deregulatory actions in railway industry, in
many countries railways are still monopolized by the governments. Therefore they do not face
direct price competition and thus they might have greater flexibility in pricing than airlines. How-
ever, also monopoly railway operators have to compete with other modes of transportation like
airlines and intercity bus operators. Consequently, even monopoly railway operator’s pricing and
capacity allocation decisions are affected by the prices of transportation substitutes. (Talluri & van
Ryzin, 2004)

29
3.2 RM applications in railways
This section presents a few examples of how RM has been used by railway operators. First I de-
scribe RM system used by the French national railway operator SNCF (Société Nationale des
Chemins de Fer Français). Then a description of RM scheme of Deutsche Bahn AG (DBAG) is
given. Thirdly I will shortly discuss on how RM has affected on Finnish national railway operator,
VR. These introductions are fairly brief and superficial because of two reasons. First, as mentioned,
there is not much published research on RM in railways, which limits the available information.
Another reason relates to the natural reluctance of railway operators to share details of their RM
applications to the public.

3.2.1 SNCF – Socrates (France)

SNCF is a public monopoly owned by the French government. It was one of the first railway op-
erators in the world that introduced some revisions to the traditional distance based pricing in
1970s. SNCF started to offer lower fares to trains with low utilization rates. The system was based
on a tricolor year calendar with three time zones: red for peak periods, blue for average periods and
white for low demand periods. The idea was to direct price sensitive passengers to use lower-priced
“white trains”. This year calendar was published and revised only once a year so the system was
not able to dynamically adjust to changes in the markets. (Mitev, 2004) Even though the system
was extremely simple, the ideology behind it was truly RM driven.

Later in 1980s SNCF expanded price differentiation by starting to use compulsory and chargeable
reservations for some special routes. Prices also varied based on the travel time so that faster high-
speed trains (TGV) were more expensive than slower intercity trains. (Mitev, 2004) However, the
big revolution in SNCF’s pricing happened in 1993 when it introduced a new computerized reser-
vation and ticketing system, called Socrates. Socrates was created in partnership with SABRE
Technology Solutions, which has earlier created RM systems for American Airlines. (Ben-
Khedher, Kintanar, Queille, & Stripling, 1998) The new reservation system enabled to incorporate
more parameters into the pricing model. For example type of the purchase and flexibility of the
product have been included into the model and thus they also have impacts on the prices. The new
system included both pricing and capacity allocation controls. The pricing model also became more
dynamic, since Socrates enabled gathering new information on bookings in real time and then ad-
just prices and capacity allocations based on new information. (Mitev, 2004) Socrates system was
30
divided into two parts: a long-term strategic schedule planning system, RailPlus, and a short-term
tactical capacity allocation system, RailCap (Ben-Khedher et al., 1998). Figure 11 illustrates the
relationship between RailCap and other RM components in Socrates system.

Figure 11 – The components of Socrates RM system. (Ben-Khedher et al., 1998).

The first steps after the introductions of Socrates system were catastrophic. So many issues went
wrong: failed bookings, impossible reservations on some trains, wrong train connections, and in-
accurate traffic information led to long queues of disappointed passengers in all main stations in
France. Also sales staff resisted the new system because it was considered too complicated and
they felt that the training had been inadequate. (Mitev, 1996) However, despite of the initial diffi-
culties, the Socrates system is estimated to provide incremental revenues of 17 million euros per
year and it has also shown to substantially decrease operating costs (Ben-Khedher et al., 1998).

3.2.2 DBAG – PEP (Germany)

In December 2002 Deutsche Bahn AG (DBAG) introduced a new fare system, called PEP, for its
long-distance passenger routes. The PEP system can be considered as an example of unsuccessful
implementation of RM system. After implementation Germans protested strongly against new
fares, which lead to such a considerable drop in economic performance that DBAG had to rectify
the system only 10 months after the introduction. (Link, 2004)

The PEP fare system was a traditional RM system with goals of maximizing profits and enhances
capacity usage. It included four main pricing elements: a base fare, early booking discounts,
BahnCard discounts and discounts for accompanying persons. In contrast to the traditional distance

31
based fare structure, new base fare was based on the demand and was also differentiated by train
type and class. Early booking discounts had been divided into three levels; a 40 % discount for
booking one week in advance, a 25 % discount for booking three days in advance, and a 10 %
discount if the booking was done one day before the departure. BahnCard was chargeable member
card, which entitled to get a 25 % discount on all tickets. Discounts for accompanying persons
meant that small groups were allowed to get a 50% discount on the base fare. Additionally to these
pricing changes, new system included high cancellation and rebooking penalties for early bookers.
(Link, 2004)

But as mentioned, the PEP system was not success. Link (2004) presents a list of matters that might
have gone wrong in the PEP system. First of all, PEP’s advanced booking discounts were not
enough to shift passengers from peak trains to low-demand trains, but also specialized peak-load
pricing should have been implemented. Secondly, Link states that even if PEP managed to segment
passengers based on their WTP, the system overlooked the issue that passengers might make a
general long-term decisions between modes of transportation instead of only considering trip spe-
cific decisions. Also complexity of railway network created highly complicated calculations of
different optimal ticket restrictions. DBAG has assumed that there are 22 million possible route
combinations in German railway network. Additionally, DBAG had problems with handling com-
plex fare structures and transparency and passenger friendliness of the system was questioned.
Lastly, Link emphasizes that in a way PEP reduced flexibility of rail transportation, which can be
considered a major advantage of railways compared to airlines.

The PEP fare system is an excellent example of why differences between railways and airlines
have to be inspected carefully when implementations of airline RM applications are considered in
railway industry.

32
3.2.3 VR Group (Finland)
VR Group is a limited company, which is completely owned by the Finnish government. The group
comprises three main business areas: passenger traffic (VR), logistics (VR Transpoint) and infra-
structure (VR Tracks). VR is responsible for operating passenger rail transportation in Finland. It
operates approximately 300 long distance trains and 900 commuter trains daily. (VR Group, 2015)

Like the other European railway operators also VR has felt pressure to respond to changing market
and demand environments. In autumn 2011 VR implemented new ticket selling system, which had
been designed to enable demand based pricing in the future (VR Group, 2012). In the following
year, VR continued to revise its pricing systems. The objective of the reform was to revise both
pricing structure and selling channels. In February 2012 first discount campaigns were launched.
Campaigns included different types of discounts, for example “two for one” and “even prices”. VR
presents in the annual report that new pricing systems increased the sales of railway tickets in 2012.
(VR Group, 2013) The change of pricing systems did not happen without problems. Serious tech-
nical and implementation problems led to dissatisfied passengers, but the firm reports that those
problems were fixed during the spring 2012 (VR Group, 2013).

Even greater changes in pricing structure occurred in 2013, when VR started to use demand based
pricing. Also mobile store was opened, which increased the share of self-service channels in ticket
sales. (VR Group, 2014) In 2014, the competitive environment tightened in a consequence of new
“low-cost operators” in the long-distance bus markets. VR responded by increasing price cam-
paigns targeted to the price sensitive passengers, for example, students got additional 50 % discount
on already discounted fares. Also investigations of new pricing reform were started in 2014. It is
stated in the annual report that VR partly failed to prepare to these changes in competition. Because
of the competition and more price-conscious consumers, it is emphasized that the pricing system
should be more transparent and customer-oriented. That’s why VR will continue to develop its
pricing so that the differences between different fare classes would be more explicit. Like in the
previous years, also in 2014 the share of self-service channels continued to grow. In the end of
2014 the share of self-service channels in ticket sales was 61.3 %, which was the highest share in
VR’s history. (VR Group, 2015)

33
Because of the limited availability of information, it is not possible to draw more precise conclu-
sions on VR’s RM systems. However, the latest annual reports indicate that VR has made remark-
able investments in the development of RM systems in the last five years and that this investment
has paid off. Figure 12 concludes the development of RM and pricing in VR since 2011.

Figure 12 – Development of VR’s RM and pricing.

34
4 Demand forecasting
This section describes the process of demand forecasting in a RM system. Some relevant methods
of demand forecasting are presented. Additionally, important issues relating demand forecasting
are considered. The emphasize is on disaggregated micro-level forecasting, because the purpose of
this thesis is to study how to construct demand forecasts for each fare class on a specific train,
which can be then used to determine efficient capacity allocation and pricing decisions.

Lee (1990) presents that the goal of airline’s micro-level demand forecasting process is to provide
reliable and accurate predictions of future demand on each date, flight and fare class for capacity
allocation and pricing optimizations. Thus, because of the industry similarities, the demand fore-
casting in the railway industry can be described correspondingly: the future demand for each train
departure and each fare class has to be forecasted in order to produce effective capacity allocation
and pricing optimizations.

The purpose of this sort of forecasting is to provide some quantitative predictions of the future
uncertain demand. Therefore, three general conditions for applying quantitative forecasting are
required: 1) information about the past must be available, 2) this information can be presented in
numerical form, and 3) this information of the past must represent the future in some level. The
third condition is known as the assumption of continuity and it forms the basis for all quantitative
forecasting methods. (Makridakis, Wheelwright, & Hyndman, 1998)

As McGill and van Ryzin (1999) present, forecasting passenger demand includes numbers of com-
plicating factors, which make it extremely difficult to accurately forecast upcoming demand. Zeni
(2001) lists some significant factors contributing to this difficulty.

Demand fluctuation and peaks

Demand for a certain travel route varies over time. This variation of demand includes variation
between seasons, day of week, time of day, and also because of the special events. Seasonal varia-
tion features variation in demand due to different time of year. For example, there is greater demand
for air and rail transportation to Lapland during the winter holiday season. Day of week variation
arises because business travelers are more likely to travel on weekdays. Business travelers also
prefer early morning and afternoon departures, which generates variation on time of day demand.

35
Special events, such as rock concerts, might create significant temporary increases in demand for
a specific route.

Sensitivity to pricing actions

Because of the price elasticity of passenger demand, price changes result in shifts in demand. While
it can be generalized that a price increase will lower the demand, the level of demand change due
to price change differ between different passenger types.

Demand dependencies between fare classes

A passenger who meets the restrictions of low fare ticket, can be forced to buy high fare ticket
because there is no available low fare tickets due to the booking limit of that class. On the contrary,
a passenger who books the low fare ticket might also have been willing to buy higher fare ticket,
if the low fare ticket had not been available.

Additionally, several other factors, such as booking cancellations, no-shows, group bookings etc.,
have their own effects on demand forecasts. Thus establishing accurate demand forecasts is cer-
tainly complicated. (Zeni, 2001)

Figure 13 – Forecasting process.

The forecasting process can be divided into four steps (Figure 13). The first step consists of the
available historical data. Next, the data has to be modified by unconstraining censored observa-
tions. Then the actual forecasts can be constructed applying suitable forecasting model. Finally,
the forecast performance has to be evaluated, so that the capacity allocation and pricing decisions

36
would be based on accurate and realistic forecasts. After the capacity allocations have been deter-
mined and new information on demand has been attained, the process is repeated over again. Thus
forecasting process, similar to RM process, is cyclical process that cycles through these four steps
over time.

4.1 Level of forecasting

As presented, in this thesis my focus is on the disaggregated demand forecasting. This is a common
practice in RM applications and Weatherford and Kimes (2003) state that the majority of the mod-
ern RM systems rely on disaggregated forecasts. Additionally, these disaggregated forecasts are
seen to perform better than general aggregated forecasts, which are then divided into the disaggre-
gated level (Weatherford & Kimes, 2003). However, also the other levels of forecasting are some-
what relevant in the context of RM and thus the next two sub-sections give a short presentation of
macro-level and passenger choice forecasting. After that the following sections consider solely the
forecasting process related to the micro-level demand forecasting.

