Revenue Management Flashcards
RM Objectives
Maximize Revenue.
Right Product to right customer at right time and right price.
RM Conditions
1 perishable product
2 low marginal cost
3 limited capacity
4 market segmentation
Classical Segmentation
Tourists
-price sensitive
book early
-low service demands
Business
- not price sensitive
- book late
- high service demands
Stages of RM process
1 Forecast
2 Optimization
3 Reservations
Why is RM getting more complex?
networked customers
alliances and competition
automatization
Traditional RM Model
and modern extensions
Independent Demand
Conditions for Price Differentiation
Multiple demand functions
supplies maximizes revenue
market not perfectly transparent or competetive
Degrees of price differentiation
1st: perfect differentiation 2nd: by self selection (find highest prices maintaining fencing) 3rd: by customer characteristics
Capacity Oriented Pricing
Price per unit dependent on number of units bought.
Segment Oriented Pricing
geographical
temporal
personal
product-oriented
Auctions
-price determined through bids
-increase with number of customers and valuation
RM potential:
near perfect 1st degree
no need for estimation
Name Your Own Price
Based on Auctions
Bids succesful if exceeding bid price
can be combined with opaque selling
Booking Limits
Maximum Number of units to be sold per class.
- partitioned
- nested
Protection Limits
Number of units to be reserved per class
- partitioned
- nested (C includes A)
Theft Nesting
Bookings in valuable classes reduce limit of other classes.
Leg Based RM with independent demand
Littlewood’s Rule and EMSR-b
Littlewood’s Rule
two classes
2 arrives before 1
p2 < p1
sell in 2 as long as p2>p1*P(D>x)
derives protection limits.
EMSR-b
1 aggregate demand for class j and more expensive 2 calculate avg price for class j and more expensive 3 apply littlewood's rule
Bid Price
Minimum Revenue to earn with next booking on resource.
=1 -> all classes are available
=inf -> no capacity left
opportunity cost of seeling an additional unit of capacity.
Bid Prices
- well-suited for?
- control what?
- dynamic?
- customers?
- > Products that require multiple resources
- > availability of tariffs rather than booking classes
- > re-calculated, updated frequently
- > don’t see bid prices, nor full set of available classes
Dynamic Demand Arrivals
possibility of product request arriving at particular time
Stochastic Dynamic Programming Model
- sales horizon partitioned (most one request)
- independent demand
- product i requires a(hi) resources of h
- state described by original capacity C and left-over capacity c(h) per resource h.
Demand Estimation
needs to know?
- how many potential customers
- in which demand segment
- which choices?
- -in the past
Demand Forecast asks?
How much demand in the future?
Faulty demand forecasts lead to?
Spoilage: selling too little, too late
Spill: Selling too much, too early
Forecast Data
Sales Transaction Data Controls Data (bid prices) Auxiliary Data (tax, demographic, weather)
Partial Data
If sales horizon and observations overlap.
Poisson Intensity indicates…
probability of a customer arriving and requesting class i
Forecasting Methods
Structural: descriptive, HW, function
Time-Series: ARIMA model!
Time Series Methods
1 Hypothesis about process generating data
2 estimate process parameters
3 apply best method for model
-> can exploit correlations in data
Pick Up Forecasting
Used for overlapping observations of subsequent offerings.
Additive: incremental gain per time-slice
Multiplicative:
relative increase per time slice (assumes correlation of future bookings with current bookings)
Independent Demand Model
1 customer only requests products defined for his segment
2 customers from different segments do not request same product
3 if product is sold out, segment stopy buying
4 pricing of not demanded products irrelevant
How to make demand (more) independent?
Fencing
- advance purchase restrictions
- minimum or weekend stay restrictions
- cancellation fees
Danger of Assuming Independent Demand?=
Spiral Down, Buy down
Dependent Demand Model
1 customers buy preferentially in their segment
2 different segments can request same product
3 if sold out, customers may be other product
4 pricing of one product may affect product choices of other segments
Relevant Questions for Dependent Demand?
1 Products targeting same customers?
2 Qualitative differences between products?
3 How much demand per product?
4 How much overlap?
FRAT 5
new fare ratio retains 50% of demand
Random Utility Models account for
1 not observable relevant variables
2 inconsistent customer choice
MNL assumptions
utility= representative component and random component
random component is Gumbel distributed
Indepdence from irrelevant alternatives
MNL scaling parameter
0 -> deterministic
inf -> entirely random
IAA example
in MNL the number of products influences choice even if products dont differ on relevant attributes.
Preference Lists
- how modeled?
- what kind of model?
directed acyclic graph
non-parametric
Estimation vs Forecasting
descriptive vs predictive
past vs future
Non-parametric vs parametric Estimation
no particular distribution or model VS assumes underlying distributin or model
shape vs parameters
Estimation Objectives
unbiased - expected value of estimator equals actual
efficient - unbiased and lowest possible variance
consistent- converges to true value as sample size increases
Estimation Challenges
Endogeneity
Constrained Observations
Biased Observations
Small Sample
Regression Estimators are…
causal predictions.
+ can explain abrupt changes
+offers behavioral explanation
-hard to define all relevant variables and relationships
Probit Model
probability of one event
Multionomial logit/probit
set of discrete probabilities
Regression Estimator vs ML Estimator
Regression (OLS):
minimize squared error resulting from estimated paramters across all observations
MLE:
Maximize probability of observing the prediction given by estimated paramters multiplied across all observations
Demand Segments.
