Revenue Management with OR

Revenue management (RM) — also called **yield management** — is the practice of selling the right product to the right customer at the right time for the right price. It is one of the most commercially valuable applications of [operations research](OperationsResearch), credited with saving American Airlines alone over $1.4 billion in its first three years (1988–1991) and reshaping pricing across airlines, hotels, car rentals, media streaming, ride-hailing, and retail.

At its core, revenue management is a stochastic optimization problem: products with fixed capacity, perishable value, heterogeneous customers, and uncertain arrival patterns. Operations research provides the models that make optimal — or near-optimal — decisions computationally tractable.

The Business Problem

Revenue management addresses industries with:

1. **Fixed, perishable capacity** — an airline seat, hotel room night, or car rental slot that goes unsold has zero revenue potential after the service date; it cannot be stored

2. **Advance sales** — customers book far in advance of consumption

3. **Heterogeneous demand** — customers have different valuations and willingness to pay

4. **Uncertain demand** — how many customers of each type will request booking is not known in advance

5. **Low variable costs** — the marginal cost of serving one more customer is small relative to the price

These conditions mean that price discrimination is both feasible and valuable. A seat reserved six months out for a leisure traveler with low willingness to pay is less valuable than the same seat reserved one week out for a business traveler who will pay a premium for last-minute access.

Single-Leg Seat Inventory Control

The canonical RM problem begins with a single flight leg with C seats and two fare classes: cheap (fare f₁) and expensive (fare f₂ > f₁). Cheap requests arrive first; expensive requests arrive later and closer to departure.

Littlewood's Rule (1972)

Frank Littlewood at British Airways derived the first RM result: protect seats for late-arriving high-fare customers until the expected revenue from one more protected seat equals the certain revenue from selling it cheap.

Let D₂ be the (random) demand for high-fare seats. Protect y seats for high-fare customers. The optimal protection level y* satisfies:

```

f₁ = f₂ × P(D₂ > y*)

```

Interpretation: protect more high-fare seats until the probability of having a high-fare customer who wants that seat equals the ratio of the cheap fare to the expensive fare.

**Example:** f₁ = $200, f₂ = $600, D₂ ~ Normal(40, 10²).

y* satisfies P(D₂ > y*) = 200/600 = 1/3.

1/3 probability corresponds to about the 67th percentile of D₂ = 40 + 10 × 0.44 ≈ 44.

Protect 44 seats for high-fare customers; sell cheap until 44 remain.

Expected Marginal Seat Revenue (EMSR)

Littlewood's rule extends to multiple fare classes through the **EMSR** family of heuristics (Belobaba, 1987). For k fare classes with f₁ < f₂ < ... < fₖ:

**EMSRb:** Aggregate demand for classes j+1, ..., k into a single composite class with blended fare and combined demand distribution, then apply Littlewood's rule. More accurate than EMSRa in practice.

EMSR heuristics remain widely used in airline revenue management despite their approximation nature — they are fast, simple to implement, and perform within a few percent of optimal on typical distributions.

Network Revenue Management

Real airlines sell itineraries (multi-leg journeys), not single legs. A passenger flying New York → Chicago → Denver consumes capacity on two legs. Accepting this passenger forecloses selling those seats on each leg to potentially higher-value passengers on other itineraries.

**The network problem:** Given a network of flights, set booking limits for each (itinerary, fare class) pair to maximize expected total revenue, subject to capacity constraints on all legs.

LP Deterministic Approximation

The simplest network RM model solves a deterministic LP:

```

Maximize: sum_j f_j * x_j

Subject to: sum_j a_ij * x_j <= C_i for all legs i

0 <= x_j <= E(D_j) for all itineraries j

```

Where a_ij = 1 if itinerary j uses leg i (0 otherwise), C_i is leg capacity, and E(D_j) is expected demand for itinerary j. The LP solution provides **bid prices** (dual variables) for each leg — the shadow price of capacity on that leg.

**Bid price control:** Accept a booking request for itinerary j if its fare fⱼ exceeds the sum of bid prices for all legs it uses.

Bid prices are recomputed periodically as demand is observed and forecasts update.

