good things take time – building up an o&d-forecaster stefan poelt presentation to the agifors...
TRANSCRIPT
Good Things Take Time –
Building up an O&D-Forecaster
Stefan Poelt
Presentation to the AGIFORS RES & YM Study Group
Berlin, 16 - 19 April 2002
O&D Forecaster • page 2
Agenda
What’s the scope of our O&D project ?
Some traps to watch
O&D Forecaster system overview
What’s the current status ?
How does a perfect project look like ?
Happy end !
O&D Forecaster • page 3
How does a perfect project look like ?
A perfect project
Is in time
Is in budget
Covers the scope
Delivers quality
Our project
Is not quite in time
Is in budget
Scope has been adjusted
Delivers quality
O&D Forecaster • page 4
What’s the scope of our O&D project ?
O&D Forecasting
Detailed level
Individual pax / groups
Gross bookings / cancellations
Aggregation of ‘exotic’ O&Ds
Implemented by Lufthansa Systems Berlin
PNR-basedNo-show
Forecasting
PNR attributes
Decision tree learning
Mixed with traditional no-show forecasts
No decision about vendor
O&D Optimization
Benchmark study
Project stopped after 11 Sep 2001
Re-activation in 2003
Implemented by CRSs
Married Segment Control
Prevents cheating
Implemented at Inventory Dept.
Coordination task
O&D Forecaster • page 5
Simple vs. complex forecast method
Separation of individual and group bookings
Separation of gross bookings and cancellations
Flexible aggregation of low demand traffic flows
Complex methods are more difficult to calibrate
Complex methods can better model reality
Trade-off between accuracy and stability
Pilot‘sWeather Forecast Station
Stone is wetStone is dryShadow at groundWhite on topStone can’t be seenStone swingsStone is hoppingStone disappeared
RainNo rainSunnySnowFogStormyEarthquakeBlizzard
O&D Forecaster • page 6
Better no-show forecasts allow better overbooking
Traditional no-show forecasts are based on booking class and routing
PNR is a richer data source (e.g. ticketing info, special service request etc.)
‘There’s gold in the PNR data mine’ (Louis Busuttil, PROS Conference 1997)
‘PNR-based no-show forecast’ (Klaus Weber, AGIFORS Conference 2000)
PNR-based methods can be applied to booked passengers only - they must be mixed with traditional no-show forecasting
PNR-based methods (e.g. decision trees) can also be applied to cancellation forecasting
O&D Forecaster • page 7
O&D forecasting without O&D optimization
Intermediate step: O&D forecasts aggregated to leg-level and used in leg-optimization
Better quality of leg demand forecasts
Better quality of (PNR-based) no-show forecasts lower overbooking risk
O&D information, better understanding of travel flows
Better communication and co-operation with Pricing and Scheduling / Flight Planning
O&D Forecaster • page 8
Reasons for delay
Too optimistic project plan
Internal staffing (6-7 project members, 4 users – part-time)
Under-estimation of noise in PNR data
Frequent change of project management
Scientific driven project – late involvement of users
Research and calibration effort for new topics (e.g. cancellations)
Re-organizations
11. September 2001
O&D Forecaster • page 9
Renaming the project does not help
There were plans to rename the project to ‘O&D-Forecaster’
This might help to confuse controllers but does not help to solve the problems
Project name: ‚Bid Price 2‘
Original scope was adjusted, O&D-optimization was stopped after 11. September 2001
O&D Forecaster • page 10
Shit happens
There are many situations where something can go wrong
The business processes haven’t been addressed at the beginning
Later on we realized that important functionalities are missing
Noise and complexity of PNR data have been under-estimated
Aggregation of no-show information from O&D to leg is problematic
O&D Forecaster • page 11
A precise specification avoidscomplaints later-on
Innovative IT-projects require a lot of research
Alternation of research and prototyping / implementation
We decided to refine the specification during the project
Problems keeping specification and documentation up to date
Testing with ‘fuzzy’ specification is challenging
You‘ve builtit up reversed !
O&D Forecaster • page 12
A good documentation is important
Documentation often is prioritized down
Good documentation improves user’s confidence
3 documents:
– system administration
– user manual / online help
– process / workflow
PRESSANY KEY
O&D Forecaster • page 13
In a perfect project: Project managers have the right perspective
Focusing on the most important topics
Understanding the subject (RM and Forecasting)
Good communication and motivation skills
Currently 2 project managers, one focusing on management, the other focusing on methods - learning from each other
O&D Forecaster • page 14
In a perfect project: Motivation of the project team is not an issue
Working for years on the same thing is boring
We are the 4. generation of project managers
The project survived 2 re-organizations
Increase motivation by changing responsibilities
Success is the best motivator !
O&D Forecaster • page 15
Our new toy
Availabili-ties
FlightSchedule
Model Mixer
Leg Optimizer
Forecaster Pool
UserInfluence
ValidatorSegment
Aggregation
ExponentialSmoothing/Brown
InductionTrees
ODI Forecast DB
Bad Data Processor
Data Loading
PNR-Data
Trip-Builder
Forecast Building
Forecast Kernel
ForecastInterface DB
Shadow
History BuilderUnconstrainingOutlier Detection
Flight Monitor
PNR-Interpreter
History Pool
Concept
Attractivities
ODO-Builder
AV-Calculator
ActualData
Event &SeasonalityManager
ForecastController
O&D Forecaster • page 16
Aggregation solves the small numbers problem
There are many low traffic O&Ds
20:90 rule - 20% of all O&Ds cover 90% of the traffic
Flexible aggregation of low traffic O&Ds is a key factor
O&D Forecaster • page 17
Effects of schedule changes are handled automatically
A major schedule change can affect thousands of O&Ds
Concept of Lufthansa Systems Berlin handles new, moved and cancelled flights
Demand is automatically adjusted based on ‘neighbored’ flights/paths
‘Distance’ to neighbored flights mainly depends on departure and arrival times
Approximation of market (passenger choice) model
new flight
neighbored flights
flight A flight B
flight C
dist(A,C) dist(B,C)
O&D Forecaster • page 18
This year we will take the last hurdle
Forecast quality (aggregated to leg-level) is better than for current system
Data loading performance is o.k.
Business processes have been re-designed and matched with the tool
Flight by flight migration of new forecast system (Jun - Nov 2002) minimizes risk
Business re-organization will go hand in hand with migration
Parallel project of O&D-reporting is started
Married Segments Control is implemented in Amadeus and Galileo
O&D Forecaster • page 19
The forecast results are very promising
Validation of O&D-forecaster at representative test flights
Distinction of open and closed flights to separate unconstraining effects
Open flights are less important but have reliable actual values
Open Flights
Err
or
Current system O&D-Forecaster
Closed Flights
Err
or
Current system O&D-Forecaster
Booking period departure Booking period departure
O&D Forecaster • page 20
Happy end !
The O&D forecaster will be completely in operation this year
Very good cooperation with Lufthansa Systems Berlin
O&D not always means Odd & Delayed !