1. PODS Update Large Network O-D Control Results
Peter Belobaba and Seonah Lee
Massachusetts Institute of Technology
AGIFORS RM STUDY GROUP MEETING
New York City
March 22-24, 2000

2. Outline Description of New Large PODS Network
Standardization of RM and O-D Method Parameters
DAVN parameters – re-optimization, virtual bucket definition
Re-optimizing rate for bid price methods (HBP and PROBP)
Results: O-D Revenue Gain Comparisons
Impacts of Average Load Factors and Distributions
Overview of Additional PODS Studies
Use of Path-Based (ODF) Forecasts in Leg/Bucket RM
Introduction of Cancellation and No-Show Behaviors
Recovery of RM Methods from Sudden Demand Shocks

3. Characteristics of Large Network
40 spoke cities with 2 hubs, one for each airline
20 spoke cities on each side, located by geographical coordinates of actual US cities
Distance to 1514 miles to the hub from spoke cities
Unidirectional -- West to east flow of traffic
Inter-hub services -- one for each direction, for each bank, for each airline
3 banks starting at 10:30, 14:00, 17:30 for each airline hub
252 flight legs, 482 O-D markets, 4 fare types per market

4. Geographical Layout H1(41) H2(42) 1 2 21 3 4 5 25 6 23 24 27 26 7 3128 30 8 29 32 33 22 9 11 34 35 38 10 12 14 15 13 16
H2(42) 36 17 18 37 19 39 20 40

5. Standardization of O-D RM Methods
“Generic” RM method parameters defined 3 years ago for smaller PODS networks (6-10 cities):
4 fare classes for Base Case EMSRb Control
6 virtual buckets per leg for GVN, HBP and DAVN
Network-wide virtual range definitions
Varying re-optimization rates for bid price methods
For new 40-city network, we updated RM methods:
“Standard” definitions to better reflect actual and feasible implementations of each method

6. Standardized RM Method Parameters
FCYM -- Fare Class Yield Management
4 fare classes grouped by yields and fare restrictions
Leg/class demand data and forecasting
EMSRb limits -- Re-optimize at 16 checkpoints
GVN -- Greedy Virtual Nesting
ODFs mapped to 8 virtual buckets based on total itinerary fare values
Network-wide virtual ranges for all legs
Leg/bucket demand data and forecasting

7. Standardized RM Method Parameters
HBP -- Heuristic Bid Price
Like GVN, ODFs mapped to 8 virtual buckets based on total itinerary fare values
Same network-wide virtual ranges for all legs
Leg/bucket demand data and forecasting
EMSRb booking limit control for local (one-leg) itineraries -- re-optimized 16 times before departure
“Bid price” control for connecting requests based on current EMSR values of last seat on each leg:
Re-optimized daily over 63-day PODS booking period

8. Standardized RM Method Parameters
DAVN -- Displacement Adjusted Virtual Nesting
ODFs mapped to 8 virtual buckets based on displacement adjusted “network” revenue values:
Network Value = ODF Fare - Displacement Cost
Leg Displacement Costs estimated by shadow prices of deterministic network LP optimization
Network re-optimized at each checkpoint (16 times)
Leg-specific virtual bucket range definitions
ODF demand forecasting (rolled up to leg/bucket)
EMSRb control of leg/buckets checkpoints

9. Standardized RM Method Parameters
PROBP--Probabilistic Network Bid Price
Nested probabilistic network convergence algorithm developed at MIT (Bratu, 1998)
Involves “prorating” total ODF value to legs traversed:
Requires ODF data demand forecasts
Estimates “critical EMSR operator” for each leg by accounting for complete nesting of ODF availabilities
Critical EMSR values used as additive bid prices for local and connecting path requests
Re-optimized daily over 63-day PODS booking period

10. Summary of New RM Parameters
Base Case Fare Class YM effectively unchanged
Enhancements to virtual bucket methods:
Number of virtual buckets increased to 8
More frequent network displacement optimization and leg-specific virtual re-bucketing for DAVN
Represents “advanced” implementations of DAVN
More realistic bid price re-optimization frequency:
Airline consensus that daily bid price updates are feasible in larger networks
Theoretically more frequent updates might be misleading

11. Demand and Load Factors Simulated
Under FCYM Base Case, simulated demand factors led to network ALFs from 70% to 87%
Load factor distributions compared well with system data provided by 2 airlines
Local traffic represents 37 to 40% of total load by flight leg, on average:
Varies by demand factor and RM methods used
Differences in load factors by connecting bank at each hub:
Highest for mid-day bank, lowest early in morning

13. ALFs by Hub Connecting Bank
3 banks per day offered at each airline’s hub:
Range of ALFs and revenue gains for each RM method
Most realistic traffic characterization in PODS to date

