Revenue Management

Revenue management (RM) is the study of how to maximize expected revenue from a perishable, capacity-constrained resource by controlling whether to accept requests (or which price to post) over time. Classic examples are airline seats, hotel rooms, spoilable advertising impressions, and limited edition product runs. The inventory is perishable because unsold units expire after the selling horizon (a flight departs; a night passes). Capacity is fixed and often small relative to demand variability, so correctly withholding capacity from low-paying customers in anticipation of high-paying customers can be valuable.

Formally, consider a finite time horizon t = 0,1,...,T (e.g. booking periods before departure), capacity (inventory) $c_0 \in \mathbb{Z}_{\ge 0}$, and a sequence of stochastic demand arrivals for $K$ fare classes or products. Let $r_k$ be the fare (price) for class $k$. At each arrival (or decision epoch) the manager chooses an action (accept or reject a request, or set a price). The goal is to choose a policy $\pi$ to maximize expected total revenue:

where $R_t$ is realized revenue at epoch $t$, subject to the constraint that the total units sold never exceed initial capacity $c_0$.

In Dynamic Programming (a prerequisite), this becomes a stochastic dynamic program: define the value function $V_t(c)$ as the maximum expected revenue from time $t$ onward with $c$ units remaining. The Bellman recursion for a single arrival in period $t$ with an arriving class $k$ request is:

with boundary $V_{T+1}(c)=0$. Example (numeric): suppose $T=1$ (one decision epoch), capacity $c_0=1$, and a single arrival that is class 1 with probability 1 and fare $r_1=150$. Then $V_1(1)=\max\{0,150\}=150$; the Bellman equation gives a trivial decision to accept. For multi-period, stochastic arrivals the Bellman recursion quickly grows in state and action branching.

Key conceptual tools that RM uses (and that you know from prerequisites):

• •From Price Discrimination: RM operationalizes third-degree segmentation as fare classes and can be seen as dynamic price discrimination across time and willingness-to-pay segments. In Price Discrimination, we learned to separate customers by elasticity; RM enforces this separation dynamically via controls like booking limits and prices.
• •From Dynamic Programming: RM uses backward induction and value functions; the marginal value of one additional unit of capacity is $V_t(c)-V_t(c-1)$ and serves as the opportunity cost (shadow price) for selling now.
• •From Bayesian Decision Theory: accepting/rejecting a request is a one-step decision where you compare immediate reward to the posterior expected opportunity cost of capacity. The decision rule that minimizes posterior expected loss is: accept iff fare $\ge$ expected opportunity cost.

Intuition: the manager should accept a low fare only if the expected future revenue from holding the unit (the option value) is less than the low fare. Littlewood's rule (derived shortly) gives a crisp form of this intuition for two classes; bid-price controls generalize it to many resources and products by pricing capacity via dual variables.

Why perishable + capacity? Because perishable inventory destroys inter-temporal substitution — unsold inventory is forever lost — so the manager faces a stochastic stopping/acceptance problem. In contrast to static price discrimination (one-shot), RM trades off present vs. future sales under capacity scarcity, leveraging Dynamic Programming and posterior expectations (Bayesian Decision Theory) to produce implementable heuristics.

Booking limits and Littlewood's rule are the foundation of two-class revenue management. Consider two fare classes: high ($H$) with price $r_H$, and low ($L$) with price $r_L$ where $r_H > r_L$. Future high-fare demand is stochastic; we want a simple rule to decide whether to accept an arriving low-fare request given remaining capacity $c$ and remaining booking horizon.

Littlewood's rule (classic statement): Maintain a protection level $y$ (number of seats reserved for high-fare passengers). Accept a low-fare booking only if the remaining capacity $c$ strictly exceeds $y$ (i.e., accept if $c > y$). The protection level $y$ is chosen to solve

where $D_H$ is the random number of arriving high-fare passengers in the remainder of the selling horizon. Equivalently, set $y$ as the smallest integer such that $P(D_H\le y) \ge 1 - r_L/r_H$.

