Why is mathematics important in airline overbooking? A flight has a fixed number of seats, yet some ticketed passengers do not arrive. Selling exactly the seat count can leave empty capacity; selling more can create a shortage when too many people show up. Probability, forecasting, expected value and risk limits make the trade-off visible.
This article uses fictional flights, prices and costs to teach the mechanism. It does not advise an airline or passenger, and it does not describe every jurisdiction’s law. Actual carriers must follow current aviation, consumer, safety and non-discrimination rules. A mathematically profitable policy can still be unfair, unlawful or operationally unwise.
Choose the capacity question you want to solve
- Why can ticket count exceed seat count?
- How does a show-up distribution work?
- Where does expected value enter?
- Why do tail risks and passenger rights matter?
- How can a student model the problem?
- What should families know?
A seat is perishable capacity
Once an aircraft departs, an empty seat on that flight cannot be stored and sold tomorrow. Economists call this perishable capacity.
The same structure appears in hotel rooms, appointments, workshops and events. Capacity is fixed for a time slot, demand is uncertain and unused space loses its immediate value.
Air travel adds safety, connections, baggage, crew, consumer rights and high disruption costs. The school model isolates one capacity mechanism without pretending that every service can use the same rule.
No-show history becomes a probability estimate
Suppose 4,750 of 5,000 booked passengers on a route arrived ready to board. The observed show-up rate is 4750/5000 = 95%.
That sample proportion estimates a future probability only if conditions are sufficiently comparable. Season, departure time, fare flexibility, weather, group travel and check-in policy can change behaviour.
Reporting the numerator, denominator and time window is better than saying “about five percent do not show” without evidence.
A rate has sampling uncertainty
Even with a stable true probability, a sample proportion varies. A rough standard error for proportion p-hat from n independent bookings is sqrt(p-hat(1−p-hat)/n).
For p-hat=0.95 and n=5,000, standard error is about sqrt(0.0475/5000)=0.00308, or 0.308 percentage points.
Independence and stable conditions may not hold, so the formula is a starting point. Model uncertainty can exceed sampling uncertainty.
The binomial model is a first approximation
If N booked passengers each independently show with the same probability p, the number who show, X, follows a binomial distribution: X~Binomial(N,p).
Its mean is Np and variance is Np(1−p). With N=105 and p=0.95, expected show-ups are 99.75 and standard deviation is sqrt(105×0.95×0.05)≈2.23.
The mean being below 100 does not guarantee everyone fits on a 100-seat flight. The distribution has a right tail above capacity.
Expected attendance is not a capacity guarantee
For 105 bookings at 95% show probability, the expected number is about 99.75. Some flights have 97 arrivals, others 102 or 105.
Capacity planning needs P(X>100), not only E[X]. Averages describe long-run centre; shortages live in the tail.
This is a common decision error: setting capacity equal to expected demand while ignoring variability creates frequent overflow.
Probability of a shortage is a tail sum
For capacity C and bookings N, shortage probability is P(X>C). Under a binomial model, add probabilities from C+1 through N.
For N=105, C=100 and p=0.95, compute sum_{x=101}^{105} binom(105,x)0.95^x0.05^(105−x). Software can evaluate the exact sum.
The probability answers whether any shortage occurs, not how many passengers are affected. Expected overflow is a different quantity.
Expected overflow counts excess passengers
Overflow is (X−C)^+, meaning max(X−C,0). Expected overflow is sum max(x−C,0)P(X=x).
Two policies can have the same shortage probability but different expected overflow. One may usually exceed capacity by one; another may rarely exceed it by five.
Both matter operationally. A binary “oversold or not” measure hides severity.
Empty seats are the opposite tail cost
Unused capacity is (C−X)^+ when show-ups fall below seats. Expected empty seats are the probability-weighted average of that quantity.
Increasing bookings generally reduces expected empties and increases expected overflow. The policy moves risk from one side to the other.
There is no free setting that eliminates both when attendance is uncertain. Mathematics clarifies the trade-off so it can be governed.