14. Revenue Gains over FCYM (Competitor uses FCYM)

15. Comparison of O-D Revenue Gains
Relative performance in line with smaller network:
Small gains for GVN, negative at higher demands
HBP revenue improvements over “greediness” of GVN
DAVN and PROBP perform best, gains of 1% or more
But, overall % gains of O-D methods are lower:
New network not designed to be “O-D friendly”
Each demand factor includes a range of ALFs by bank, with lower % gains for lower demand banks
More path choices without airline preferences or re-planning disutilities result in greater passenger shifts among paths

16. Revenue Gains by Connecting Bank (Network ALF=83%, Competitor uses FCYM)

17. Competitive Impacts of O-D Methods (Network ALF=83%, Competitor uses FCYM)

18. Competitive Impacts of O-D Methods
O-D control can have substantial revenue impacts on competitor:
Continued use of FCYM against O-D methods results in revenue losses for Airline B
Interesting is GVN result, where Airline B’s revenue loss is greater than Airline A’s gain
Still not a zero-sum game, as revenue gains of Airline A exceed revenue losses of Airline B
Other simulation results show both airlines can benefit from using more sophisticated O-D control

19. Lessons from Larger Network
Demand characteristics affect O-D benefits:
No explicit effort to design “bottleneck” legs that favor GVN
More realistic distribution of load factors across legs
Different load factors for connecting banks by time of day
Misleading to focus comparisons on peak connecting banks
Characterization of O-D methods also critical:
More sophisticated DAVN parameters, more realistic PROBP re-optimization frequency
Robustness of DAVN even with periodic re-optimization
O-D control has important competitive impacts

20. Large Network in PODS: Next Steps
Alternative demand and network characteristics:
Proportion of local vs. connecting O-D demand
Load factor distributions
Business vs. leisure traffic mix
Impacts of passenger choice disutility parameters:
Increase re-planning costs for changing preferred times
Modify airline preference factors from 50/50
Introduce path quality options (non-stops) and disutilities
Less structured and more “realistic” O-D fares:
Not necessarily tied to O-D market distances

21. Overview of Other PODS Studies
Path-Based (ODF) Forecasting in Leg-Based RM
Introduction of Cancellation and No-Show Rates
Impacts of Sudden Demand Shocks
Competitive Studies Planned and Under Way

22. Path-Based Forecasting in Leg RM
Preliminary results show potential gains from use of path-based (ODF) forecasts in leg-based RM:
ODF database to keep historical booking data
Tested simple moving average “pick-up” forecasts with “booking curve” unconstraining
ODF forecasts “rolled up” to leg/class or leg/bucket
ODF forecasts not necessarily more “accurate”:
Error relative to mean forecast is large due to small numbers
But ability to unconstrain demand by ODF path appears to contribute in large part to revenue gains

23. Example: Path Forecasts for Leg RM (Previous Large Network ALF=75%)

24. Cancellation and No-show Rates
Over past several months, we have incorporated cancellation and no-show processes into PODS:
“Memory-less” daily cancellation probability
Gaussian distributions of no-show rates at departure
Probabilistic overbooking model to determine AUs
Neither process has a large impact on revenue gains of O-D methods:
Relative performance of methods stays the same at similar load factors; O-D methods do slightly better at lower ALFs
Now testing gross vs. net booking forecast models

25. Impacts of Sudden Demand Shock
Simulated “overnight” demand shifts of +/- 20%:
Extreme test of robustness of each RM method to changes in actual demand vs. forecast
Compared percentage revenue gains of each method vs. FCYM before and after demand shock
After 20% sudden demand decrease:
GVN benefited, showing immediate revenue increase
DAVN and PROP suffered, due to over-forecasts by ODF
HBP maintained relative revenue gains
Relative performance stabilized after samples

27. Competitive Studies with PODS
Introduction of third “new entrant” airline in one or more spoke-hub local markets:
What are impacts on hub carrier that uses leg vs. O-D RM?
What are “rational” vs. “predatory” responses by hub carrier in terms of prices, capacity and RM controls?
System-wide reduction of aircraft capacity (6%?) by one hub airline to increase legroom:
Revenue and load impacts with leg-based vs. O-D RM?
What increase in airline preference is needed to make up for revenue losses?

28. Summary: PODS RM Research
After four years of development, PODS network is now approaching “realistic” characterization.
Change in recent emphasis of PODS simulations:
Away from O-D method “competitions”
Towards understanding major impacts on RM performance
Ability to simulate larger networks opens up even greater potential for PODS research:
Airline alliances and other competitive strategies
Impacts of pricing and schedule changes on RM methods
Inclusion of scheduling and fleet assignment models

29. PODS Revenue Management Research at MIT
MIT PODS Consortium of 6 international airlines
Major accomplishments in past year:
Expansion of PODS network cities, 2 airlines, multiple banks per day
Establishment of “implementable” O-D methods
Focus on sell-up models and interaction with forecasts
Impacts on RM method performance of forecasting, demand shocks, fare structures, cancellations
New competitive studies involving RM
Alliance RM Strategies
Impacts of New Entrant Airlines