Macro-level (aggregated level)

The broadest level of forecasting is macro-level or aggregated forecasting. In demand forecasting
these forecasts construct predictions of total industry level demand. In the transportation industry,
the groundbreaking studies of macro-level forecasting are presented by Kanafani (1983) and Taneja
(1978) (Lee, 1990). Taneja (1978) provides an extensive study of airline traffic forecasting. His
emphasize is on forecasting total airline traffic and future national traffic growth. Kanafani (1983)
as for gives overview of transportation demand theories.

Macro-level forecasts have been also compiled in the railway industry. For example, Rao (1978)
constructs a forecasting system for railway freight services. The model predicts the impacts of
macro-economic activity and intermodal competition on rail freight demand. Wardman (2006)
compiles a model for forecasting aggregated passenger railway demand. He tries to explain the
growth of railway demand in the 1990s in the United Kingdom and comes to the conclusion that
the most significant demand stimulator is the GDP growth.

Passenger choice models

The most disaggregated type of demand forecasting is related to passenger behavior. Passenger
choice models are interested in forecasting demand based on passengers’ choices between available

37
alternatives (competitors’ products and different transportation modes). (Lee, 1990) Ben-Akiva
and Lerman (1985) present a general description of a discrete choice modeling in transportation
industry. Their general choice process is based on four steps. At first, individual decision-maker
has to be defined. It is relevant to be able to explain the variation of preferences between different
decision-makers and thus characteristics like income, age, gender, education, etc. have to be in-
cluded in the model. The next step is to separate different alternatives that are available for the
decision-maker. When analyzing individual’s decision making, it is relevant to not only know what
has been chosen, but also the alternatives that could have been chosen. These alternatives include
all alternatives that can be considered as substitutes. For example, a decision maker planning to
travel from A to B has to choose whether to use car, bus, train or airplane. These transportation
means additionally include numerous options. Decision maker must also decide which company to
use if competitive alternatives are available and also which time to depart on what day. All these
available alternatives are characterized by different attributes, which represent the costs and bene-
fits of each alternative to the decision maker. (Ben-Akiva & Lerman, 1985) In this example the
airplane would probably be the fastest alternative but also most likely the most expensive. Moreo-
ver some consumers would prefer travelling in early morning while some other would prefer late
night departure. Decision maker compares the costs and benefits of different attributes of available
alternatives and determines which alternative would maximize her or his utility. After these three
steps are determined, a decision rule can be applied. The decision rule describes how the decision
maker makes the choice between alternatives. (Ben-Akiva & Lerman, 1985) Thus the discrete
choice model predicts a decision maker’s choice between different alternatives based on the utility
of each alternative (Rasouli & Timmermans, 2012).

Talluri and van Ryzin (2004b) are one of the first who develop a methodology to analyze passenger
choices in a single-leg RM problem. They create a general choice model that determines the prob-
ability of purchase of each fare class as a function of all available fare classes. Thus the assumption
of independency between fare classes is not needed, making the model more realistic. The initial
model does not take account of substitutive products, but Talluri and van Ryzin (2004b) emphasize
that their approach could be revised so that the analyze of passenger choices between substitutes
would be possible. Then the information regarding both firm’s own products and competitors’
products should be included into choice set model.

38
4.2 Historical data
Data is the heart of a forecasting system. Therefore, the first step of generating any sort of forecasts
is to identify what sort of data is available and from which sources. The most used data to construct
demand forecasts is historical sales or booking data. This data covers information on previously
occurred sales and bookings. Often historical booking data is relatively easily available and thus it
creates efficient data collection, estimation and forecasts processes. (Talluri & van Ryzin, 2004)
The available historical booking data can be divided into two types: complete and partial data. The
complete booking data includes information on bookings that have been realized. (Lee, 1990) For
example in the railway context, this type of data includes information on the bookings of a train
that has already departed. Alternatively, the partial booking data is comprised of booking infor-
mation on trains that have not yet departed. Thus partial booking curve is constructed by the de-
mand data available at the current time, t days before departure. (Lee, 1990) Talluri and van Ryzin
(2004) note that partial booking data can be for instance used to forecast the short-term increments
of demand for each booking day.

Besides historical booking data, also other data sources are often available for a firm. Lee (1990)
mentions that for example additional information about seasonality can offer a tool for a firm to
account the seasonal variation in the bookings. Seasonality can be addressed by the seasonal index,
which describes the variation of the demand of a particular day from the average day. RM databases
also gather information on the booking process itself. Information on past prices, overbookings,
bid prices, promotions and when each fare classes was closed, are commonly recorded. Different
industry level factors, such as the availability of competitors’ bookings and prices, can also give
additional information to forecasting. (Talluri & van Ryzin, 2004)

4.3 Constrained data

RM applications use different capacity allocation methods to optimize revenues of the inventory.
These methods are based on controlling the available capacity of each fare class. For making these
capacity allocations firms have to construct forecasts of the future demand. The problem of con-
strained data arises, since these forecasts are generally based on the historical booking data, which
might be constrained because of the previous capacity controls (Figure 14). Weatherford and Pölt
(2002) explain that booking requests can be rejected because of the booking limits of each fare

39
class and thus the historical booking data might not reflect the real demand for each class. After a
certain booking class is closed, the additional information of demand for this class is lost. In statis-
tics, this problem is called constrained or censored data (Guo, Xiao, & Li, 2012).

Figure 14 – Constrained booking data caused by the booking limit (Guo et al., 2012).

If a demand forecast is generated using constrained data, it is probable that the forecast would give
underestimated predictions, because it does not take account of the demand information after the
booking class is closed (Guo et al., 2012). According to Weatherford and Belobaba (2002), under-
estimating demand can result in significant losses in revenue. They estimate that underestimating
demand by 12.5 – 25 % can lead to a loss of revenue from 1 – 3 %. Additionally, because of the
relationship between forecasts and capacity controls, the constrained data might create a spiral-
down effect on the total revenues. This spiral-down effect means that if underestimated forecasts
have been used repeatedly to construct the capacity controls, revenues will systematically decrease
over time. (Cooper, Homem-de-Mello, & Kleywegt, 2006)

To prevent this problem, the constrained data can be extrapolated to match the true demand distri-
bution. This process is termed demand unconstraining. (Weatherford & Pölt, 2002; Guo et al.,
2012) Demand unconstraining has been seen to have significant impacts on demand forecasting
accuracy and consequently on revenues. Weatherford and Plöt (2002) report that the demand un-
constraining process is related to 2 – 12 % increase in revenues in the airline context.

40
Several mathematical methods for demand unconstraining have been proposed and various studies
have evaluated the performance of these methods (see e.g. Weatherford & Pölt, 2002). A widely
used mathematical method is Expected Maximization (EM), which is especially used in quantity-
based RM (Talluri & van Ryzin, 2004). It was first applied to unconstraining censored demand by
Salch (1997).

Expected Maximization (EM)

The EM model is an iterative model including two steps. In the first step, the invisible demand
observations are replaced with the estimated values of the true demand. This estimation is done
using a maximum likelihood estimation. The second step is to re-estimate the original probability
distribution of demand. These steps are then repeated until convergence. (Weatherford & Pölt,
2002) Salch (1997) assumes that the data is normally distributed 𝑁(𝜇, 𝜎 2 ) and the capacity control,
C, determines the bound of unconstrained and constrained observations. Then the best estimate of
the unconstrained value U is determined by 𝐸(𝑋|𝑋 > 𝐶). Figure 15 illustrates the expectations of
constrained demand. Zeni (2001) gives an extensive description of the EM model in his doctoral
thesis.

Figure 15 – Expectation of constrained demand (Zeni, 2001).

41
4.4 Forecasting methods
This section describes some of the commonly used micro-level demand forecasting models in pas-
senger transportation. Like the research of RM in general, also the research of demand forecasting
in transportation industry is strongly focused on the air travel demand (see e.g. Sa, 1987; Lee, 1990;
Weatherford & Belobaba, 2002). However, later researches have also been conducted solely in the
railway industry, for example Tsai, Lee and Wei (2005) and Sharif Azadeh (2013) in his doctoral
thesis, study demand forecasting in railway markets.

In a railway RM system, the objective of demand forecasting is to construct predictions of future

total demand or bookings for each departure time, given the partial and complete historical demand
data. This data can be illustrated by an imaginary booking matrix (Table 1).

Week DB0 DB7 DB14 DB21 DB60 DB90 DB180

-4 100 90 70 55 25 10 0

-3 97 85 65 50 23 9 0

-2 94 80 60 45 21 8 0

-1 91 77 57 42 19 7 0

0 88 73 54 39 17 6 0

1 - 70 50 35 15 5 0

2 - - 45 30 13 4 0

3 - - - 25 11 3 0

Table 1 – An illustrative booking matrix (Zeni, 2001).

Table 1 describes the booking history of a route with a single fare class over an eight-week period.
The negative week numbers present past departures, while the positive week numbers are future
departures. Week 0 refers to the last departure. The DB columns present the current bookings at
certain days before the departure. Hence for example, the three weeks ago departed train had 65
bookings on hand 14 days before the departure.

DP 0 indicates the final number of bookings at the departure time. The booking data of future trains
(week numbers 1, 2, and 3) is partial, because of the missing values of the booking matrix.

42
Forecasting methods can be divided into two types depending on which data in the booking matrix
they use to perform forecasts. Historical models use the booking data of previous departures to
forecast future bookings (Lee, 1990). These are time series models and the objective of these mod-
els is to recognize the patterns in the historical data and extrapolate that pattern into the future
demand. For example in Table 1, a historical model would use booking data in column DB 0 to
forecast future total bookings of Week 1. Alternatively, advanced models determine demand fore-
casts using both historical and advanced booking data, the rows of the booking matrix (Lee, 1990).
In Table 1, an advanced model would use historical booking data and advanced booking data in
row Week 1 to forecasts total bookings of Week 1 train at the time of departure.

The most of the models used in transportation demand forecasting do not attempt to detect any
causal factors that have impacts on the demand. The demand system is treated as a black box from
which forecasting models attempt to recognize the patterns for future forecasts (Makridakis et al.,
1998). Two reasons for using this type of forecasting methods are presented by Makridakis et al.
(1998). First, as previously noted, the demand is impacted by several different factors. This multi-
dimensionality creates such a complex system that it might be impossible to measure and separate
the impacts of different factors. Secondly, the main objective of short-term demand forecasting is
to know what the demand will be, rather than knowing which factors affect the demand. The long-
term forecasting is more interested in identifying the underlying factors.

Based on the studies of Zeni (2001) and Weatherford, Gentry and Wilamowski (2002), there are
four forecasting model groups that are particularly used in transportation demand forecasting.
These groups are moving averages, exponential smoothing, pickup, and regression models. These
models are considered as traditional forecasting models and the extensive descriptions can be found
from any forecasting handbook (see e.g. Makridakis et al., 1998). Moving average and exponential
smoothing models are pure historical time series models, whereas regression and pickup models
can be considered as advanced models.

Booking histories can also be presented in graphical forms, which are called booking profiles (Fig-
ure 16). Each curve in booking profile reflects the advance of bookings of each departure. The
booking curves tend to increase as the departure date approach, but they might also decrease be-
cause of the booking cancellations. (Zeni, 2001)

43
 Figure 16 – Booking profiles (Zeni, 2001).

Figure 17 defines the notations used in the following descriptions of different methods.

Figure 17 – Forecasting notations (remodeled from Makridakis et al., 1998).

44
4.4.1 Moving averages
Weatherford et al. (2002) use two different moving average models in their comparative analysis:
simple and weighted moving averages. Both models are based on the assumption that the demands
of recent departures are good predictors for the demand of future departure.