Worked example with a small aircraft
Consider a fictional flight with 10 seats, 11 bookings and independent show probability 0.9. Expected show-ups are 9.9.
All 11 show with probability 0.9^11≈0.3138, the only way to exceed capacity in this tiny example. Expected overflow is therefore about 0.314 passengers per flight.
That expectation means roughly 314 excess passengers per 1,000 similar flights under the model, not a fraction of a person on each flight.
With exact capacity, empty seats remain likely
If the same flight sells only 10 bookings, everyone fits, but expected show-ups are 9. The expected number of empty seats is one.
Probability all ten show is 0.9^10≈0.3487. Thus roughly 65.1% of flights have at least one empty seat under the independent model.
This comparison shows why operators consider controlled overbooking. It does not establish that any particular policy is acceptable.
Correlation breaks the independent model
Passengers do not always decide independently. A family of four may all arrive or all miss a connection. Weather or a transport disruption can affect many people together.
Positive correlation increases variability relative to a binomial model with the same individual rate. More flights may have very high or very low show-ups.
Using a binomial tail while ignoring group dependence can underestimate shortages. Booking-party and itinerary structure deserve modelling.
Different passengers have different probabilities
A flexible business ticket, a group booking and a connecting passenger may have different show-up probabilities. The equal-p binomial assumption compresses those differences.
With independent but unequal probabilities, the total follows a Poisson-binomial distribution. Its mean is the sum of individual probabilities and variance is the sum of p_i(1−p_i).
Segmentation can improve forecasts, but it also creates fairness and privacy concerns. Sensitive attributes should not be used casually.
Group bookings change the unit of uncertainty
If four passengers travel together, modelling four independent Bernoulli trials overstates the number of mixed outcomes where only some show.
A simple alternative treats the party as one group event with size four, though individual changes can still occur. Simulation can represent group-level and person-level probabilities.
The correct unit depends on how bookings behave. Counting records is not the same as modelling decisions.
Connections create conditional arrivals
A passenger starting at the airport may have a high show probability, while a connecting passenger depends on an inbound flight arriving in time.
If inbound on-time probability is 0.88 and a passenger boards given on-time arrival with probability 0.98, simplified show probability is 0.88×0.98≈0.8624, before other causes.
These events may share weather and network disruptions. Multiplying probabilities requires a conditional structure, not blind independence.
Time-to-departure updates the forecast
Check-in, app activity and inbound status provide new evidence as departure approaches. A booking-level probability can be updated.
At 24 hours, one forecast may expect 98 arrivals; at 45 minutes, checked-in counts could make the distribution much tighter. Operational decisions should use the latest valid information.
Updating must not become covert or discriminatory surveillance. Evidence should be relevant, governed and protected.
Expected profit balances empty seats and shortages
Let an additional booking produce expected fare contribution R if it uses an otherwise empty seat. Let expected shortage cost be B when it displaces someone.
A simplified marginal rule accepts the booking when R×P(seat would be empty) > B×P(booking creates overflow), adjusted for other effects.
The probabilities depend on current bookings. As N rises, another booking becomes less likely to fill emptiness and more likely to create overflow.
Costs include more than immediate compensation
Displacement can create hotel, meal, re-routing, staff and legal costs. It can also damage trust, disrupt connections and impose large personal costs on passengers.
Some costs are difficult to price. Missing a wedding or medical appointment cannot be captured by a generic voucher amount.
An optimisation that counts only the carrier’s immediate cash cost is incomplete. The objective boundary is an ethical and business choice.
Expected value is not a complete policy
Expected profit averages many flights. It can hide rare days with large disruption or concentrated harm to a few passengers.
Risk limits might cap shortage probability, expected overflow or a high quantile such as the 99th percentile. Service rules can require protection for connections or special needs.
An expected-value winner can be rejected because tail risk, fairness or law makes it unacceptable.