Derivation (short, rigorous sketch): Consider the moment a single low-fare request arrives, with remaining capacity $c$ and future high-fare demand random variable $D_H$ (not counting the current low request). Rejecting the low fare leaves capacity $c$ to capture future high fares; accepting reduces capacity to $c-1$. The expected incremental value of keeping one seat for high fares equals the high-fare revenue times the probability that at least one high-fare customer will arrive to use that seat: $r_H \cdot P(D_H > c-1)$. If we accept the low fare, we get $r_L$ now; if we reject we expect to earn $r_H$ with probability $P(D_H > c-1)$. Accept iff

Rearrange and define $y = c-1$; then accept iff $P(D_H > y) \le r_L/r_H$. Choosing $y$ to satisfy the equality gives the canonical protection level. This uses Bayesian decision thinking: the posterior expectation of future reward (here $r_H \cdot P(\text{sale})$) is compared to immediate reward $r_L$ and the action minimizing expected loss (regret) is chosen.

Concrete numeric example: Suppose $r_H=200$, $r_L=100$ so $r_L/r_H=0.5$. Suppose future high-fare demand is Poisson($\lambda=1.2$). Then compute tail probabilities:

• •$P(D_H>0) = 1 - e^{-1.2} = 1 - 0.3010 = 0.6990$ (approx)
• •$P(D_H>1) = 1 - (P(0)+P(1)) = 1 - (e^{-1.2}+1.2 e^{-1.2}) = 1 - (0.3010 + 0.3612) = 0.3378$
• •$P(D_H>2) = 1 - (P(0)+P(1)+P(2)) = 1 - (0.3010 + 0.3612 + 0.2167) = 0.1211$

We need $P(D_H>y) = 0.5$. From above, $P(D_H>0)=0.699>0.5$, $P(D_H>1)=0.338<0.5$. So solution is $y=1$. Interpretation: protect $y=1$ seat for high fare; accept a low-fare booking only when $c>1$ (i.e., when at least 2 seats remain). If $c=2$, accept low; if $c=1$, reject low.

Mechanistic intuition: Littlewood uses only the distribution of future high demand and the two prices – it doesn't require computing the full DP. It is exact for two classes under the assumption that high-fare requests are i.i.d. in remaining horizon and that high fares are first-come-last-served in reservation logic. It is also expressible via a marginal-value comparison: accept iff $r_L \ge$ expected increment from saving the seat, $r_H P(D_H>c-1)$.

Multiple classes and nested protection levels: If there are more than two fare classes ordered $r_1 < r_2 < ... < r_m$, we can sequentially apply Littlewood's rule to compute a nested set of protection levels $y_{m-1} \le y_{m-2} \le \cdots \le y_1$. For example, to decide whether to accept class $k$ we compute the protection level against all higher classes combined (treating their aggregate demand distribution) and accept only if $c>y_k$.

Limitations and when Littlewood is exact: Littlewood is exact under independent, stationary arrivals and for the two-class case. For more general arrival processes (choice-based demand, nonstationary Poisson), it is an approximation; it is often a very good one in practice when demand forecasts are accurate.

Connection to Dynamic Programming and Bayesian Decision Theory: The inequality $r_L \ge r_H P(D_H>c-1)$ is a one-step Bayes-risk minimizing rule: the posterior expected value of rejecting equals $r_H P(D_H>c-1)$; accept if immediate reward exceeds that posterior expectation. From Dynamic Programming, this rule is equivalent to comparing fare $r_L$ with the marginal value of capacity $V_t(c)-V_t(c-1)$ when the value function has particular structure (monotone marginal values) — this is why Littlewood emerges as a closed-form policy.

Real problems rarely have a single resource and two fares. Airline problems involve many itineraries that consume multiple flight legs; hotels combine room-types and dates; advertising sales allocate impression bundles across campaigns. In multi-resource, multi-product settings we use bid-price controls derived from a deterministic linear program (DLP) approximation, whose dual variables act as shadow prices (bid prices) for scarce resources.

Network model: Let there be $I$ resources (e.g., flight legs) with capacities $C_i$ for $i=1,...,I$. There are $J$ products (offers), each with price $r_j$ and a deterministic or expected demand $\bar{D}_j$ over the horizon (the DLP uses expected demand). Product $j$ consumes $a_{ij}$ units of resource $i$ (often 0 or 1). The deterministic LP (DLP) allocates expected sales $x_j$ to maximize expected revenue:

Numeric example: Two legs ($I=2$) with capacities $C=(100,100)$ seats; three products $J=3$: product 1 is nonstop on leg 1 only ($a_{1,1}=1,a_{2,1}=0$) with price $r_1=200$ and expected demand $\bar{D}_1=80$; product 2 is nonstop on leg 2 only ($a_{1,2}=0,a_{2,2}=1$) with $r_2=150$, $\bar{D}_2=90$; product 3 is a connecting itinerary using both legs ($a_{1,3}=1, a_{2,3}=1$) with $r_3=300$, $\bar{D}_3=120$. The DLP is:

maximize $200 x_1 + 150 x_2 + 300 x_3$ s.t.