Simple moving average:

1
F𝑡+1 = 𝑛 ∑𝑡−𝑛+1
𝑡 𝑋𝑡 (4.1)

The term F𝑡+1 is the forecasted demand for the next period. Values of 𝑋𝑡 represent the observed
values of demand, up to a given number of n historical periods. Determining the appropriate num-
ber of n is the key decision in using moving average model. (Weatherford et al., 2002) Makridakis
et al. (1998) emphasize that the large number of n increases the likelihood that the model will
eliminate the randomness from the forecast. On the other hand, the model will also waste infor-
mation regarding the seasonality and trends, when using large number of n.

Weighted moving average:

The weighted moving average model emphasizes more recent demand over more distant demand.
The weights are set so that they equal to one.

𝐹𝑡+1 = 𝑊1 𝑋𝑡 + 𝑊2 𝑋𝑡−1 + 𝑊3 𝑋𝑡−2 … 𝑊𝑁 𝑋𝑡−𝑁+1 (4.2)

The model is similar than the simple moving average model, but instead of averaging by the num-
bers of n, the averaging is done by weighting each historical demand. The sum of these weights
equals one and generally more weight is put on more recent observations. The advantage of using
weighted average is that it results in smoother trend-cycle than the simple moving average model.
(Weatherford et al., 2002)

4.4.2 Exponential smoothing

Exponential smoothing model is similar to weighted moving average model in the sense that it puts
less weight on more distant demand observations. Only a small amount of data is required to com-
pile forecasts using exponential smoothing model, since it is based only on the most recent demand
observation and forecast. It is also simple to implement and thus it has its advantages over moving
average methods. (Zeni, 2001; Weatherford et al., 2002) Three different exponential smoothing
models are described next.

45
Simple exponential smoothing:
𝐹𝑡+1 = 𝛼𝑋𝑡 + (1 − 𝛼)𝐹𝑡 (4.3)

The forecast 𝐹𝑡+1 is compiled by weighting the most recent demand observation, 𝑋𝑡 , with a smooth-
ing constant, α, and the most recent forecast, 𝐹𝑡 , with (1 – α). The smoothing constant is set between
zero and one. (Makridakis et al., 1998)

Exponential smoothing with trend: Holt’s method

Holt first presented exponential smoothing with trend component in 1957. (Holt, 2004) The model
is an extended version of simple exponential smoothing method:

𝑇𝑡 = 𝛽(𝐹𝑡 − 𝐹𝑡−1 ) + (1 − 𝛽)𝑇𝑡−1 (4.4)

𝐹𝑡+1 = 𝛼𝑋𝑡 + (1 − 𝛼)𝐹𝑡 + 𝑇𝑡 (4.5)

Equation 4.4 defines the trend component. The value of β represents the smoothing constant of the
trend. Then Equation 4.5 provides a complete formula for exponential smoothing with trend. (Ma-
kridakis et al., 1998)

Exponential smoothing with trend and seasonality: Holt-Winter’s method

Winters (1960) extends further Holt’s model to capture seasonality. The model is similar to simple
exponential smoothing and Holt’s models, but it also includes seasonal component:

𝑆𝑡 = 𝛾(𝑋𝑡 − 𝐹𝑡 ) + (1 − 𝛾)𝑆𝑡−𝑠 (4.6)

𝑇𝑡 = 𝛽(𝐹𝑡 − 𝐹𝑡−1 ) + (1 − 𝛽)𝑇𝑡−1 (4.7)

𝐹𝑡+1 = 𝛼𝑋𝑡 + (1 − 𝛼)𝐹𝑡 + 𝑇𝑡 + 𝑆𝑡 (4.8)

Equation 4.6 defines the seasonal component, the lowercase s represents the length of seasonality
and γ is the smoothing constant of seasonality. Like in the Holt’s model, Equation 4.7 provides a
complete formula for the forecast. (Makridakis et al., 1998)

The choice of smoothing constant α (and also β and γ in the extended models) has an effect on the
responsiveness of the model. If small value of alpha is used, the model will respond slowly to
changes in the observed demand. This will result in relatively stable forecast. Alternatively, if large
value of alpha is used the forecast will be more responsive to changes in observed demand. (Zeni,
2001) Mean squared error (MSE) is commonly used to determine smoothing constant. Smoothing
constant is simply set so that it minimizes the MSE of the forecast. (Makridakis et al., 1998)

46
4.4.3 Pickup models
Pickup models are advanced forecasting models. The main idea of pickup models is to produce a
forecast of incremental demand between time periods and then aggregate these increments to com-
pile a forecast of total demand at the departure (Talluri & van Ryzin, 2004). Therefore forecasted
demand is a function of the current bookings and the amount of incremental bookings between the
current time and the departure. These incremental bookings are illustrated by the amount of book-
ings picked up between the current time and the departure. (Zeni, 2001)

Zeni (2001) presents two basic pickup models: additive and multiplicative pickups. Both of these
models can be performed using classical pickup or advance pickup. The difference between these
model specifics is how they use the historical data. The classical pickup uses only complete histor-
ical booking data to compile incremental bookings between periods, while the advanced pickup
uses all available data, including partial booking data. The formulations of pickup models are il-
lustrated here using the classical versions.

Additive pickup model:

̅̅̅̅𝐷𝐵(𝑋,0) = 𝐵̅𝐷𝐵0 − 𝐵̅𝐷𝐵𝑋
𝑃𝑈 (4.9)

Equation 4.10 presents the average pickup between day X before departure and departure
̅̅̅̅𝐷𝐵(𝑋,0) ). The average bookings at departure (𝐵̅𝐷𝐵0) and X days before departure (𝐵̅𝐷𝐵𝑋 ) are
(𝑃𝑈
compiled from the historical data of previous departures. For example moving average method can
be used to calculate these averages. The forecast of bookings at departure can be then presented by
combining the bookings at X days before departure and the average pickup between day X and
departure:

𝐵𝐷𝐵0 = 𝐵𝐷𝐵𝑋 + ̅̅̅̅

𝑃𝑈𝐷𝐵(𝑋,0) (4.10)

This model is called additive pickup, since it adds the average incremental bookings to current
bookings at certain time. In addition to this, also multiplicative pickup models have been devel-
oped. They are similar to additional models in the sense that they use information on incremental
bookings to forecast future bookings, but instead of adding the average pickups they multiplies
current bookings by the average pickup ratio. (Zeni, 2001)

47
Multiplicative pickup model:
The average pickup ratio is determined by:

𝐵 ̅
̅̅̅̅̅̅
𝑃𝑈𝑅𝐷𝐵(𝑥,0) = 𝐵̅ 𝐷𝐵0 (4.11)
𝐷𝐵𝑋

A forecast of bookings at departure can be the compiled from equation:

𝐵𝐷𝐵0 = 𝐵𝐷𝐵𝑋 ∗ ̅̅̅̅̅̅

𝑃𝑈𝑅𝐷𝐵(𝑋,0) (4.12)

4.4.4 Regression
In RM, regressions can be used to predict the future demand of a certain departure, without making
any conclusions about the underlying factors that might affect the demand. Regression forecasting
in RM assumes that there exists some relationship between the final bookings and the bookings of
a certain time before the departure. (Zeni, 2001) Both historical and advanced booking data are
used to compile regression forecasts. Zeni (2001) presents a simple regression using both historical
and advanced data. The dependent variable is the total demand at the time of departure and the
demand at a specific time before departure has been used as an explanatory variable. For example
using demand at day X as an independent variable, the linear regression model is:

𝐵𝐷𝐵0 = 𝛽0 + 𝛽1 𝐵𝐷𝐵𝑋 (4.13)

Where, 𝐵𝐷𝐵0 is the number of total bookings at the departure, 𝐵𝐷𝐵𝑋 is bookings at X days before
the departure. 𝛽0 and 𝛽1 are regression parameters that are estimated from the historical data using
some estimation technique, for example ordinary least square (OLS). (Zeni, 2001) The demand
forecast of bookings at the day of departure is then compiled inserting the number of bookings at
X days before the departure in the regression equation. In this phase advanced booking data can be
used to obtain more accurate forecasts.

Zeni (2001) emphasizes that this regression model differs from the general econometric models of
demand, since it does not include any economic causal variables, such as price. The forecast of
total demand is only dependent on the previous bookings and thus demand forecasts should be
defined to all fare classes separately.

While previous linear regression model assumes linear relationship between dependent variable
and independent variable, also other regression models could be applied. Weatherford et al. (2002)
state that demand forecast can also be presented by quadratic or cubic regressions.
48
4.5 Forecast performance
As noted, demand forecasting plays a crucial role in RM. In order to produce efficient capacity
allocation and pricing decisions, these decisions should be based on demand forecasts. However,
it does not provide any additional value to base these decisions on inaccurate forecasts – quite the
contrary, resting inaccurate forecasts might lead to even more inefficient capacity allocation and
pricing decisions. Talluri and van Ryzin (2004) underline that the analysis of the forecast perfor-
mance is as important as the forecasting itself. It is suggested that improvement of forecast accuracy
by 10 % can increase revenue by 0.5 – 3.0 % (Lee, 1990).

The term forecast performance or accuracy refers to how well the forecasting model is able to
extrapolate the available data to the future (Makridakis et al., 1998). The forecast performance is
measured using forecast errors. The idea of forecast error is simple:

𝑒𝑡 = 𝑋𝑡 − 𝐹𝑡 (4.14)

The forecast error, 𝑒𝑡 , is the difference between the observed demand, 𝑋𝑡 , and the forecasted de-
mand, 𝐹𝑡 , at time t. This is an example of one-step forecast error, since it takes only account of one
period, t. (Makridakis et al., 1998)

Besides that forecast errors can be used to evaluate forecast performance, errors can be useful also
for other reasons. The variance of demand can be estimated using forecast errors, which would
enable to estimate the demand distribution. Errors can also be used to identify and exclude outliers
from the data. Moreover, errors can give the signals of unusual events or instability in the demand.
(Talluri & van Ryzin, 2004)

Since the forecasting system generally includes observations and forecasts for more than one time
period, it is necessary to compile a measure of total forecast error. There are several commonly
used methods to measure the total forecast error, which all assume that there is available infor-
mation on forecasts and observed demands for n periods. (Makridakis et al., 1998; Talluri & van
Ryzin, 2004)

49
Mean error (ME):
1
𝑀𝐸 = 𝑛 ∑𝑛𝑡=0 𝑒𝑡 (4.15)

The mean error is the average of the errors of each period n. It represents an estimate of the forecast
bias. When the forecasting system is unbiased, the mean error will converge to zero once N in-
creases. (Talluri & van Ryzin, 2004) Each 𝑒𝑡 can be calculated using Equation 4.14. The problem
of the mean error method is that positive and negative errors tend to offset each other. The mean
error represents only over- or under-forecasting of the model and thus it does not provide a good
illustration of the size of the errors. (Makridakis et al., 1998)

Mean squared error (MSE):

1
𝑀𝑆𝐸 = 𝑛 ∑𝑛𝑡=0 𝑒𝑡2 (4.16)

The mean squared error can be used to overcome the problem of the mean error method. The model
makes all errors positive by squaring them, making it possible to interpret the size of total error.
Consequently it gives a better illustration of the forecast accuracy. (Makridakis et al., 1998)

Mean percentage error (MPE):

𝑋𝑡 −𝐹𝑡
𝑃𝐸𝑡 = ∗ 100 (4.17)
𝑋𝑡

1
𝑀𝑃𝐸 = 𝑛 ∑𝑛𝑡=0 𝑃𝐸𝑡 (4.18)

In order to compile the mean percentage error, the percentage error, 𝑃𝐸𝑡 , must be calculated first
(Equation 4.17). The mean percentage error is then formulated by averaging the percentage errors.
(Makridakis et al., 1998)

Mean absolute percentage error (MAPE):

The mean percentage error suffers the same problem as the mean error method and thus the mean
absolute percentage error can be more efficient to describe the forecast accuracy.
1
𝑀𝐴𝑃𝐸 = 𝑛 ∑𝑛𝑡=0 |𝑃𝐸𝑡 | (4.19)

The mean absolute percentage error is constructed by averaging the absolute values of percentage
errors. (Makridakis et al., 1998)

50
4.6 Network forecasting
Because of the prevailing network effect in the railway industry, leg-level demand forecasts may
not be sufficient to set efficient capacity allocations. The network capacity allocation models re-
quire that also demand forecasts are compiled in the network level (Walczak et al., 2012). As men-
tioned in Section 2.2.3 the network effect makes the capacity allocation problem extremely com-
plex and thus also the forecasting process becomes more challenging.