Quantiles answer a service-level question
The 95th percentile of show-ups is a number q such that the model assigns at least 95% probability to X≤q. Setting capacity at or above q limits shortage probability to about 5% under the model.
When capacity is fixed, choose bookings so the show-up quantile does not exceed seats under a defined threshold.
The threshold is a policy choice. Model error means an estimated 5% is not a promise.
Value at Risk can hide tail severity
A 99% value-at-risk measure says a loss threshold is not exceeded in 99% of cases. It says little about how bad the remaining 1% becomes.
Conditional tail expectation averages losses beyond a high quantile and therefore retains some severity information.
Both depend on the cost model and data. A neat risk number cannot restore omitted passenger consequences.
Simulation handles complex attendance rules
Monte Carlo simulation generates many fictional flights. For each booking, sample attendance under the chosen probability and dependence model, then compute empty seats, overflow and cost.
After 100,000 trials, averages and quantiles approximate model outcomes. Use a fixed random seed for reproducibility and verify the simulator against a tiny exact case.
Simulation does not make assumptions true. It efficiently calculates their consequences.
A normal approximation needs caution
For large N with p away from zero and one, a binomial can be approximated by a normal distribution with mean Np and variance Np(1−p).
Continuity correction improves a threshold calculation: approximate P(X≤100) using P(Y≤100.5) for continuous normal Y.
When tail probabilities are small or p is near one, approximation error may matter. Exact computation or simulation is often feasible.
Forecast calibration checks probability honesty
Collect flights assigned shortage probability near 10%. If roughly 10% actually experience shortages over enough comparable cases, the forecasts are calibrated in that band.
A model can rank flights correctly but be poorly calibrated, systematically turning a true 20% risk into a reported 10%.
Reliability plots and Brier scores help evaluate probability forecasts, but operational changes can shift the data-generating process.
The Brier score measures squared probability error
For binary outcome y in {0,1} and forecast p, Brier score is (p−y)^2. Lower is better.
If shortage forecast is 0.2 and no shortage occurs, score is 0.04. If shortage occurs, score is 0.64. Over many flights, average the scores.
The Brier score mixes calibration and discrimination. It should be compared with simple baselines such as the historical base rate.
Demand forecasting and show-up forecasting are separate
Demand forecasting estimates how many people want to book at each price. Show-up forecasting estimates attendance among booked passengers.
Overbooking can matter only after sales, but booking limits also depend on future demand and fare classes. Selling one low-fare seat now may displace a later high-fare request.
Revenue management therefore combines inventory control, demand uncertainty and attendance risk. This article owns the show-up capacity mechanism, not the whole airline business.
Nested fare classes create opportunity cost
A carrier may protect some seats for later demand while selling lower fares earlier. The value of keeping a seat open is uncertain.
If probability of a later request is 0.4 and its additional contribution is $150, simplified expected value of protection is $60. Compare that with the current offer’s contribution.
Real systems handle networks, cancellations and competition. The small calculation introduces opportunity cost without prescribing pricing.
Cancellations and no-shows differ
A cancellation before departure can free inventory for resale; a no-show may be learned too late. Refundability and rebooking rules affect behaviour and revenue.
Data should distinguish cancelled, rebooked, checked-in and absent records. Combining them into one “did not fly” rate erases timing.
The model’s state transitions can be represented with conditional probabilities or survival analysis.
Survival analysis models time until change
Time-to-cancellation can be treated as a survival process. The hazard rate describes the instantaneous chance of cancellation given that the booking remains active.
Covariates and time effects can change the hazard. Holiday bookings may behave differently from weekday business travel.
Censoring matters because bookings still active at the data cutoff have not yet revealed a cancellation time. Treating them as permanent distorts the estimate.
Rebooking creates network effects
A displaced passenger must be placed on another service with its own capacity. One oversold flight can move demand into later flights.
If the next flight is nearly full, expected reaccommodation cost rises. A network model considers downstream seats rather than one departure in isolation.
Local optimisation can therefore create system-wide congestion. This is closely related to queueing and flow problems.