What to estimate?
How many?
What share?
What behaviour?
Behaviour of a demand segment.
Classical?
Modern?
classical:
one segment per booking class
modern:
random choice
different utility models
non-parametric (preference lists)
Methods to unconstrain demand?
Naive Projection Detruncation -estimate mean and sd -replace closed with heuristial solution -test for convergence Estimation Maximisation -estimate mean and sd -test for convergence -replace xconst by xk via ML estimation
Bid Prices under dependent demand. Problems:
and Solution.
demand may be dependent on offers with buy-dwon between classes (customers dont spend as much as they might have)
bid prices may decrease over time - cheaper classes become available
->Solution: Adjusted Fares
Bid Prices adjusted to account for buy -down
Nested Capacity Controls or Bid Prices that increase monotonously, RM ensures increasing prices over sales horizon.
Why would this not hold?
Sales Interventions
Forecast Updates
Competitor Matching
Dependent Demand Problem: Strategic Customers (time-lag-spiral-down)
Customers may delay purchase, expecting price dip in the future (as they have learned from past observations)
->updating forecasts can induce a time-laag-spiral-down effect
Dependent Demand Problem:
Parallel Flights
Customers switching booking class use same resource. Intractable to optimize dependencies in choice modelling.
Dependent Demand Problem:
Competition
Automated RM system would have to forecast demand AND competitor otffers -hardly feasible
Prisoner’s Dilemma:
underbidding is dominant
leads to total buy-down
->need thresholds
Capacity Controls.
1 single resource and independent demand?
2 multiple resources and independent demand
3 dependent demand
1 EMSR-b
2 Bellman Function and Bid Prices
3 Bellman Function and adjusted Fares
Considered Trade-Off when Overbooking?
Service Degree vs Incurred Cost
Service Degree Type 1
Probability that given overbooking limit and capacity all customers can be serviced as promised
Service Degree Type 2
Percentage of Customers that did not receive service as promised
Overbooking gone wrong.
Components of Costs?
1 actual monetary compensation
2 reputation demage
->often increased as exponentially increasing
Cancellation vs No Shows
no shows forego service without announcing in advance.
Static Overbooking Models
1 Determine overbooking limit
2 use result for capacity allocation as if it was real
3 update overbooking limit and allocations as actual bookings and cancellations come in
Dynamic Overbooking Model…does what?
updates constantly
Conditions for Dynamic Pricing
1 prices can be adjusted quickly without high transactional costs
2 customers and policymakers do not expect fixed prices
3 models remain solvable
Dangerous simplyfing assumptions:
myopic customers
time-independent demand
monopoly
Deterministic Dynamic Pricing
KKT
1 MR constant across periods
2 either MR=0 or C sold out
3 if demand >C; MR>0
3 Types of Uncertainty
1 Ignorance
2 Incertitude (probs unknown)
3 Risk
RM Indicators
Revenue Yield (revenue per passenger) Bookings Capacity Utilization Forecasted Demand Inventory Controls Interventions
MAPE
1/n * SUM [abs(FCi-OBSi)/OBSi]
RMSE
root(1/n*SUM[(FCi-OBSi)^2])
Challenges to Evaluating RM Performance
1 dynamic economy 2 evaluating effect of competition 3 what went wrong and where 4 how good is good 5 identifying comparable markets
Revenue Opportunity Model
Problematic: assumes demand prediction was absolutely accurate
Air Cargo Paper
- two dimensional problem: weight and volume
- uncertainty makes overbooking neccessary
- itineraries only defined by start and end
- decomposition heuristic performs best
Attended Home Delivery Paper
Trade-Off:
high service degree vs efficient logistical operations
->dynamic management
offer time slot if generated profit > opportunity costs
using generalized attraction model to capture dissatisfaction if preferred slot not offered
Risk Averse RM Paper
Why?
low number of repetitions
critical impact
management constraints
Bid Price with Factor performed best
Non-Traditional Settings
Classical Practices not entirely transferable
inventory sensitive vs price sensitive demand
golf courses with many early bookings and conversion management performed best
Trade In or Upgrade Paper
Result: cash payout and refund should reversely depend on inventory level of refurbished product
low: both trade in options offers
medium: both with medium payout
high: no trade in for cash, upgrade only
Competition in Airlines Paper
vertical competition
multiple airlines, one leg, one company
horizontal competition
multiple companies per leg
Dangerous simplifying assumptions for dynamic pricing
- myopic demand
- time independent demand
- monopoly
Integrating Overbooking and Capacity Controls
Iterative:
1 calculate ob-limit 2 allocate capacity
Integrative
static
1 calculate current ob-limit 2 allocate capacity during time slice
dynamic
1 calculate service degree 2 decide request acceptance
Assumption of exact OB models:
cancellation and no shows probabilities:
- same for all customers
- iid among customers
- independent of booking time
Integrating Overbooking and Capacity Controls MEANS…
balancing the value of accepting requests with the potential costs of denied boardings.
Advantages of Price-based RM
- better demand curve fitting
- more flexible
- more intuitive
Advantages of Capacity-based RM
- well established-algorithms
- probably better fit with existing technology
- fewer simplifying assumptions
