MIT
International
MIT
ICAT
Center
for
Air
Transportation
WILLINGNESS TO PAY AND COMPETITIVE
REVENUE MANAGEMENT
Peter P. Belobaba
With help from the MIT PODS Team:
Maital Dar, Valenrina Soo, Thierry Vanhaverbeke
12th PROS Conference
Houston, TX
March 5-8, 2006
MIT
ICAT
Traditional RM Systems Are Struggling
Under Fare Simplification
y Simplified fare structures characterized by
ƒ One-way fares with little or no product differentiation, priced at
different fare levels
ƒ Existing RM systems employed to control number of seats sold at
each fare level
y But RM systems were developed for restricted fares
ƒ Assumed independent fare class demands, because restrictions
kept full-fare passengers from buying lower fares
ƒ Time series forecasting models used to predict future demand
based on historical bookings in each fare class
ƒ Given independent demand forecasts, top-down protection for
highest classes, extra seats made available to lowest class
2
MIT
ICAT
Revenue Impacts of Fare Simplification
with Traditional RM Models
65,000
60,000
.5%
50,000
.8%
-16
-0
55,000
.6%
-29
45,000
40,000
%
-45
35,000
30,000
Fully
Restricted
Remove AP
Remove Sat
Night Min Stay
Remove All
Restr, Keep
AP
Remove All
Restr and AP
3
MIT
ICAT
Traditional RM Models “Spiral Down”
without Product Differentiation
100
88.1
87.8
90
81.6
79.8
82.7
80
70
60
50
40
FC 6
30
FC 5
20
FC 4
FC 3
10
FC 2
0
Fully Restricted
Remove AP
Remove Sat
Night Min Stay
Remove All
Restr, Keep AP
Remove All Restr
and AP
FC 1
4
MIT
ICAT
Existing Airline RM Systems Need to be
Modified for Changing Fare Structures
y Without modification, these RM systems will not
maximize revenues in less restricted fare structures
ƒ Unless demand forecasts are adjusted to reflect potential sell-up,
high-fare demand will be consistently under-forecast
ƒ Optimizer then under-protects, allowing more “spiral down”
y Current RM system limitations are negatively
affecting airline revenues
ƒ Existing systems, left unadjusted, generate high load factors but
do not increase yields
ƒ Many legacy carriers are using “rule-based” RM practices
y RM forecasting models must be changed to reflect
passenger willingness to pay (WTP)
5
MIT
ICAT
Forecasting by Willingness to Pay
y With undifferentiated fares, forecasting must focus
on demand by willingness to pay higher fares
ƒ Approach is to “transform” all historical bookings at different fares
to maximum demand potential at lowest fare
y Total demand potential at lowest fare converted to
demand forecasts by fare level for future flights
ƒ Based on estimates of sell-up potential/WTP
y But, most simplified fare structures still retain some
product differentiation
ƒ Lowest fares can have more several cancel/change restrictions
ƒ Higher fares can offer full flexibility and additional amenities
6
MIT
ICAT
Hybrid Forecasting For Partially
Differentiated Fare Structures
• Generate separate forecasts for price (“priceable”)
and product (“yieldable”) oriented demand
A passenger is price-oriented if the next lower class from the one
booked is closed
A passenger is product-oriented if the next lower class from the
one booked was open.
• Combine standard forecasts and WTP forecasts for
input to RM optimizers
For product-oriented demand, bookings are treated as a historical
data for the given class, and standard time series forecasting
applied.