Crew and aircraft changes create other shortages
Passengers can be denied a seat for reasons beyond oversales, including aircraft substitutions or operational restrictions. These mechanisms should not be mixed in an overbooking evaluation.
US DOT guidance distinguishes oversales from several other situations when describing passenger rights. The legal treatment depends on circumstances and jurisdiction.
Clean data labels prevent a forecasting model from learning the wrong cause.
Passenger rights are jurisdiction-specific and current
The US Department of Transportation’s Bumping and Oversales guide explains that airlines must first seek volunteers before involuntary denied boarding on covered oversold flights and describes current U.S. conditions and compensation rules.
That page was current when checked in October 2026, but rules can change and do not automatically apply to departures elsewhere. Travellers should consult the responsible authority and carrier’s current contract.
The classroom model never overrides a right. Legal constraints enter before profit optimisation.
Volunteering creates an auction-like process
When a carrier asks for volunteers, increasing compensation can attract more passengers willing to travel later. This resembles a reverse auction for scarce seats.
If one volunteer accepts at $300 and another would accept at $500, the carrier does not know every private valuation in advance. Timing and information affect the outcome.
A fair process needs clear terms, confirmed alternative travel and disclosure. Mathematical efficiency is not enough without informed consent.
Selection rules require fairness auditing
If volunteers are insufficient, a carrier may have boarding-priority rules within applicable law. Criteria can create unequal impacts across passenger groups.
Audit denial rates by relevant protected and service categories where lawful and appropriate. Raw counts need exposure denominators: compare denials per eligible passenger, not total denials alone.
A statistically small disparity may still matter ethically, and a large disparity needs investigation rather than automatic causal accusation.
Families and groups create indivisible constraints
A party of four may prefer to remain together. Solving a one-seat shortage by moving one child alone is not an acceptable generic operation.
Capacity allocation becomes an integer programme with group, accessibility and connection constraints. Feasible solutions are not every combination of passengers.
The human structure of bookings matters as much as the number 100.
Accessibility needs can change consequences
A later flight may not provide the same assistance, seating or connection. Average compensation does not measure individual disruption.
Policies should account for passengers whose alternatives are constrained. The model can include larger costs or hard rules, but the values require human and legal governance.
Treating everyone as an identical unit makes the spreadsheet simpler and the decision less accurate.
Data drift changes show-up behaviour
A new check-in rule, pandemic, rail disruption, festival or schedule change can alter attendance patterns. A model trained on last year may miscalibrate suddenly.
Monitor rates over time and by route. Use control charts or rolling estimates to detect persistent shifts without reacting to every random fluctuation.
Retraining alone is not enough; teams need to understand which process changed and whether historical variables retain meaning.
Seasonal averages can hide peaks
An annual 95% show rate can combine 90% in one season and 99% in another. Using the annual figure for both misstates tail risk.
Segmenting by season helps, but too many small segments create noisy estimates. Hierarchical models can partially pool information across related routes.
The aim is neither one giant average nor a separate fragile model for every flight. It is an evidence-based level of detail.
Small routes need uncertainty-aware estimates
A route with 19 shows out of 20 has the same observed 95% rate as 4,750 out of 5,000, but far less evidence.
Confidence intervals or Bayesian shrinkage can pull the small-route estimate toward a broader network pattern while preserving local data.
Reporting both rate and sample size prevents false precision. Three decimal places cannot compensate for twenty observations.
Prediction intervals matter more than point forecasts
Saying “100 passengers will show” hides variability. A prediction distribution might assign substantial probability from 96 to 103.
Capacity decisions use that distribution to calculate shortages and empty seats. The interval should include parameter and process uncertainty where practical.
A narrow interval built from an overconfident model is more dangerous than an honest wider one.
Sensitivity analysis tests fragile assumptions
Calculate the policy under p=0.94, 0.95 and 0.96. A one-point change may sharply increase shortage probability near capacity.