Walczak et al. (2012) present how the network effect creates challenges for demand forecasting.
Perhaps the most significant complication is that the number of necessary forecasts increases sub-
stantially. This complication can be illustrated by a simple example. Let’s consider a railway net-
work, comprising seven stations (Figure 18).

Figure 18 – A simple railway network.

If a train travels from A to G without making any intermediate stops, the itinerary A – G can be
considered as a single-leg and a demand forecast only for leg A – G would be needed to allocate
capacity efficiently. But if the train stops at all five intermediate stations, the itinerary A – G should
be considered as a network, which consists of six separated legs (A – B, B – C … F – G). These
legs then form 21 different OD pairs (A – B, A – C, A – D … E – G, F – G). Thus in order to make
efficient capacity allocation for this train, 21 different demand forecasts should be performed. Ad-
ditionally, if we take into consideration multiple fare classes, the number of forecasts needed in-
creases even more. Also the second dimension of railway network (presented in Figure 10) would
increase the number of necessary forecasts.

But while the number of forecasts increases in the network context, the forecasted demands for
each OD pairs decreases (Walczak et al., 2012). The problem is that the demand for a certain OD
can be so small that it would be impossible to draw accurate demand forecast for that itinerary.
Walczak et al. (2012) propose that this problem can be addressed by forecasting only the most
significant ODs directly while forecasting demands for less demanded ODs at more aggregate
level.

51
Third challenge relates to demand unconstraining. In a network system, the demand unconstraining
process is far more complicated than in a single-leg system because the demand can be constrained
by the booking limits on one or more of the legs that are overlapped. Hence all leg-level constraints
over the network at any given time must be tracked, so that the network unconstraining can be
performed. (Walczak et al., 2012)

52
5 Illustrative examples
Following sections give illustrative examples of demand forecasting and capacity allocation and
their relationship in RM system. Examples consider singe-leg OD forecasting and capacity alloca-
tion problems and do not take account of network effects. Because of limited availability of real
demand data, I use hypothetical booking histories to represent how different forecasting methods
can be applied. I also show how the performance of different methods can be evaluated. Addition-
ally, I show how demand forecasts would impact on capacity allocation.

The main objective is to show how different forecasting methods are applied. I perform demand
forecasts using four types of methods: moving average, exponential smoothing, regression, and
pickup methods. These methods differ in whether they use only historical booking data or also
available information on advanced bookings. The performance of each method is then evaluated
using basic measures of forecast errors: mean error, mean squared error, and mean absolute per-
centage error. Since forecasts are based on the hypothetical booking data, these evaluations do not
provide any conclusions on the performance of these methods for real demand data. They are only
presented in order to illustrate how the performance of different methods should be evaluated with
real data.

In addition to demand forecasts, I show how capacity allocation can be conducted using EMSR –
model. The objective of this example is to illustrate the link between demand forecasting and ca-
pacity allocation. The example pictures how demand environment effects on capacity allocation.

5.1 Data
The data that is used in demand forecasting examples is based on hypothetical booking history
matrix presented by Wickham (1995) (Table 2). The booking matrix represents unconstrained
booking profiles over 18 weeks period for single booking class for single-leg OD itinerary. Rows
in the matrix represent booking profiles for each departure and columns show the number of book-
ings on hand at particular day before the departures.

53
 Week DB0 DB7 DB14 DB21 DB28 DB35 DB42 DB49 DB56
1 25 22 10 5 3 3 2 0 0
2 30 21 17 15 12 7 3 1 0
3 25 23 14 9 8 5 5 2 1
4 40 34 30 16 11 6 3 0 0
5 35 29 20 13 12 8 3 1 0
6 39 33 30 21 14 6 4 2 1
7 45 28 22 18 10 5 3 0 0
8 50 42 18 11 10 7 4 2 1
9 33 29 21 15 9 8 6 6 2
10 46 40 29 22 11 7 3 2 0
11 49 37 25 17 10 9 8 5 2
12 25 22 10 5 3 3 2 0 0
13 30 21 15 17 12 7 3 1 0
14 23 22 14 9 8 5 5 2 1
15 40 34 30 16 11 6 3 0 0
16 35 29 20 13 12 8 3 1 0
17 39 33 30 21 14 6 4 2 1
18 45 28 22 18 10 5 3 0 0
Table 2 – Hypothetical booking history matrix (Wickham, 1995).

In order to determine the performance of different methods, the forecasted bookings should be
compared to the actual booking values. Therefore the historical booking data is divided into his-
torical and future departures by setting some artificial present week. Week 10 is chosen to be a
present week and thus it is expressed as Week 0 in Table 3, which represents the division of the
original booking matrix.

All negative week numbers are considered past departures and booking data above the Week 0
thick line represents historical booking data. The positive week numbers denotes the future depar-
tures so that e.g. Week 5 represents a train, which departs in five weeks. The shaded cells under
the Week 0 line picture advanced bookings of trains that are not yet departed. The bookings in the
unshaded cells are bookings that are not yet recorded because they are considered to take place in
the future. They are used to evaluate the forecast performance of different methods. Thus all shaded
cells together represent available data that can be used forecasting at present time (Week 0). The
procedure is similar to the method that Wickham (1995) uses in his thesis.

54
 Week DB0 DB7 DB14 DB21 DB28 DB35 DB42 DB49 DB56
-9 25 22 10 5 3 3 2 0 0
-8 30 21 17 15 12 7 3 1 0
-7 25 23 14 9 8 5 5 2 1
-6 40 34 30 16 11 6 3 0 0
-5 35 29 20 13 12 8 3 1 0
-4 39 33 30 21 14 6 4 2 1
-3 45 28 22 18 10 5 3 0 0
-2 50 42 18 11 10 7 4 2 1
-1 33 29 21 15 9 8 6 6 2
0 46 40 29 22 11 7 3 2 0
1 49 37 25 17 10 9 8 5 2
2 25 22 10 5 3 3 2 0 0
3 30 21 15 17 12 7 3 1 0
4 23 22 14 9 8 5 5 2 1
5 40 34 30 16 11 6 3 0 0
6 35 29 20 13 12 8 3 1 0
7 39 33 30 21 14 6 4 2 1
8 45 28 22 18 10 5 3 0 0
Table 3 – Division of booking matrix.

Grey lines in Figure 19 illustrate historical booking profiles and the black line is the average of
historical profiles. The descriptive statistics of historical bookings are given in Table 4.

55
50
45
40
35
Bookings

30
25
20
15
10
5
0
DB0 DB7 DB14 DB21 DB28 DB35 DB42 DB49 DB56
Days before departure

Historical booking profiles Average historical booking profile

Figure 19 – Historical booking profiles.

55
 Statistic DB0 DB7 DB14 DB21 DB28 DB35 DB42 DB49 DB56
Mean 37 30 21 15 10 6 4 2 1
Std. Error 2,76 2,28 2,17 1,66 0,94 0,49 0,37 0,56 0,22
Median 37 29 21 15 11 7 3 2 0
Std. Dev. 8,72 7,22 6,85 5,25 2,98 1,55 1,17 1,78 0,71
Variance 75,96 52,10 46,99 27,61 8,89 2,40 1,38 3,16 0,50
Min. 25 21 10 5 3 3 2 0 0
Max. 50 42 30 22 14 8 6 6 2
Table 4 – Descriptive statistics of historical bookings.

5.2 Forecasting
Using the data in Table 3, demand forecasts are compiled in two different scenarios using selected
methods. Table 5 presents selected methods and necessary model specifics.

Method Model specifics

Moving average Last 10 observations are used: n = 10
Historical
methods

Exponential smoothing Smoothing parameter = 0,5

(α=0,5) Average of last 10 observations is used as a last forecast (Ft)
Exponential smoothing Smoothing parameter = 0,2
(α=0,2) Average of last 10 observations is used as a last forecast (Ft)
Additive pickup Classical additive pickup
Advanced
methods

Multiplicative pickup Classical multiplicative pickup

Dependent variable: Bookings at DB0
Regression (DB21)
Independent variable: Bookings at DB21
Table 5 – Selected forecasting methods and specifics.

5.2.1 Scenario 1
In the first scenario, the objective is to compile forecasts of final bookings for the next eight weeks.
The forecast of final bookings for each week is constructed at present time (Week 0) using all
available booking data at that time. Selected methods are divided into two groups, historical and
advanced methods, depending on which booking data they use in forecasting (Table 5). Table 6
illustrates the difference between methods using Week 3 as an example. Grey shaded areas repre-
sent booking data used in forecasting for both groups.

56
 Historical methods Advanced methods
Week DB0 DB7 DB14 DB21 Week DB0 DB7 DB14 DB21
-9 25 22 10 5 -9 25 22 10 5
-8 30 21 17 15 -8 30 21 17 15
-7 25 23 14 9 -7 25 23 14 9
-6 40 34 30 16 -6 40 34 30 16
-5 35 29 20 13 -5 35 29 20 13
-4 39 33 30 21 -4 39 33 30 21
-3 45 28 22 18 -3 45 28 22 18
-2 50 42 18 11 -2 50 42 18 11
-1 33 29 21 15 -1 33 29 21 15
0 46 40 29 22 0 46 40 29 22
1 37 25 17 1 37 25 17
2 10 5 2 10 5
3 F 17 3 F 17
Table 6 – Scenario 1: Data used in forecasts of Week 3 final bookings.

As can be seen, advanced methods use the number of bookings at DB21 in current week to forecast
DB0 bookings. Table 7 gives the specifics of the forecasts of Week 3 bookings.

Forecasts of Week 3 bookings at DB0

Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 37 30 -7 46 23 %
Exponential smoothing (α=0,5) 41 30 -11 130 38 %
Exponential smoothing (α=0,2) 39 30 -9 75 29 %
Additive pickup 39 30 -9 86 31 %
Multiplicative pickup 43 30 -13 173 44 %
Regression (DB21) 39 30 -9 86 31 %
Table 7 – Scenario 1: Forecasts and errors for Week 3.

Similar forecasts are compiled for each week separately (Appendix 1). The performance of differ-
ent methods can be evaluated using methods presented in Section 4.5.1. Both MSE and MAPE
indicate that additive pickup method works best in this hypothetical data (Table 8). Also regression
(DB21) performs well, but it has to be taken account that it can be compiled only for first three
weeks, since the number of bookings on hand at DB21 is needed to compile regression forecast.
Both exponential smoothing methods perform weakly. It is not surprising because exponential
smoothing is designed to be used with long time series and moving forecast process. Multiplicative
pickup model gets the largest MSE, 149, which is result of extremely bad forecast in Week 8. If

57
forecast of Week 8 is omitted, the MSE of multiplicative pickup decreases significantly and the
performance becomes nearly as good as additive pickup model.

Forecasts performance over 8 weeks: Ranked by MSE

Method ME MSE MAPE
Additive pickup 0 43 16 %
Regression (DB21)* -1 62 20 %
Moving average (n=10) -1 76 24 %
Exponential smoothing (α=0,2) -3 84 25 %
Exponential smoothing (α=0,5) -6 107 29 %
Multiplicative pickup -4 149 27 %
* Forecasts only for first three weeks
Table 8 – Scenario 1: Forecast performance.

