revmng

Revenue management for Python: the building blocks for selling fixed, perishable capacity well. Single-leg seat protection, overbooking, pricing, network bid prices and the standard performance metrics, computed from the textbook methods and validated against published worked examples.

The core has no third-party dependencies. Every result carries provenance (library version, input hash, timestamp) and a JSON-safe to_dict().

pip install revmng

Quick start

import revmng

# Two-class seat protection (Littlewood's rule)
revmng.littlewood(fare_high=1000, fare_low=400, mean_high=50, sd_high=18)
# -> protect this many seats for the high fare; sell the low fare below it

# Multi-class nested booking limits (EMSR-b, the industry-standard heuristic)
a = revmng.emsr_b(
    [(1000, 30, 12), (700, 40, 15), (400, 60, 20)],   # (fare, demand mean, sd)
    capacity=120,
)
a.booking_limits_int          # nested whole-seat booking limits
print(a.summary())

# Overbooking: balance the cost of bumping against the cost of empty seats
o = revmng.overbooking_limit(100, no_show_rate=0.12,
                             denied_cost=400, spoilage_cost=180)
o.authorization_limit         # how many reservations to accept

# Pricing: the newsvendor stocking quantity and the profit-maximising price
revmng.newsvendor(price=10, cost=4, demand_mean=100, demand_sd=30).optimal_quantity
revmng.optimal_price(revmng.LinearDemand(intercept=100, slope=2), unit_cost=10)

# Score booking limits against perfect-hindsight and no-control benchmarks
revmng.revenue_opportunity([1000, 800, 600, 200], [25, 30, 20, 50], capacity=100,
                           booking_limits=[100, 70, 45, 32]).rom

# Exact optimal protection levels (and the expected revenue EMSR leaves behind)
opt = revmng.optimal_protection_levels(
    [(1000, 30, 12), (700, 40, 15), (400, 60, 20)], capacity=120)
opt.expected_revenue

# Should we take this group? (displacement analysis)
cost = revmng.group_displacement(group_size=30, capacity=100,
                                 demand=[70, 95, 85], value_per_unit=70)
revmng.evaluate_group(group_rate=60, units=90, displacement_cost=cost,
                      variable_cost=50).accept

# Should we accept this multi-night stay? (length-of-stay bid-price control)
revmng.evaluate_stay({"Mon": 80, "Tue": 120, "Wed": 90},
                     ["Mon", "Tue", "Wed"], total_rate=320).accept

Network bid prices need SciPy and the charts need matplotlib:

pip install "revmng[network]"   # bid_prices via the deterministic LP
pip install "revmng[plot]"      # booking-limit, overbooking and ROM charts

from revmng import Product
n = revmng.bid_prices(
    [Product(fare=200, uses={"AB": 1}, demand=60),
     Product(fare=300, uses={"AB": 1, "BC": 1}, demand=40)],
    capacities={"AB": 100, "BC": 100},
)
n.bid_prices                       # marginal value of each resource
n.accept(fare=250, uses={"AB": 1, "BC": 1})

What it covers

optimal_protection_levels computes the exact optimal nested protection levels by dynamic programming (the policy the EMSR heuristics approximate) and reports the optimal expected revenue; policy_revenue scores any policy against it. Group and length-of-stay decisions share one idea: accept a request when its contribution covers the opportunity cost of the capacity it displaces (a marginal seat value or a sum of nightly bid prices).

Demand for the capacity-control methods is assumed normal and independent, with the classic low-before-high booking assumption. EMSR-a and EMSR-b are heuristics; for two classes both reduce to Littlewood's optimal rule.

What is out of scope

revmng computes decisions from the demand you supply. It does not forecast demand, estimate willingness-to-pay, or unconstrain censored bookings, and it does not connect to any reservation, property-management or distribution system. Those belong upstream; keeping them out is what lets every result here be reproduced and audited.

Validation

The methods are checked against published and hand-derived worked examples (see tests/validation_cases.json):

A clean multi-class EMSR-b worked example with both inputs and outputs is hard to find in the open literature, so the heuristics are also cross-checked two independent ways (tests/test_crosscheck.py): their protection levels match a separate SciPy implementation of the same formulas to 1e-9, and over many random instances EMSR-b never exceeds revmng's exact dynamic program and stays within about 0.5% of it, matching Belobaba's reported near-optimality. | Group displacement | standard displacement analysis, arithmetic shown | 40 displaced room-nights, cost 2800, break-even rate 81.11 | | Length-of-stay bid price | Talluri and van Ryzin (2004), arithmetic shown | 290 hurdle, accept at rate 320 | | Newsvendor quantity | critical-fractile rule | reproduced | | Overbooking (service level) | deterministic no-show formula | reproduced |

Definitions and methods

• K. Littlewood, "Forecasting and control of passenger bookings" (1972).
• P. Belobaba, "Application of a probabilistic decision model to airline seat inventory control," Operations Research (1989). EMSR-a and EMSR-b.
• K. Talluri and G. van Ryzin, The Theory and Practice of Revenue Management (2004).
• R. Phillips, Pricing and Revenue Optimization (2005).

License

MIT. Written and maintained by Atakan Arikan, MSc Student at Tsinghua University and Politecnico di Milano.