Change the cost of displacement, group correlation and re-accommodation availability. If the recommended booking limit changes often, the decision is fragile.
Robust policies perform acceptably across plausible inputs rather than winning only under one convenient estimate.
Scenario planning keeps rare disruptions visible
A normal day and a severe-weather day have different dependence, cancellations and rebooking capacity. Mixing them into one average can hide operational stress.
Model scenarios separately and attach probabilities cautiously. Stress tests can ask what happens when every inbound connection is delayed or the substitute aircraft is smaller.
Stress tests are not forecasts of exact events. They reveal where the system fails and what contingency is needed.
Expected revenue needs realised outcomes
Forecasted revenue from an additional booking may disappear through refund, rebooking or compensation. Track realised contribution after operational costs.
If a policy estimates $10 extra per booking but realised net is $2 with more complaints, the model boundary was too narrow or its probabilities changed.
Post-decision measurement closes the loop. A forecast is accountable to outcomes, not impressive equations.
Goodhart’s law warns against one target
When a measure becomes the sole target, people and systems can optimise it in damaging ways. Maximising load factor alone may create too many shortages.
Balancing load factor, involuntary denials, voluntary re-accommodation, complaints and on-time performance gives a fuller view.
No dashboard eliminates judgement. Metrics should reveal trade-offs rather than combine every value into one flattering score.
A beta-binomial model allows extra variation
The binomial assumes one fixed show probability p. A beta-binomial model instead treats p as varying across flights according to a beta distribution.
This creates more variance than a simple binomial and can represent unobserved day-to-day conditions. Parameters are estimated from historical counts, not chosen to make a preferred booking limit look safe.
Extra flexibility is valuable only when validation supports it. A complex distribution can overfit a short history just as easily as a complex prediction model.
Bayesian updating can learn from route history
Begin with a beta prior Beta(α,β) for show probability. After observing s shows and f no-shows under a simplified independent model, the posterior is Beta(α+s,β+f).
With prior Beta(19,1), mean is 95%. Observing 90 shows and 10 no-shows gives posterior Beta(109,11), mean about 90.8%.
The calculation combines prior information and new evidence visibly. The prior’s strength should reflect real knowledge, and dependence can invalidate the simple conjugate model.
Posterior predictive risk includes parameter uncertainty
Plugging one estimated p into a binomial treats that estimate as certain. A posterior predictive distribution averages binomial outcomes across plausible p values.
This usually produces wider tails than the plug-in model, especially with little data. A small route therefore receives more cautious shortage estimates.
Uncertainty should influence decisions, not appear only in a technical appendix after the booking limit is fixed.
The newsvendor analogy compares underage and overage
The classical newsvendor problem chooses inventory before uncertain demand. Ordering too little loses sales; ordering too much leaves unsold stock.
Overbooking reverses some labels but shares the critical-fractile logic: compare marginal empty-capacity cost with marginal shortage cost and choose a quantile.
Airline seats add cancellations, networks and passenger rights, so the analogy is structural rather than complete. A useful analogy reveals a mechanism without erasing domain differences.
Robust optimisation protects against probability error
Instead of trusting one p, define a plausible interval such as 0.94 to 0.97 and evaluate every candidate booking limit across it.
A robust policy may maximise worst-case value or minimise worst-case regret relative to the best policy for each p. It often accepts less average profit for more stability.
The interval must be evidence-based. Making it arbitrarily wide can force an overly conservative solution, while a narrow interval creates false confidence.
Chance constraints make a risk limit explicit
A chance-constrained policy might require P(X>C)≤1%. The optimisation then chooses the largest booking count meeting that modelled limit.
This is clearer than hiding passenger risk inside a monetary cost. The organisation can debate whether 1% is acceptable before viewing profit results.
The guarantee is only as good as the distribution. Backtesting checks whether realised shortages stay near the stated tolerance.
Standby passengers are not confirmed overbookings
A standby list offers travel if capacity becomes available. A confirmed reservation carries different expectations and legal treatment.