35%
140
29%
30%
120 27%
25%
24%
25%
100
20%
20%

MAPE
80 16%
MSE

149
15%
60
107

40 84 10%
76
62
20 43 5%

0 0%
Moving Exponential Exponential Additive Multiplicative Regression
average smoothing smoothing pickup pickup (DB21)*
(n=10) (α=0,5) (α=0,2)

MSE MAPE * Forecasts only for first three weeks

Figure 20 – Scenario 1: Forecast performance.

58
5.2.2 Scenario 2
While in Scenario 1 the forecasts of final bookings for the next eight weeks are compiled at static
“present time” (Week 0), in Scenario 2 the forecasts of the same eight weeks are compiled but in
moving “present time”. Thus the objective is to produce forecast of final bookings for next period.
For example the forecast of Week 3 final bookings is compiled at Week 2 using all available data
at that time. This type of modification gives better illustration on how methods might work dynam-
ically in long-term. Same six methods than in Scenario 1 are used to compile the forecasts of final
bookings. Again Week 3 is used as an example to illustrate how historical and advanced methods
differ (Table 9).

Historical methods Advanced methods

Week DB0 DB7 DB14 DB21 Week DB0 DB7 DB14 DB21
-9 25 22 10 5 -9 25 22 10 5
-8 30 21 17 15 -8 30 21 17 15
-7 25 23 14 9 -7 25 23 14 9
-6 40 34 30 16 -6 40 34 30 16
-5 35 29 20 13 -5 35 29 20 13
-4 39 33 30 21 -4 39 33 30 21
-3 45 28 22 18 -3 45 28 22 18
-2 50 42 18 11 -2 50 42 18 11
-1 33 29 21 15 -1 33 29 21 15
0 46 40 29 22 0 46 40 29 22
1 49 37 25 17 1 49 37 25 17
2 25 22 10 5 2 25 22 10 5
3 F 21 15 17 3 F 21 15 17
Table 9 – Scenario 2: Data used in forecasts of Week 3 final bookings.

The forecasts of Week 3 final bookings and corresponding errors are presented in Table 10.

Forecasts of Week 3 bookings at DB0

Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 39 30 -9 76 29 %
Exponential smoothing (α=0,5) 35 30 -5 26 17 %
Exponential smoothing (α=0,2) 38 30 -8 57 25 %
Additive pickup 28 30 2 4 7%
Multiplicative pickup 26 30 4 19 15 %
Regression (DB21) 39 30 -9 84 31 %
Table 10 – Scenario 2: Forecasts and errors for Week 3.

59
The forecasts and errors for other seven weeks are presented in Appendix 2. The performance of
historical and advanced methods are clearly different in Scenario 2 using the hypothetical data
(Table 11). Advanced methods appear to perform significantly better than methods that rely only
on historical data, when compiling the forecasts of the final bookings for the next period.

Forecast performance over 8 weeks: Ranked by MSE

Method ME MSE MAPE
Multiplicative pickup 1 25 11 %
Additive pickup 0 28 12 %
Regression (DB21) -1 35 14 %
Exponential smoothing (0,2) -1 91 27 %
Moving average (n=10) -2 96 27 %
Exponential smoothing (0,5) 0 103 27 %
Table 11 – Scenario 2: Forecast performance.

120 30%
27% 27% 27%
100 25%

80 20%

MAPE
MSE

14%
60 12% 15%
103 11%
96 91
40 10%

20 35 5%
28 25
0 0%
Moving Exponential Exponential Additive Multiplicative Regression
average smoothing smoothing pickup pickup (DB21)
(n=10) (0,5) (0,2)

MSE MAPE

Figure 21 – Scenario 2: Forecast performance.

60
5.3 Capacity allocation
In this section, the effects of demand forecasts on capacity allocation have been illustrated using
EMSR –model. Single-leg capacity allocation with four fare classes is compiled under four differ-
ent demand cases. Four nested fare classes with average fares are Business (100€), Economy (60€),
Discount (30€) and Super discount (20€). The average fares of each class are assumed to be stable
in all cases. This assumption indicates that price changes cannot be used to stimulate demand.
Additionally, the total capacity is fixed to 150 seats. Capacity is allocated to these fare classes
assuming that the general assumptions presented in Section 2.2 hold. Demand for each fare class
is also supposed to be independent between fare classes. This denotes that each passenger is de-
manding only one fare class. Although this is extremely theoretical assumption, it is necessary in
order to present capacity allocation in such a simplistic form. Four demand cases are described
next.

Case 1:
In the first case, forecasted aggregate demand of all fare classes equals the total capacity of imag-
inary train. Demand forecasts also suggest that the demands for higher fare classes (B and E) will
be lower than for lower fare classes (D and S). (Table 12)

Case 1 Business (B) Economy (E) Discount (D) Super discount (S)
Fare 100 € 60 € 30 € 20 €
Mean 15 30 50 55
Std. Dev. 5 10 17 19
Forecasted total demand: 150
Table 12 – Case 1 details.

Case 2:
In the second case, forecasted aggregate demand equals the total capacity, but now the forecasted
total demand is evenly distributed across the fare classes. (Table 13).

Case 2 Business (B) Economy (E) Discount (D) Super discount (S)
Fare 100 € 60 € 30 € 20 €
Mean 37,5 37,5 37,5 37,5
Std. Dev. 5 5 5 5
Forecasted total demand: 150
Table 13 – Case 2 details.

61
Case 3:
The third case represents the situation where forecasted aggregate demand exceeds the total capac-
ity. Similarly to the first case, the forecasted demands for lower fare classes are higher than for
higher fare classes. (Table 14)

Case 3 Business (B) Economy (E) Discount (D) Super discount (S)
Fare 100 € 60 € 30 € 20 €
Mean 30 60 100 150
Std. Dev. 5 15 17 20
Forecasted total demand: 340
Table 14 – Case 3 details.

Case 4:
The last case pictures the demand environment where the total demand is lower than the total ca-
pacity. Like in cases 1 and 3, the forecasted demands for lower fare classes are higher than for
higher fare classes. (Table 15)

Case 4 Business (B) Economy (E) Discount (D) Super discount (S)
Fare 100 € 60 € 30 € 20 €
Mean 8 30 43 55
Std. Dev. 3 9 12 17
Forecasted total demand: 136
Table 15 – Case 4 details.

Basic EMSR –model is used to construct capacity allocations for all four cases. Protection levels
between fare classes are determined respectively and these distinct protection levels are then
summed to obtain the joint protection levels for classes B, E and D. Booking limits for each fare
class are then compiled subtracting the joint protection level of higher fare class from the total
capacity. For example in Case 1, the booking limit for class D can be expressed:

𝐵𝑜𝑜𝑘𝑖𝑛𝑔 𝑙𝑖𝑚𝑖𝑡𝐷 = 𝑇𝑜𝑡𝑎𝑙 𝐶𝑎𝑝𝑎𝑐𝑖𝑡𝑦 − 𝐽𝑜𝑖𝑛𝑡 𝑝𝑟𝑜𝑡𝑒𝑐𝑡𝑖𝑜𝑛 𝑙𝑒𝑣𝑒𝑙𝐷𝑖 (5.1)

𝐵𝑜𝑜𝑘𝑖𝑛𝑔 𝑙𝑖𝑚𝑖𝑡𝐷 = 150 − (14 + 34) = 102 (5.2)

More detailed description of EMSR –allocations are presented in Appendix 3.

Figure 22 pictures capacity protection levels for Cases 1 and 2 and Figure 23 shows respective
booking limits. These figures illustrate well how capacity allocations depend on how total demand

62
is distributed between fare classes. When demands for lower fare classes are significantly higher
than demands for higher fare classes, more capacity should be allocated to lower fares.

49
41 39
34 36

14

Case 1 Case 2

B E D

Figure 22 – Protected capacity from lower fare classes (Case 1 & 2).

For example, booking limits for classes D and S are considerable larger in Case 1 than Case 2
(Figure 23). For example, in Case 1 the booking limit of class D is 103 and the corresponding
booking limit in Case 2 is only 73.

Booking limits: Case 1 Booking limits: Case 2

150 B 150 B

136 E 114 E

102 D 73 D

54 S 33 S

Figure 23 – Booking limits (Case 1 & 2).

63
Capacity protection levels and booking limits in Cases 3 and 4 reflect another considerable issue
of demand forecasts. While the forecasted total demand is quite evenly distributed across fare clas-
ses in Cases 3 and 4, the forecasted total demands are significantly different. Case 3 can be seen as
a peak period. The total demand is expected to be higher than the total capacity. Because of this
excess demand, the expected revenues are maximized if more capacity is allocated to higher fare
classes. In Case 3 nearly one thirds of the capacity (93 seats) is allocated to classes B and E (Figure
24). Actually, the expected demands for higher classes in Case 3 are so large that capacity alloca-
tion model suggests that no capacity should be allocated to the lowest fare class S. (Figure 25) On
the contrary, Case 4 can be considered as a low demand state, since the forecasted total demand is
less than the total capacity. In order to maximize expected revenues and load factor, capacity allo-
cation model suggests that less capacity should be protected to higher fare classes. Thus in Case 4
only 45 seats are protected to higher classes B and E (Figure 24).

64
58
43
29 32

Case 3 Case 4

B E D

Figure 24 – Protected capacity from lower fare classes (Case 3 & 4).

64
 Booking limits: Case 3 Booking limits: Case 4

150 B 150 B

121 E 143 E

58 D 111 D

S 68 S

Figure 25 – Booking limits (Case 3 & 4).

5.4 Conclusions of examples

None of previous results can be interpreted as generalized results, since those are totally based on
the hypothetical booking data. Nevertheless, these examples reflect some significant issues related
to demand forecasting in RM.

Firstly, the time horizon of necessary forecasts has definite effect on method selection. Some meth-
ods might be suitable for long-term forecasting while some others are more suitable for forecasting
in shorter-term. Thus the evaluation of different possible methods is essential in the selection of
the most suitable method for each particular forecasting process. It is especially important to eval-
uate, could forecast performance be improved by utilizing advanced forecasting method when ad-
vanced booking data is available. It is evident that in my hypothetical data, advanced methods
provide explicitly better results than methods that only rely on historical booking data.

Second implication is that demand forecasts should be performed continuously. New corrective
forecasts should be compiled dynamically when new information on demand becomes available.
This is supported by the demand forecast examples, since the booking forecasts appear to be more
precise in Scenario 2 compared to the forecasts in Scenario 1. Also this implication relates to the
use of advanced booking information.

65
Thirdly, the EMSR – capacity allocation examples demonstrate two reasons why demand forecast-
ing plays such a significant role in the RM systems. In order to compile effective capacity alloca-
tions, demand forecasts should be able to give precise predictions of the total demand and the
distribution of total demand across fare classes. Cases 1 and 2 illustrate how the distribution of
total demand affects capacity allocation. When the forecasted total demand equals the total capac-
ity, the number of allocated seats for a specific fare class corresponds to the relative shares of total
demand of that particular class. Cases 3 and 4 account for the impacts of the forecasted total de-
mand. In order to optimize capacity allocation under excess demand the capacity allocation model
allocates more capacity to higher fares. On the contrary, in scarce demand case more capacity is
allocated to lower fare classes so that higher load factor and thus higher revenues can be attained.

66
6 Conclusions
More than 30 years RM has gathered lots of interest among researchers in different subjects. RM
has been studied from the perspective of economics, transportation research and operation research.
The fundamental economic concepts, willingness-to-pay, price discrimination and product differ-
entiation, form the ground for RM and RM research for other subjects. The majority of studies are
focused on the airline industry, most likely because of the importance of the industry in the U.S.
Demand forecasting has been one of the main research areas in RM because of its tight relations to
other RM components; capacity allocation, overbooking and pricing.