Stochastic Programming Approaches

The deterministic LP approximation ignores demand uncertainty. Stochastic alternatives include:

- **Re-solving:** Update demand forecasts and resolve the LP as bookings arrive

- **Randomized linear program:** Run the LP on multiple demand scenarios; combine booking limits

- **Approximate dynamic programming:** State = (remaining capacities, time), action = accept/reject; compute approximately optimal value function

The tradeoff is computation time vs. solution quality. Airlines re-solve LP models every few hours; real-time ADP would require sub-second solve times for large networks.

Dynamic Pricing

Revenue management moves beyond booking limits to **price-setting**: adjust the price continuously based on current demand, remaining capacity, and time to service date.

The Basic Dynamic Pricing Model

Let the system be in state (n, t): n units of capacity remaining, t periods until the deadline. Customers arrive according to a Poisson process with rate λ. A customer at price p accepts the offer with probability d(p) (the demand function — typically decreasing in p).

The optimal pricing policy satisfies the **Hamilton-Jacobi-Bellman (HJB) equation** (in continuous time):

```

V'(n,t) = λ × max_p { d(p) × (p + V(n-1,t) - V(n,t)) }

```

Where V(n,t) is the expected revenue from optimal pricing with n units remaining and t time until deadline. The optimal price at each state balances immediate revenue against the option value of preserving capacity for later customers.

**Key properties of the optimal policy:**

- Prices decrease as deadline approaches (when capacity is likely to go unsold)

- Prices increase as capacity becomes scarce (high demand relative to remaining supply)

- The structure is called a **threshold policy** in the discrete version

Internet Pricing Applications

Airline, hotel, and car rental dynamic pricing are now fully automated. Prices on a flight can change thousands of times in the weeks before departure, responding to booking pace, competitor prices, and demand signals.

Ride-hailing extends this to real-time surge pricing: Uber and Lyft use dynamic pricing to balance driver supply and rider demand in real time, with prices adjusting every few minutes.

Overbooking

A consequence of the perishable capacity problem is **no-shows** and cancellations: customers who book but don't appear. For airlines, this historically left seats empty on heavily demanded flights.

**Overbooking:** Accept more bookings than physical capacity, relying on a no-show forecast to avoid flying over-full.

The Overbooking Optimization

Let C be physical capacity and x the number of bookings accepted (x > C). Let N ~ F(·) be the random number of passengers who actually show up. Passengers denied boarding receive compensation value v.

The optimal overbooking level x* balances the revenue from selling an extra seat against the expected compensation cost of a denied boarding. The result is structurally identical to Littlewood's rule and the newsvendor critical ratio:

```

P(show-ups >= capacity | overbook by k) = revenue per seat / (revenue per seat + denied boarding cost)

```

**In practice:** Airlines segment no-show probabilities by fare class (cheap tickets cancel more), day-of-week, season, and departure time. Dynamic overbooking models update levels as the departure date approaches and cancellation patterns emerge.

Revenue Management Beyond Airlines

Hotels

Hotel RM shares the structure but differs in key ways:

- **Length of stay:** A hotel room for 3 nights uses capacity on multiple nights (analogous to network RM)

- **Room type differentiation:** King vs. double, view vs. no view — multiple products sharing capacity

- **Group bookings:** Large blocks reserved far in advance at discounted rates

Hotel systems track occupancy forecasts by night, set rack rates and discount tier availability, and control group bookings as network LP problems.

Streaming and Media

Streaming services use RM principles for advertising inventory:

- Ad slots are perishable (unsold inventory = zero revenue)

- Advertisers value different audiences differently

- Programmatic advertising uses real-time bidding — effectively a continuous dynamic pricing auction

Retail Markdown Optimization

Fashion retailers commit to inventory before knowing demand. Unsold items must be marked down. Markdown optimization is a sequential pricing problem solved as an MDP: state = (inventory remaining, time in season), actions = (keep price, mark down to level x).

The Value of Revenue Management

| Industry | RM Application | Documented Impact |

|----------|---------------|-------------------|

| Airlines | Seat inventory control, dynamic pricing | American Airlines: $1.4B in 3 years (1988-91) |

| Hotels | Room type & duration optimization | 2-8% revenue lift typical |

| Car rental | Fleet pricing and availability | Significant fleet utilization improvement |

| Cruise lines | Cabin pricing by deck/type/duration | Material yield improvement |

| Broadcasting | Ad inventory pricing | Improved sellthrough and CPM |