Combining both in one “booked” count misstates obligation. The state model should distinguish confirmed, standby, checked-in, boarded, rebooked and cancelled passengers.
Precise categories improve both forecasting and communication. A data table can be numerically complete and conceptually wrong if statuses are blurred.
Reminder messages can change the process
An airline or event organiser might send check-in reminders. If reminders reduce no-shows, the historical rate without reminders no longer predicts the new policy.
A randomised test can estimate the causal effect, provided assignment, outcome and interference are handled properly. Travelling parties may share reminders, violating individual independence.
Operational interventions should be evaluated as process changes, not treated as free predictors that leave behaviour untouched.
Strategic behaviour can respond to rules
If compensation becomes predictable, some flexible travellers may volunteer more often. If check-in priority determines boarding, people may check in earlier.
The system is not a passive probability machine; published rules change incentives. Game-theoretic reasoning asks how participants adapt.
A policy estimated from old behaviour can shift the very distribution it relied on. Monitoring after rollout is part of the model lifecycle.
Confidence intervals for rare denials need care
If only two involuntary denials occur among 100,000 passengers, the observed rate is 0.002%. Normal approximations can behave poorly near zero.
Exact binomial or suitable interval methods give asymmetric uncertainty. Rates should also be stratified carefully without creating tiny, unstable subgroups.
“Almost zero” is not the same as impossible, and one severe incident can matter beyond its frequency.
Load factor and denial rate use different denominators
Load factor often compares occupied seat capacity with available capacity. Denial rate may compare denied boardings with eligible passengers or enplanements under a defined reporting rule.
A 90% load factor and 0.1% denial rate cannot be subtracted or combined directly. They describe different populations and questions.
Every dashboard should place denominator and reporting period beside the percentage. This protects readers from attractive but meaningless comparisons.
Independent audit improves accountability
Booking-limit code can be tested on synthetic edge cases: zero no-shows, all no-shows, one seat, one group larger than remaining capacity and a sudden aircraft change.
Separate teams can reproduce probability calculations and inspect whether rules match approved policy. Logs should record model version, inputs and override reasons.
Auditability does not prevent every bad outcome, but it makes hidden arithmetic harder to use as an excuse after harm occurs.
Did You Know? A mean below capacity can still overflow often
Students often assume an expected attendance of 99.75 means a 100-seat flight is safe. But expectation is a long-run average, not a ceiling.
The probability mass above 100 may remain meaningful. A small standard deviation does not remove the tail.
This is why probability distributions matter in staffing, inventory, flood planning and exam-room capacity—not only airlines.
A student project: model a fictional school coach
Use a fictional 40-seat coach for a school event, not a real transport decision. Estimate attendance probability from supplied classroom data rather than personal records.
Compare 40, 41, 42 and 43 registrations under a binomial model. Calculate expected attendance, shortage probability, expected overflow and expected empty seats.
Add a hard rule that no child can be left behind. Explain why this makes deliberate overbooking inappropriate even if expected cost looks favourable.
A second project: test dependence with simulation
Simulate 50 individual travellers at 95% show probability. Then replace 20 of them with five groups of four that show together with 95% group probability.
Compare means, standard deviations and probabilities above a chosen capacity. The expected total may remain similar while variability changes.
Repeat with a common disruption affecting everyone. The project shows why independence is an assumption, not a default fact.
Practical learning steps for students
Begin with proportions, combinations, binomial probability, expected value and standard deviation. Calculate tiny cases exactly before using software.
Next learn conditional probability, simulation, calibration, optimisation, survival analysis and time series. Keep training and evaluation periods separate.
Always state capacity, show-up definition, forecast horizon, costs, constraints and rights. The arithmetic becomes trustworthy when the decision boundary is explicit.
Parent guidance: keep passenger rights separate from the model
A probability exercise can explain why overbooking exists without excusing poor treatment. Business logic and passenger rights are separate layers.