The purpose of this thesis was to investigate demand forecasting in the context of railway RM. At
first I answered two questions, what is RM and how is it applicable to railway industry, based on
the existing literature. Then demand forecasting issues were covered in the context of railway RM.
Finally, I performed hypothetical demand forecasting and capacity allocation examples and illus-
trated the specifics of demand forecasting and its impacts on capacity allocations.

Even though RM is generally linked to airlines, I find that actually it is applicable to many service
industries. Like Kimes (1989) presents, RM is basically a tool for all capacity constrained service
firms to manage their inventories. Studies are widely agreed on the six general conditions that are
necessary to effective applicability of RM. A firm employing RM has to have relatively fixed ca-
pacity in short-term and it must be able to somehow segment its customers according to their pref-
erences. The products are also considered perishable and the purchases of the products occur before
the actual consumption. Additionally, demand can be seen changing substantially over time and
these changes happen with some level of uncertainty. Finally, because the capacity is considered
fixed in short-term, the marginal cost of additional consumer is low and the costs of capacity
changes are high.

Taking account of these conditions it seems evident that RM is applicable in railway markets. This
is in line with the study of Armstrong and Meissner (2010) and with the fact that many railway
operators have used different RM techniques several decades. Railways and airlines appear to func-
tion rather similar market environments, but also a few significant differences exist. Thus it is
important to evaluate the differences between these industries and also consider how these differ-
ences might effect on RM. I identified four main differences. Firstly, the network effect is more
dominant in railway markets because of the two dimensional network structure. This creates highly
67
complex forecasting and capacity allocation processes and thus efficient RM applications are even
more difficult to construct than in airline markets. Secondly, railway operators have historically
used more product differentiation techniques and less weight has been put on price differentiation
methods. This is a result of differences in passenger mixes between routes. Third difference relates
to the conventions of ticket purchasing between airlines and railways. Even if the development of
information technology and self-service selling platforms in the railway industry has increased the
number of advanced ticket purchases, the number of passengers who buy tickets just before the
departure is still significant in railway markets. This is probably the most distinctive consideration
between these two industries in RM context, since the advanced purchase restrictions have tradi-
tionally been one of the main passenger segmentation method used by the airlines. The last issue
relates to the differences in competitive environments. In many countries railway operators are
monopolized by the governments and despite regulatory changes, especially in Europe, movement
towards more competitive markets has been slow. One significant reason for this slow change of
market environment is extremely high entry costs in the railway industry. These concentrated mar-
kets indicate that railway operators might have higher pricing power than airlines have in furiously
competitive airline markets. This suggests that railways might actually have more potential to uti-
lize RM techniques than airlines, since price differentiation is more powerful under monopoly mar-
kets.

As RM techniques are shown to be effective in the railway industry, also the importance of demand
forecasting is apparent. Like demand forecasting in general, also railway demand forecasting faces
some considerable difficulties. One significant problem relates to the constrained demand data. The
methods of how this problem can be resolved are out of my topic, but it is extremely important to
understand how constrained demand data might have effects on demand forecasts. I focused on the
differences between forecasting methods and presented some methods of how the performance of
these methods can be evaluated. The main difference between methods is how they use available
demand data in forecasting process. The illustrative examples indicate that the time horizon of
forecasts relates to the performance of different methods. Advanced methods seem to perform rel-
atively better than historical models when short-term forecasts are compiled dynamically to the
next period. However, it has to be emphasized that any generalizations cannot be drawn as hypo-
thetical booking data was used in forecasting instead of real booking information.

68
Capacity allocation examples highlight other important issues. The effects of the forecasted total
demand and demand distribution between fare classes on capacity allocation, and thus revenue
optimization, are notable. Even if only EMSR –model was used to allocate capacity in the illustra-
tive example, also other capacity allocation models, such as probabilistic linear programming, re-
quire information on the total demand and demand distribution between fare classes. Examples
demonstrate how significantly capacity allocations can differ between demand states. Thus it is
extremely important that demand forecasting process manages to identify fluctuations in demand
over time. Especially important is to identify abnormal demand states, such as outstandingly high
and low demands.

In this thesis, an extensive overview of RM applications and demand forecasting in a railway RM

system was compiled. However, the results involve some serious drawbacks that should be taken
into account. The most notable drawback relates to the empirical demonstration. As the data used
in the illustrative examples is completely hypothetical, any general conclusions could not be pro-
duced. This thesis only manages to point out some significant issues that should be considered in
demand forecasting process. Further empirical analysis with real demand data should be needed to
give more precise conclusions on which forecasting method should be applied in which circum-
stances. Additionally, the issue of constrained demand data could not be covered sufficiently and
therefore more research on demand unconstraining methods with real railway demand data should
be needed. Third significant drawback relates to existing network effects in railway markets. In
this thesis capacity allocation and demand forecasting processes were mainly discussed under sin-
gle-leg environment. Because the network effect is evidently relevant in railway markets, even
more outstanding than in airline markets, it should be definitely incorporated into further study.

The last consideration relates to the level of forecasting. The existing literature on RM demand
forecasting, as well this thesis, is mainly interested in micro-level demand forecasting. Since rail-
ways compete with other transportation modes, it would be beneficial to incorporate passenger
choice models to RM demand forecasting. This is especially relevant issue at the moment in Fin-
land because the deregulation of the bus industry has increased substitutive competition signifi-
cantly and thus some competitive counteractions should be done in the railway industry.

69
References
Alderighi, M. (2010). Fare dispersions in airline markets: A quantitative assessment of theoretical
explanations. Journal of Air Transport Management 16, 144-150.

Alstrup, J., Boas, S., Madsen, O. B., & Vidal, R. V. (1986). Booking policy for flights with two types of
passengers. European Journal of Operations Research, vol. 27, 274-288.

Armstrong, A., & Meissner, J. (2010). Railway revenue management: overview and models. Working paper.

Belobaba, P. P. (1987). Air Travel Demand And Airline Seat Inventory Management. PhD Thesis, Flight
Transportation Laboratory, MIT.

Belobaba, P. P. (1992). Optimal vs. heuristic methods for nested seat allocation. Presentation at ORSA/TIMS
Joint National Meeting.

Belobaba, P. P. (2009). Fundamentals of Pricing and Revenue Management. In P. Belobaba, A. Odoni, &
C. Barnhard, The Global Airline Industry. John Wiley & Sons, Ltd.

Ben-Akiva, M. E., & Lerman, S. R. (1985). Discrete choice analysis: theory and application to travel
demand. Vol. 9, MIT press.

Ben-Khedher, N., Kintanar, J., Queille, C., & Stripling, W. (1998). Schedule Optimization at SNCF: From
Conception to Day of Departure. Interfaces 28(1), 6-23.

Bharill, R., & Rangaraj, N. (2008). Revenue management in railway operations: A study of the Rajdhani
Express, Indian Railways. Transportation Research Part A 42, 1195-1207.

Bitran, G., & Caldentey, R. (2003). An Overview of Pricing Models for Revenue Management.
Manufacturing & Service Operations Management 5(3), 203-229.

Borenstein, S., & Rose, N. L. (1994). Competition and Price Dispersion in the U.S. Airline Industry. Journal
of Political Economy, Vol. 102, No. 4, 653-683.

Botimer, T. C., & Belobaba, P. P. (1999). Airline pricing and fare product differentiation: A new theoretical
framework. The Journal of the Operation Research Society, Vol. 50, No. 11, Yield Management,
1085-1097.

Brumelle, S. L., & McGill, J. I. (1993). Airline Seat Allocation with Multiple Nested Fare Classes.
Operation Research, Vol. 41, No. 1, 127-137.

Chuang, H. M., Chu, C. P., & Niu, W. F. (2010). A study on revenue management of Taiwan High Speed
Railway. Industrial Engineering and Engineering Management (IEEM), (pp. 999-1003).

70
Ciancimino, A., Inzerillo, G., Lucidi, S., & Palagi, L. (1999). A mathematical programming approach for
the solution of the railway yield management problem. Transportation science 33 (2), 168-181.

Coase, R. H. (1972). Durability and Monopoly. Journal of Law and Economics, Vol. 15, No. 1, 143-149.

Cooper, W. L., Homem-de-Mello, T., & Kleywegt, A. J. (2006). Models of the Spiral-Down Effect in
Revenue Management. Operations Research, Vol. 54, No. 5, 968-987.

Cross, R. G. (1998). Revenue Management: hard-core tactics for market domination. Broadway Books.

Dana, J. D. (1999). Equilibrium Price Dispersion under Demand Uncertainty: The Roles of Costly Capacity
and Market Structure. The RAND Journal of Economics, Vol. 30, No. 4, 632-660.

European Commission. (1991). Directive 91/440. Brussels: Official Journal of the European Communities,
No L 237/25.

Friebel, G., Ivaldi, M., & Vibes, C. (2010). Railway (De)regulation: A European Efficiency Comparison.
Economica, 77(305), 77-91.

Gallego, G., & van Ryzin, G. (1997). A Multiproduct Dynamic Pricing Problem and Its Applications to
Network Yield Management. Operations Research, Vol. 45, No. 1, 24-41.

Glover, F., Glover, R., Lorenzo, J., & McMillan, C. (1982). The passenger-mix problem in the scheduled
airlines. Interface 12.3, 72-80.

Guo, P., Xiao, B., & Li, J. (2012). Unconstraining Methods in Revenue Management Systems: Research
Overview and Prospects. Hindawi Publishing Corporation: Advances in Operations Research.

Holt, C. C. (2004). Forecasting seasonals and trends by exponentially weighted moving averages.
International Journal of Forecasting 20, 5-10.

Kanafani, A. (1983). Transportation demand analysis. McGraw-Hill, Incorporated.

Kimes, S. E. (1989). Yield Management: A Tool for Capacity-Constrained Service Firm. Journal of
Operations Management Vol. 8, No. 4, 348-363.

Lee, A. O. (1990). Airline Reservations Forecasting: Probabilistic and Statistical Models of the Booking
Process. PhD Thesis, Flight Transportation Laboratory, MIT.

Link, H. (2004). Pep - a yield management scheme for rail passenger fares in Germany. Japan Railways &
Transport Review 38(3), 50-55.

Littlewood, K. (1972). Forecasting and Control of Passenger Bookings. Terrigal, Israel: Agifors
Symposium Proc.

71
Makridakis, S., Wheelwright, S. C., & Hyndman, R. J. (1998). Forecasting methods and applications - 3rd
ed. John Wiley & Sons, Inc.

McGill, J. I. (1989). Optimization and Estimation Problems in Airline Yield Management. PhD Thesis,
Faculty of Commerce and Business Administrations, University of British Columbia.

McGill, J. I., & van Ryzin, G. J. (1999). Revenue Management: Research Overview and Prospects.
Transportation Science 33 (2), 233-256.

Mitev, N. (1996). More than a failure? The computerized reservation systems at French Railways.
Information Technology & People, Vol. 9, 8-19.

Mitev, N. (2004). Trains, planes and computers: from high-speed trains to computerized reservation systems
at French railways. Journal of transport history, 25 (2), 101-123.

Pak, K., & Piersma, N. (2002). Airline Revenue Management: An Overview of OR Techniques 1982-2001.
ERIM Report Series Research in Management.

Phlips, L. (1983). The Economics of Price Discrimination. Cambridge: the Press Syndicate of the University
of Cambridge the Pitt Building.

Rao, P. S. (1978). Forecasting the demand for railway freight services. Journal of Transport Economics and
Policy, 7-26.

Rasouli, S., & Timmermans, H. (2012). Uncertainty in travel demand forecasting models: literature review
and research agenda. Transportation Letters 4.1, 55-73.

Rothstein, M. (1971). An airline overbooking model. Transportation Science 5.2, 180-192.

Sa, J. (1987). Reservation forecasting in airline yield management. Master's Thesis, Flight Transportation
Laboratory, MIT.