• •leg 1: $x_1 + x_3 \le 100$
• •leg 2: $x_2 + x_3 \le 100$
• •demand bounds: $0 \le x_1 \le 80$, $0 \le x_2 \le 90$, $0 \le x_3 \le 120$.

Solving by inspection: we would prefer to sell connecting product 3 whenever both legs have capacity because $r_3=300$ is greater than $r_1+r_2=350$? (Note: check arithmetic: $r_1+r_2=200+150=350>300$, so in this example selling two separate nonstops generates more revenue than one connecting ticket — this is a modeling choice.) Suppose connecting revenue is attractive relative to single legs; we find the LP solution using simplex or intuition. For this particular example, assume the LP's optimal corner solution is $x_1=80, x_2=90, x_3=0$ (because nonstops individually have higher combined revenue than connection). The dual variables (shadow prices) associated with leg capacities, call them $\pi_1,\pi_2$, satisfy dual constraints:

Dual (informal): minimize $100\pi_1 + 100\pi_2 + \sum_j \bar{D}_j \mu_j$ subject to the covering constraints; the dual values $\pi_i$ give the marginal value of an extra unit of capacity on leg $i$. In our (assumed) primal solution $x_1=80,x_2=90,x_3=0$, both legs are not at full capacity simultaneously (leg 1 used 80/100, leg 2 used 90/100), so dual prices may be $\pi_1=0, \pi_2=0$ if demand bounds bind; more realistic examples will have binding capacity and positive duals.

How to use duals as a control: At runtime, when a booking for product $j$ arrives, compute its resource cost under duals: $\text{opportunity cost} = \sum_{i} \pi_i a_{ij}$. Accept the booking if

This is exactly the expected-revenue criterion: accept if fare exceeds the expected value of capacity consumed (the posterior expected opportunity cost). Numeric demo: suppose the LP dual yields $\pi_1=60$ and $\pi_2=40$. For product 3 (uses both legs), opportunity cost = $60+40=100$. Compare to $r_3=300$; since $300 \ge 100$, we accept; for product 1 (uses leg 1 only), opportunity cost $=60$, compare to $r_1=200$ accept.

Relation to Littlewood: For a single resource ($I=1$) with two fares, the dual price $\pi$ reduces to $\pi = r_H P(D_H > y)$ interpreted as the marginal expected value of capacity, and Littlewood's inequality $r_L \ge r_H P(D_H>c-1)$ is the same as $r_L \ge \pi$. Thus the bid-price rule generalizes Littlewood to networks by replacing the scalar $r_H P(\cdot)$ with dual prices that aggregate expected marginal values across resources.

Why the DLP? The exact DP for networks suffers from curse of dimensionality (state space is inventory vector in $\mathbb{Z}^I$). The DLP is a fluid (deterministic) approximation: it optimizes expected usage of capacity and ignores stochastic sequencing. Its dual gives economically interpretable shadow prices. Even though the DLP ignores stochasticity, the resulting bid-price controls are often very good in practice, especially when re-solved frequently and combined with stochastic protection updates (EMSR variations).

Refinements and theory:

• •EMSR (Expected Marginal Seat Revenue) is an operational heuristic that computes per-leg protection levels by approximating the distribution of high-fare demand per leg and applying Littlewood in an aggregated way.
• •Re-solve heuristics: periodically re-solve the DLP with updated remaining capacity and updated expected demands to get time-varying bid prices — this asymptotically tracks the optimal DP under mild conditions.
• •Value-function approximations and Approximate Dynamic Programming: approximate $V_t(c)$ by a parametric family (e.g., linear in capacity with coefficients equal to dual prices), and then use policy improvement or rollout to refine.