For price-oriented demand, forecasts by WTP based on expected
sell-up behavior
7
MIT
ICAT
PODS Simulations in Network D
y Simplified Fare Structure
ƒ 6 fare classes
ƒ Compressed fare ratio of 4.1
y 2 competing hub airlines
ƒ 40 Spoke Cities
ƒ 252 legs
ƒ 482 OD markets
Class
Average Fare
1
412.85
2
293.34
Class
AP
Min
Stay
Chg
Fee
Non
Ref
1
0
0
0
0
2
3
0
1
0
3
7
0
1
1
4
14
0
1
1
5
14
0
1
1
6
21
0
1
1
3
179.01
4
153.03
5
127.05
6
101.06
8
MIT
ICAT
PODS “Network D”
1
2
H1(41)
3
33
4
5
6
23
7
28
8
9
22
11
10
14
16
12
17
32
40
24
29
34
26
31
37
21
27
35
38
15
18
13
36
H2(42)
19
20
30
25
39
Traffic Flows
9
MIT
ICAT
Baseline Results: Standard RM Models
Standard pickup (moving average) forecaster, booking curve
unconstraining and EMSRb optimizer
Revenue is $1,124,782 ; Yield is $0.1061
Load factor is 86.45%
Loads by Fare Class
30.0
25.9
25.0
23.7
20.0
13.8
15.0
10.0
10.2
5.6
7.4
5.0
0.0
1
2
3
4
5
6
Fare Class
10
MIT
ICAT
Spiral Down is Evident in Standard Forecasts
Forecast + Bookings by Fare Class
Very High Initial
Forecasts in
Classes 5 + 6
140
120
100
80
60
40
20
0
1
2
3
4
FC1
5
6
FC2
7
FC3
8
9 10 11 12 13 14 15 16
FC4
FC5
FC6
11
MIT
ICAT
Airline 1 Implements Hybrid Forecasting
Airline 1
∆ from
Trad RM
Airline 2
∆ from
Trad RM
$1,163,408
+ 3.4%
$1,118,299
- 0.2%
Yield
$0.1107
+ $0.005
$0.1016
- $0.001
Loads
85.70%
- 0.75
86.40%
+ 0.26
Revenue
Loads by Fare Class
30.0
25.0
Hybrid
Standard
20.0
15.0
10.0
5.0
0.0
1
2
3
Fare Class
4
5
6
12
MIT
ICAT
Both Competitors Use Hybrid Forecasting
Airline 1
Revenue
∆ from
Trad RM
Airline 2
∆ from
Trad RM
$1,158,167
+ 2.97%
$1,152,222
+ 3.03%
Yield
$0.1098
+ $0.004
$0.1055
+ $0.004
Loads
85.95%
- 0.5
85.72%
- 0.68
Loads by Fare Class
30.0
25.0
Hybrid
Standard
20.0
15.0
10.0
5.0
0.0
1
2
3
Fare Class
4
5
6
13
MIT
ICAT
WTP Approach Brings Forecasts Back into Line
Forecast + Bookings by Fare Class
120
100
80
60
Higher
Forecasts40
Classes 1-3
20
0
1
2
3
4
FC1
5
6
FC2
7
FC3
8
9 10 11 12 13 14 15 16
FC4
FC5
FC6
14
MIT
ICAT
Distorted Forecasts Affect RM Performance
y Spiral-down leads to high forecasts in lower classes
ƒ And, in turn, forecasts that are too low in higher classes
y In EMSR-based leg RM, at least the lowest class
demand is rejected when demand is high
ƒ Not revenue maximizing, but some benefit from booking limits
y Recent PODS simulations illustrate impacts of
distorted forecasts on O+D method performance:
ƒ Mismatch between independent path/class demands assumed by
network optimizer and reality of passenger sell-up
ƒ Network optimizers over-protect for forecast low-class connecting
traffic, leading to distorted bid prices or displacement costs
15
MIT
ICAT
Impacts on O+D Revenue Gains Under
Simplified Fare Structure
y Revenue performance of O+D methods is affected:
Revenue Gain over EMSRb
ƒ Standard forecasting assumes path/class independence
ƒ Incorrect forecasts fed into network optimizers (LP, ProBP)
ƒ Network optimization methods more affected than Heuristic BP
2.00%
1.93%
1.75%
1.50%
1.25%
1.00%
0.75%
0.50%
0.74%
0.43%
0.25%
0.00%
DAVN
PRO BP
H BP
16
MIT
ICAT
All RM Methods Benefit From Hybrid
Forecasting
y Use of Hybrid Forecasting with WTP component improves RM
revenue gains by 2-4%
y O+D RM Methods once again outperform EMSRb leg controls
by 1% or more
Revenue Gain over EMSRb
2.6%
4.0%
3.3%
3.7%
4.5%
4.0%
3.3%
3.5%
2.9%
3.0%
2.5%
2.0%
1.7%
2.6%
1.93%
1.5%
1.0%
0.5%
0.74%
0.0%
EMSRb
DAVN
0.43%
ProBP
H BP
17
MIT
ICAT
Simulations in Larger Network R