For a real trip, check the carrier and relevant authority’s current conditions, arrive by required deadlines and keep documents. Do not rely on a classroom compensation example.
Why Mathematics? | Queues, Waiting Times and Service Decisions explains waiting systems, while the current article focuses on uncertain attendance against fixed seats.
Learning and careers should remain open
Capacity-risk mathematics appears in aviation, hospitality, healthcare scheduling, events, operations research, revenue management, data science and logistics.
These roles also require law, safety, software, communication and domain experience. A probability model does not authorise operational decisions by itself.
The authorised mathematics keyword cloud includes “maths for pilots” as a genuine reader intent. Students can explore aviation mathematics while keeping pathways broad through statistics, computing and geography.
Questions students and parents often ask
Why sell more tickets than seats?
Because some booked passengers do not travel and empty capacity expires at departure. The policy trades empty-seat risk against shortage risk.
Does expected attendance below capacity mean everyone will fit?
No. Expected value is an average. The attendance distribution can still exceed capacity.
Are show-ups independent?
Often not. Families, connections, weather and common disruptions create correlation.
Is the most profitable booking limit automatically best?
No. Law, fairness, service, safety, tail risk and passenger consequences constrain the decision.
Do the same passenger rights apply everywhere?
No. Rules depend on jurisdiction, route, cause and current law. Use the relevant authority and carrier’s current information.
What mathematics should a student learn next?
Practise binomial probability, expectation, variance, quantiles, simulation, calibration, optimisation and sensitivity analysis.
Mathematics makes capacity risk visible
Airline overbooking matters mathematically because one fixed seat map meets uncertain human behaviour. Probability describes the show-up distribution. Expected value compares competing costs. Quantiles and simulations expose tails. Constraints protect rights and service.
The hopeful lesson is not that every shortage can be eliminated. It is that assumptions can be tested, policies can be audited and avoidable harm can be measured rather than dismissed as bad luck.
A responsible model keeps the passenger inside the mathematics. It records groups, connections, alternatives, uncertainty and rights instead of treating everyone as identical points. That wider frame turns optimisation from a narrow revenue exercise into accountable decision support.
The decision also changes across time. Months before departure, the useful quantity is a broad demand-and-cancellation forecast. Near check-in, confirmed connections and party status narrow uncertainty. At the gate, the task becomes a constrained allocation with actual people, actual alternatives and enforceable rights. Reusing one static probability at every stage mistakes convenience for consistency. A mature system updates the forecast while preserving a trace of which evidence changed the decision.
For students, the topic is a compact laboratory for probability literacy. The expected number can sit below capacity while the upper tail crosses it. Two groups can share the same mean and have different variance. A lower shortage probability can still carry a worse tail when it occurs. Those distinctions transfer to hospital appointments, theatre tickets, restaurant bookings and cloud-computing capacity, but each application needs its own constraints and ethics.
Most importantly, a number such as “one percent risk” should trigger three follow-up questions: one percent of which flights, under which data and what happens to whom in that one percent? The first identifies the denominator, the second tests the model and the third restores human consequence. Probability becomes responsible when all three remain visible beside the optimisation result.
The same transparency should govern overrides. Operations staff sometimes possess current information that the model lacks, such as a delayed connection carrying a large school group or an aircraft change that invalidates the seat count. An override can be sensible, but its reason should be recorded and reviewed. If overrides consistently improve outcomes, the model is missing a feature; if they consistently worsen outcomes, the workflow needs coaching or tighter rules. Human judgement and automation learn best when disagreement becomes data rather than blame.
Finally, evaluate the policy across a full journey. A passenger rebooked “successfully” may arrive too late for the purpose of travel, lose prepaid transport or face an inaccessible replacement. Counting only whether another seat was found understates consequence. Good mathematics follows the decision to its real outcome and keeps service quality beside load factor.
Continue through the Mathematics Learning Hub or read Why Mathematics? | Voting, Fair Shares and Apportionment for another problem where indivisible capacity and explicit rules determine who receives a place.