Connection to Bayesian Decision Theory: The dual price $\pi$ is an estimator of the shadow value of capacity. The decision "accept if $r_j \ge \sum_i \pi_i a_{ij}$" minimizes expected one-step regret when $\pi$ equals the posterior expected marginal value of capacity. Thus bid-price controls are just Bayes-optimal greedy rules under an approximation where the future value function is replaced by its linear estimate derived from the DLP.

Revenue management is pervasive across perishable-inventory industries. I list key applications, extensions, and practical considerations, showing how the theory connects to implementation.

Airline and Transportation: Airlines were the originators of modern RM. Each flight has multiple legs (resources), itineraries (products) that consume subsets of legs, and many fare classes. Practical systems implement nested booking limits and bid-price or EMSR controls derived from DLPs that are frequently re-optimized. Real systems face cancellations, overbooking, no-shows, and fare class granularities with integer seat blocks.

Example numeric application: A low-cost carrier manages a single flight with capacity 150 and a forecast that 60 passengers will buy a refundable flexible ticket at $300, and 200 will buy a nonrefundable budget ticket at$100 with time-varying arrival rates over 30 days. Using Littlewood one can compute protection for flexible fares; using DLP one can allocate expected sales across fare types and set bid prices; the dual price (say $\pi=180$) implies reject any request below $180$ (so accept only flexible or large group budget requests).

Hotels and Perishable Retail: Hotels have per-night per-room perishability. Markdown optimization (dynamic pricing) is typically used near-date where remaining inventory is large relative to remaining demand. A canonical markdown model: continuous price $p_t$, demand $D_t(p_t)=\alpha - \beta p_t$ per remaining period. The DP chooses $p_t$ to maximize expected revenue subject to inventory depletion. While closed-form solutions exist for special cases, in practice heuristic rules like "price so that expected sales per remaining period equal remaining inventory divided by remaining periods" are used, or numerical DP is employed. Concrete numeric example: inventory 10, periods 2, demand $D(p)=6-p$. One computes the optimal $p_2$ and $p_1$ by enumerating feasible sales and their values.

Advertising and Online Platforms: Perishable inventory is impressions for a given time window. Demand is stochastic and often modeled via auctions. RM in this setting involves real-time bid prices for budgets and pacing; dual prices come from a DLP that allocates impressions to campaigns subject to budget constraints.

Choice-based demand and robustification: Real customers choose among offered prices or fare classes; choice models (MNL, nested logit) replace independent arrival models and change optimal controls from booking limits to assortment/price decisions. Bayesian estimation is used to update demand model posteriors; decisions are then Bayes-optimal given current beliefs.

Machine Learning + RM: Modern systems use ML to forecast demand (conditional on context) and feed those forecasts into DLP or DP approximations. Combination is nontrivial because forecasts are inevitably biased; safe policies use robust optimization, constrained re-optimization, or end-to-end policies trained by reinforcement learning.

Limitations and practical adjustments:

• •Overfitting forecast noise: DLP with point forecasts can produce poor duals; stochastic LPs or safety buffers (protection margins) are used.
• •Strategic customers: Customers may delay purchases expecting markdowns; this requires modeling strategic behavior (game-theoretic extensions) or commitment to posted-pricing rules.
• •Cancellation and refunds: Overbooking policies need to be combined with RM.

Theoretical frontier: Network revenue management remains an active research area in applied probability and optimization. Results include asymptotic optimality of static bid-price policies under scaling (fluid and diffusion limits), performance bounds for EMSR heuristics, and improved policies via approximate dynamic programming with provable regret bounds.

Connection to your prerequisites and next steps: The DP backbone is essential to derive value-function-based policies (Dynamic Programming). Bayesian Decision Theory explains decision thresholds as posterior expected value comparisons (accept iff price >= expected shadow price). Price Discrimination provides the microeconomic intuition about segmenting customers by willingness-to-pay. Looking forward, mastering RM enables work in choice-based revenue management, reinforcement-learning-powered dynamic pricing, and robust optimization for uncertain demand.

Practical rule-of-thumb summary:

• •Two classes: use Littlewood — compute protection level via demand tails and price ratio.
• •Many products or network: solve (or re-solve) DLP to get duals — use bid-price accept/reject criterion.
• •Near-departure: prefer markdown/dynamic pricing with demand elasticity models.
• •When in doubt: compute the marginal value of one extra unit of capacity (via DP or dual) and accept if price >= marginal value.

This completes the core mechanics and their connection to practical revenue management.