4 Competitors with 4 Hubs
y All methods benefit from Hybrid Forecasting
ƒ ProBP again is most affected by simplified fare structure, but
benefits most from use of Hybrid Forecasts based on WTP
Revenue Gain over EMSRb
1.3%
2.4%
3.0%
2.4%
3.5%
3.0%
2.3%
2.5%
0.8%
2.0%
1.9%
1.5%
1.0%
1.3%
1.51%
0.5%
0.65%
0.44%
0.0%
EMSRb
DAVN
ProBP
H BP
18
MIT
ICAT
Existing RM Forecasts Are Inadequate for
Simplified Fare Structures
y Existing RM systems need to be modified
ƒ Mismatch between RM model assumptions and fare structures
y Price/product hybrid forecasting increases revenues
ƒ Compared to use of standard RM forecasting methods
ƒ Gains come from higher forecasts in upper/middle classes,
increasing protection and helping to reduce “spiral down”
y Modified forecasters require estimates of passenger
WTP by time to departure for each flight
ƒ Approach is to forecast maximum demand potential at lowest
fare, and convert into “partitioned” forecasts for each fare class
y But, WTP forecasting is much more difficult…
19
MIT
ICAT
Standard RM Forecasts Assume
Independent Demands by Class
Demand observable
in each DCP in any
fare class
Fare class
1
2
3
4
5
6
1
2
3
4
5
6
7
8
DCP
Current
DCP
y Product-oriented demand observed for any open
class by DCP
20
MIT
ICAT
Forecasts of Demand to Come by Class
Used as Inputs by Optimizer
Demand to
come in any
fare class
Fare class
1
2
3
4
5
6
1
2
3
4
5
6
7
8
DCP
Current
time-frame
y Demand to come by class assumed to be
independent of what classes are open or closed.
21
MIT
ICAT
Price-Oriented Demand for
Undifferentiated Fares
Historical information obtained
when j was the lowest open class
Fare class
1
2
3
4
5
6
1
2
3
4
5
6
7
8
DCP
y On a single flight departure, bookings in each class
observed only when lower class was closed down.
22
MIT
ICAT
Sample of Price-Oriented Demand over
Multiple Departures
Fare class
1
2
3
4
5
6
1
2
3
4
5
6
7
8
DCP
y With information about class closures and observed
bookings, we can estimate WTP and sell-up
y But we can’t estimate what we’ve never seen!
23
MIT
ICAT
Factors Affecting WTP Estimates
y Time to departure
ƒ Business travelers with higher WTP book closer to departure date
y Peak vs. off-peak periods
ƒ Higher demand periods tend to have higher WTP
y Market characteristics
ƒ Limited competition and low capacity (relative to demand) means
a higher WTP among consumers
y RM seat availability – your own and your competitors’
ƒ Weak RM controls on previous flights mean historical bookings
don’t reflect higher WTP of consumers
ƒ High availability of low-fare seats on competitor makes it difficult
(impossible) to observe high-fare bookings
24
MIT
ICAT
Willingness to Pay Relative to Lowest
Fare Changes over the Booking Process
6.00
HIGH
5.00
4.00
MEDIUM
3.00
2.00
LOW
1.00
0.00
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
BOOKING PERIODS (DCPs)
25
MIT
ICAT
WTP Estimates are Critical to Effective
RM under Simplified Fares
y Nothing simple about RM models required by these
simplified fare structures
ƒ Traditional RM methods do not maximize revenues
ƒ Without forecasting modifications, even O+D control gains are
affected
y New approaches to “hybrid” forecasting of price- vs.
product-oriented demand show good potential
ƒ Significant revenue gains over standard forecasting methods
y Recent PODS research shows potential for
conditional WTP estimates:
ƒ Requires separate estimation of WTP for each scenario of class
closure, for own airline and potentially competitor(s)
26