Salch, J. (1997). Unconstraining passenger demand using the EM algorithm. Proceedings of the INFORMS
Conference.

Sharif Azadeh, S. (2013). Demand Forecasting in Revenue Management Systems. PhD Thesis, École
Polytechnique de Montréal.

Simpson, R. W. (1989). Using Network Flow Techniques to Find Shadow Prices for Market and Seat
Inventory Control. Flight Transportation Laboratory, MIT.

Smith, B. C., Leimkuhler, J. F., & Darrow, R. M. (1992). Yield Management at American Airlines.
Interfaces, (pp. 8-31).

72
Talluri , K., & van Ryzin, G. (2004b). Revenue Management Under a General Discrete Choice Model of
Consumer Behavior. Management Science 50(1), 15-33.

Talluri, K., & van Ryzin, G. (1998). An Analysis of Bid-Price Controls for Network Revenue Management.
Management Science, Vol 44., part 1, 1577-1593.

Talluri, K., & van Ryzin, G. (2004). The Theory and Practice of Revenue Management. Kluwer Academic
Publishers.

Taneja, N. K. (1978). Airline traffic forecasting: A regression analysis approach. Lexington Books.

Tsai, T. H., Lee, C. K., & Wei, C. H. (2005). Design of dynamic neural networks to forecast short-term
railway passenger demand. Journal of the Eastern Asia Society for Transportation Studies, Vol 6.,
1651-1666.

Walczak, D., Boyd, A. E., & Cramer, R. (2012). Chapter 3 - Revenue Management. In C. Barnhart, & B.
Smith, Quantitative Problem Solving Methods in the Airline Industry - A Modeling Methodology
Handbook. Springer Science+Business Media, LLC.

van Ryzin, G., & Vulcano, G. (2008). Simulation-Based Optimization of Virtual Nesting Controls for
Network Revenue Management. Operations Research 56(4), 865-880.

Wang, Y., Lan, B., & Zhang, L. (2012). A Revenue Management Model for High-Speed Railway.
Proceedings of the 1st International Workshop on High-Speed and Intercity Railways (pp. 95-103).
Springer Berlin Heidelberg.

Wardman, M. (2006). Demand for rail travel and the effects of external factors. Transportation Research
Part E 42, 129-148.

Varian, H. R. (1989). Price discrimination. In R. Schmalensee, & R. D. Willing, Handbook of Industrial

Organization, Volume I (pp. 597-654). Elsevier Science Publisher B.V.

Weatherford, L. R., & Belobaba, P. P. (2002). Revenue Impacts of Fare Input and Demand Forecast
Accuracy in Airline Management. The Journal of the Operational Research Society, Vol. 53, No.
8, 811-821.

Weatherford, L. R., & Bodily, S. E. (1992). A Taxonomy and Research Overview of Perishable-Asset
Revenue Management: Yield Management, Overbooking, and Pricing. Operations Research, Vol.
40, No. 5, 831-844.

Weatherford, L. R., & Kimes, S. E. (2003). A comparison of forecasting methods for hotel revenue
management. International Journal of Forecasting, 19, 401-415.

73
Weatherford, L. R., & Pölt, S. (2002). Better unconstraining of airline demand data in revenue management
systems for improved forecast accuracy and greater revenues. Journal of Revenue and Pricing
Management, Vol. 1, No. 3, 234-254.

Weatherford, L. R., Gentry, T. W., & Wilamowski, B. (2002). Neural network forecasting for airlines: A
comparative analysis. Journal of Revenue and Pricing Management, Vol. 1, No. 4, 319-331.

Wickham, R. R. (1995). Evaluation of Forecasting Techniques for Short-Term Demand of Air

Transportation. Master's Thesis, Flight Transportation Laboratory, MIT.

Williamson, E. L. (1992). Airline Network Seat Inventory Control: Methodologies and Revenue Impacts.
PhD Thesis, Flight Transportation Laboratory, MIT.

Winters, P. R. (1960). Forecasting Sales by Exponentially Weighted Moving Averages. Management

Science, Vol. 6, No. 3, 324-342.

VR Group. (2012). Annual report 2011. Helsinki: VR Group.

VR Group. (2013). Annual report 2012. Helsinki: VR Group.

VR Group. (2014). Annual report 2013. Helsinki: VR Group.

VR Group. (2015). Annual report 2014. Helsinki: VR Group.

Zeni, R. H. (2001). Improved forecast accuracy in revenue management by unconstraining demand

estimates from censored data. PhD Thesis, The State University of New Jersey.

74
Appendices
Appendix 1: Forecasting Scenario 1
Formulations used in forecasts of moving average, exponential smoothing and pickup models are
presented in Section 4.4.

Regression (DB21)
Simple ANOVA – regression using dependent variable Bookings at the departure and independent
variable Bookings at 21 days before departure is estimated using regression tool in Excel. Booking
data available at present time (Week 0) is used in estimation:

SUMMARY OUTPUT
Regression Statistics
Multiple R 0,5944
R Square 0,3533
Adjusted R
Square 0,2725
Standard Error 7,4335
Observations 10

ANOVA
df SS MS F Significance F
Regression 1 241,5493 241,5493 4,3714 0,0699
Residual 8 442,0507 55,2563
Total 9 683,6

Lower Upper
Coefficients Standard Error t Stat P-value 95% 95%
Intercept 22,5042 7,2303 3,1125 0,0144 5,8312 39,1772
Bookings DB21 0,9859 0,4716 2,0908 0,0699 -0,1015 2,0733

Estimated coefficients are β0 = 22,5042 and β1 = 0,9859. Thus the model used in demand forecast-
ing is 𝐵𝑜𝑜𝑘𝑖𝑛𝑔𝑠𝐷𝐵0 = 22,5042 + 0,9859 ∗ 𝐵𝑜𝑜𝑘𝑖𝑛𝑔𝑠𝐷𝐵21.

According to P-values both coefficients are statistically significant at 10 % significance level. How-
ever low R2 indicates that the estimated model fits quite badly to my hypothetical dataset. But
because the purpose of these forecast examples is only to provide illustrations of each method, I
will not consider the reasons of this problem more closely.

75
 Forecasts and errors for each week in Scenario 1 using selected methods:
Forecasts of Week 1 bookings at DB0 Forecasts of Week 5 bookings at DB0
Method Forecast Actual Error Sq. Error Abs % - Error Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 37 49 12 149 25 % Moving average (n=10) 37 40 3 10 8%
Exponential smoothing (α=0,5) 41 49 8 58 16 % Exponential smoothing (α=0,5) 41 40 -1 2 4%
Exponential smoothing (α=0,2) 39 49 10 107 21 % Exponential smoothing (α=0,2) 39 40 1 2 3%
Additive pickup 44 49 5 28 11 % Additive pickup 37 40 3 12 9%
Multiplicative pickup 45 49 4 14 8% Multiplicative pickup 36 40 4 19 11 %
Regression (DB21) 39 49 10 95 20 % Regression (DB21) N/A 40 N/A N/A N/A

Forecasts of Week 2 bookings at DB0 Forecasts of Week 6 bookings at DB0

Method Forecast Actual Error Sq. Error Abs % - Error Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 37 25 -12 139 47 % Moving average (n=10) 37 35 -2 3 5%
Exponential smoothing (α=0,5) 41 25 -16 269 66 % Exponential smoothing (α=0,5) 41 35 -6 41 18 %
Exponential smoothing (α=0,2) 39 25 -14 186 55 % Exponential smoothing (α=0,2) 39 35 -4 13 10 %
Additive pickup 26 25 -1 0 3% Additive pickup 36 35 -1 1 3%
Multiplicative pickup 17 25 8 57 30 % Multiplicative pickup 31 35 4 19 12 %
Regression (DB21) 27 25 -2 6 10 % Regression (DB21) N/A 35 N/A N/A N/A

Forecasts of Week 3 bookings at DB0 Forecasts of Week 7 bookings at DB0

Method Forecast Actual Error Sq. Error Abs. % - Error Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 37 30 -7 46 23 % Moving average (n=10) 37 39 2 5 6%
Exponential smoothing (α=0,5) 41 30 -11 130 38 % Exponential smoothing (α=0,5) 41 39 -2 6 6%
Exponential smoothing (α=0,2) 39 30 -9 75 29 % Exponential smoothing (α=0,2) 39 39 0 0 1%
Additive pickup 39 30 -9 86 31 % Additive pickup 37 39 2 3 5%
Multiplicative pickup 43 30 -13 173 44 % Multiplicative pickup 46 39 -7 49 18 %
Regression (DB21) 39 30 -9 86 31 % Regression (DB21) N/A 39 N/A N/A N/A

Forecasts of Week 4 bookings at DB0 Forecasts of Week 8 bookings at DB0

Method Forecast Actual Error Sq. Error Abs. % - Error Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 37 23 -14 190 60 % Moving average (n=10) 37 45 8 67 18 %
Exponential smoothing (α=0,5) 41 23 -18 339 80 % Exponential smoothing (α=0,5) 41 45 4 13 8%
Exponential smoothing (α=0,2) 39 23 -16 245 68 % Exponential smoothing (α=0,2) 39 45 6 40 14 %
Additive pickup 35 23 -12 139 51 % Additive pickup 36 45 9 76 19 %
Multiplicative pickup 29 23 -6 41 28 % Multiplicative pickup 74 45 -29 818 64 %
Regression (DB21) N/A 23 N/A N/A N/A Regression (DB21) N/A 45 N/A N/A N/A

76
 Forecasts and errors using selected methods in Scenario 1: Weeks 1 - 8
Moving average (n=10) Additive pickup
Week Forecast Actual Error Sq.Error Abs. % - Error Week Forecast Actual Error Sq.Error Abs. % - Error
1 37 49 -12 149 25 % 1 44 49 -5 28 11 %
2 37 25 12 139 47 % 2 26 25 1 0 3%
3 37 30 7 46 23 % 3 39 30 9 86 31 %
4 37 23 14 190 60 % 4 35 23 12 139 51 %
5 37 40 -3 10 8% 5 37 40 -3 12 9%
6 37 35 2 3 5% 6 36 35 1 1 3%
7 37 39 -2 5 6% 7 37 39 -2 3 5%
8 37 45 -8 67 18 % 8 36 45 -9 76 19 %
Mean 37 36 1 76 24 % Mean 36 36 0 43 16 %

Exponential smoothing (α=0,5) Multiplicative pickup

Week Forecast Actual Error Sq.Error Abs. % - Error Week Forecast Actual Error Sq.Error Abs. % - Error
1 41 49 -8 58 16 % 1 45 49 -4 14 8%
2 41 25 16 269 66 % 2 17 25 -8 57 30 %
3 41 30 11 130 38 % 3 43 30 13 173 44 %
4 41 23 18 339 80 % 4 29 23 6 41 28 %
5 41 40 1 2 4% 5 36 40 -4 19 11 %
6 41 35 6 41 18 % 6 31 35 -4 19 12 %
7 41 39 2 6 6% 7 46 39 7 49 18 %
8 41 45 -4 13 8% 8 74 45 29 818 64 %
Mean 41 36 6 107 29 % Mean 40 36 4 149 27 %

Exponential smoothing (α=0,2) Regression (DB21)

Week Forecast Actual Error Sq.Error Abs. % - Error Week Forecast Actual Error Sq.Error Abs. % - Error
1 39 49 -10 107 21 % 1 39 49 -10 95 20 %
2 39 25 14 186 55 % 2 27 25 2 6 10 %
3 39 30 9 75 29 % 3 39 30 9 86 31 %
4 39 23 16 245 68 % 4 N/A 23 N/A N/A N/A
5 39 40 -1 2 3% 5 N/A 40 N/A N/A N/A
6 39 35 4 13 10 % 6 N/A 35 N/A N/A N/A
7 39 39 0 0 1% 7 N/A 39 N/A N/A N/A
8 39 45 -6 40 14 % 8 N/A 45 N/A N/A N/A
Mean 39 36 3 84 25 % Mean 35 36 1 62 20 %

77
Appendix 2: Forecasting Scenario 2
Formulations used in forecasts of moving average, exponential smoothing and pickup models are
presented in Section 4.4.

Regression (DB21)
Simple ANOVA – regression using dependent variable Bookings at the departure and independent
variable Bookings at 21 days before departure is estimated using Excel. Data over 18 weeks is
used to estimate the model parameters:

SUMMARY OUTPUT
Regression Statistics
Multiple R 0,6593
R Square 0,4347
Adjusted R Square 0,3994
Standard Error 6,7512
Observations 18

ANOVA
df SS MS F Significance F
Regression 1 560,7469 560,7469 12,3029 0,0029
Residual 16 729,2531 45,5783
Total 17 1290

Lower Upper
Coefficients Standard Error t Stat P-value 95% 95%
Intercept 19,9735 4,9281 4,0529 0,0009 9,5263 30,4207
Bookings DB21 1,1283 0,3217 3,5076 0,0029 0,4464 1,8102

Estimated coefficients are β0 = 19,9735 and β1 = 1,1283. Thus the model used in demand forecast-
ing is 𝐵𝑜𝑜𝑘𝑖𝑛𝑔𝑠𝐷𝐵0 = 19,9735 + 1,1283 ∗ 𝐵𝑜𝑜𝑘𝑖𝑛𝑔𝑠𝐷𝐵21.

According to P-values both coefficients are statistically significant at 5 % significance level. How-
ever low R2 indicates that the estimated model fits quite badly to my hypothetical dataset. But like
in Scenario 1 I will not consider the reasons of this problem more closely, because the purpose of
this example is to provide illustrations of how each method is compiled.

78
 Forecasts and errors for each week in Scenario 2 using selected methods:
Forecasts of Week 1 bookings at DB0 Forecasts of Week 5 bookings at DB0
Method Forecast Actual Error Sq. Error Abs. % - Error Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 37 49 12 149 25 % Moving average (n=10) 38 40 3 6 6%
Exponential smoothing (α=0,5) 41 49 8 58 16 % Exponential smoothing (α=0,5) 28 40 12 149 31 %
Exponential smoothing (α=0,2) 39 49 10 107 21 % Exponential smoothing (α=0,2) 33 40 7 43 16 %
Additive pickup 44 49 5 28 11 % Additive pickup 41 40 -1 1 3%
Multiplicative pickup 45 49 4 14 8% Multiplicative pickup 42 40 -2 4 5%
Regression (DB21) 39 49 10 97 20 % Regression (DB21) 38 40 2 4 5%

Forecasts of Week 2 bookings at DB0 Forecasts of Week 6 bookings at DB0

Method Forecast Actual Error Sq. Error Abs. % - Error Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 39 25 -14 202 57 % Moving average (n=10) 38 35 -3 9 9%
Exponential smoothing (α=0,5) 45 25 -20 408 81 % Exponential smoothing (α=0,5) 34 35 1 1 3%
Exponential smoothing (α=0,2) 41 25 -16 247 63 % Exponential smoothing (α=0,2) 35 35 0 0 1%
Additive pickup 30 25 -5 21 18 % Additive pickup 36 35 -1 1 3%
Multiplicative pickup 27 25 -2 5 9% Multiplicative pickup 36 35 -1 1 2%
Regression (DB21) 26 25 -1 0 2% Regression (DB21) 35 35 0 0 1%

Forecasts of Week 3 bookings at DB0 Forecasts of Week 7 bookings at DB0

Method Forecast Actual Error Sq. Error Abs. % - Error Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 39 30 -9 76 29 % Moving average (n=10) 38 39 1 2 4%
Exponential smoothing (α=0,5) 35 30 -5 26 17 % Exponential smoothing (α=0,5) 34 39 5 21 12 %
Exponential smoothing (α=0,2) 38 30 -8 57 25 % Exponential smoothing (α=0,2) 35 39 4 18 11 %
Additive pickup 28 30 2 4 7% Additive pickup 40 39 -1 1 3%
Multiplicative pickup 26 30 4 19 15 % Multiplicative pickup 41 39 -2 3 5%
Regression (DB21) 39 30 -9 84 31 % Regression (DB21) 44 39 -5 22 12 %

Forecasts of Week 4 bookings at DB0 Forecasts of Week 8 bookings at DB0

Method Forecast Actual Error Sq. Error Abs. % - Error Method Forecast Actual Error Sq. Error Abs. % - Error
Moving average (n=10) 39 23 -16 262 70 % Moving average (n=10) 37 45 8 64 18 %
Exponential smoothing (α=0,5) 33 23 -10 91 42 % Exponential smoothing (α=0,5) 37 45 8 69 18 %
Exponential smoothing (α=0,2) 36 23 -13 170 57 % Exponential smoothing (α=0,2) 36 45 9 88 21 %
Additive pickup 30 23 -7 45 29 % Additive pickup 34 45 11 119 24 %
Multiplicative pickup 27 23 -4 19 19 % Multiplicative pickup 34 45 11 132 25 %
Regression (DB21) 30 23 -7 51 31 % Regression (DB21) 40 45 5 22 10 %

79
 Forecasts and errors using selected methods in Scenario 2: Weeks 1 – 8
Moving average (n=10) Additive pickup
Week Forecast Actual Error Sq. Error Abs. % - Error Week Forecast Actual Error Sq. Error Abs. % - Error
1 37 49 12 149 25 % 1 44 49 5 28 11 %
2 39 25 -14 202 57 % 2 30 25 -5 21 18 %
3 39 30 -9 76 29 % 3 28 30 2 4 7%
4 39 23 -16 262 70 % 4 30 23 -7 45 29 %
5 38 40 3 6 6% 5 41 40 -1 1 3%
6 38 35 -3 9 9% 6 36 35 -1 1 3%
7 38 39 1 2 4% 7 40 39 -1 1 3%
8 37 45 8 64 18 % 8 34 45 11 119 24 %
Mean 38 36 -2,25 96 27 % Mean 35 36 0,41 28 12 %

Exp. Smooth. (0,5) Multip. pickup

Week Forecast Actual Error Sq. Error Abs. % - Error Week Forecast Actual Error Sq. Error Abs. % - Error
1 41 49 8 58 16 % 1 45 49 4 14 8%
2 45 25 -20 408 81 % 2 27 25 -2 5 9%
3 35 30 -5 26 17 % 3 26 30 4 19 15 %
4 33 23 -10 91 42 % 4 27 23 -4 19 19 %
5 28 40 12 149 31 % 5 42 40 -2 4 5%
6 34 35 1 1 3% 6 36 35 -1 1 2%
7 34 39 5 21 12 % 7 41 39 -2 3 5%
8 37 45 8 69 18 % 8 34 45 11 132 25 %
Mean 36 36 -0,13 103 27 % Mean 35 36 1,03 25 11 %

Exp. Smooth. (0,2) Reg. (DB21)

Week Forecast Actual Error Sq. Error Abs. % - Error Week Forecast Actual Error Sq. Error Abs. % - Error
1 39 49 10 107 21 % 1 39 49 10 97 20 %
2 41 25 -16 247 63 % 2 26 25 -1 0 2%
3 38 30 -8 57 25 % 3 39 30 -9 84 31 %
4 36 23 -13 170 57 % 4 30 23 -7 51 31 %
5 33 40 7 43 16 % 5 38 40 2 4 5%
6 35 35 0 0 1% 6 35 35 0 0 1%
7 35 39 4 18 11 % 7 44 39 -5 22 12 %
8 36 45 9 88 21 % 8 40 45 5 22 10 %
Mean 36 36 -0,70 91 27 % Mean 36 36 -0,58 35 14 %

80
Appendix 3: EMSR – capacity allocation
Forecasted Allocated seats Allocated seats Allocated seats
Class (i) Fare(i)
Mean Std. Dev. (S,i) (D,i) (E,B)
B fB µB σB 𝐸𝑀𝑆𝑅𝐵 𝑠𝑆𝐵 = 𝑓𝐸 𝐸𝑀𝑆𝑅𝐵 𝑠𝐷𝐵 = 𝑓𝐸 𝐸𝑀𝑆𝑅𝐸 𝑠𝐸𝐵 = 𝑓𝐸
E fE µE σE 𝐸𝑀𝑆𝑅𝐸 𝑠𝑆𝐸 = 𝑓𝐷 𝐸𝑀𝑆𝑅𝐸 𝑠𝐷𝐸 = 𝑓𝐷
D fD µD σD 𝐸𝑀𝑆𝑅𝐷 𝑠𝑆𝐷 = 𝑓𝑆
S fS µS σS
Allocated seats (s): 𝑠𝑆𝑖 𝑠𝐷𝑖 𝑠𝐸𝐵

An example of determination of allocated seats: 𝒔𝑫

𝑺

The number of seats allocated to fare class D from fare class S are presented by 𝒔𝑫
𝑺 . The number

of allocated seats comes from the equation 𝑬𝑴𝑺𝑹𝑫 (𝒔𝑫

𝑺 ) = 𝒇𝑺 .

This means that the number of allocated seats is the number of seats that equals the previous equa-
tion.

Because EMSR –model assumes that demand for each fare class is normally distributed, EMSR
can be presented by 𝑬𝑴𝑺𝑹𝑫 (𝒔𝑫
𝑺 ) = 𝒇𝑫 ∙ 𝑷𝑫 (𝒔𝑫 ).

Thus the problem is to determine the value of 𝒔𝑫 ̅ 𝑫

𝑺 for which: 𝒇𝑺 = 𝒇𝑫 ∙ 𝑷𝑫 (𝒔𝑺 )

Following table will describe how the number of allocated seats can be found:

Fare S Fare D Mean Std. Dev.

fS fD µD σD
s CDF 1 - CDF EMSR
1,2,3,…,s 𝑃𝐷 (𝑠𝑆𝐷 ) 𝑃𝐷 (𝑠𝑆𝐷 ) 𝑓𝐷 𝑃̅𝐷 (𝑠𝑆𝐷)

CDF is the cumulative density function of selling s seats. In EMSR –model normal distribution
N (µ, σ) is used to create CDF.

As all allocated seats between each fare classes have been calculated, the sum of allocated seats
for each class is determined by ∑ 𝑠𝑆𝑖 , ∑ 𝑠𝐷𝑖 , 𝑠𝐸𝐵 . These represent protection levels of fare classes
D, E and B.

81
Then each booking limit can be determined by:

Booking limit = Total capacity - Allocated seats (s)

Class: B E D S
Booking limits: 𝐶 𝐶 − 𝑠𝐸𝐵 𝐶− 𝑠𝐷𝑖 𝐶− 𝑠𝑆𝑖

Calculated protection levels and bookings limits for each case:

Case 1: Capacity allocation Total capacity: 150

Class Fare Joint protection Booking limits Protection level
B 100 € 14 150 14
E 60 € 48 136 34
D 30 € 96 102 49
S 20 € - 54 -
Case 2: Capacity allocation Total capacity: 150
Class Fare Joint protection Booking limits Protection level
B 100 € 36 150 36
E 60 € 78 114 41
D 30 € 117 73 39
S 20 € - 33 -
Case 3: Capacity allocation Total capacity: 150
Class Fare Joint protection Booking limits Protection level
B 100 € 29 150 29
E 60 € 93 121 64
D 30 € 194 58 58
S 20 € - 0 -
Case 4: Capacity allocation Total capacity: 150
Class Fare Joint protection Booking limits Protection level
B 100 € 7 150 7
E 60 € 40 143 32
D 30 € 82 111 43
S 20 € - 68 -

82