01 / State the question
Choose an observable target
A ticket-demand forecast should begin with a sentence that a venue team can challenge: estimate paid bookings by event day, for a defined fixture and product, conditional on a specified offer. That is narrower than predicting “fan interest,” but it gives analysts a measurable target. Tickets sold, seats issued and gate entries are different quantities. Choose one primary target and explain how refunds, transfers and complimentary tickets affect it.
Also specify when the forecast is made. An estimate produced before general sale has different information available from one produced shortly before kickoff. A model that sees final prices, eventual attendance or late campaign results during training may appear accurate while being unusable at the real decision point. Preserve historical snapshots rather than rebuilding the past from today’s ticketing dashboard.
02 / Make inputs comparable
Reconstruct the offer, not just the orders
Record inventory availability, advertised prices, mandatory fees, section definitions and eligibility rules over time. Membership windows and season-ticket entitlements change who can purchase. A block released late has a shorter selling opportunity than an otherwise similar block offered from launch. Without that distinction, a model may learn the release policy and mislabel it as weak demand.
Context matters, but more columns do not guarantee better evidence. Opponent, competition, event time, public holidays and known transport disruption can be useful if consistently defined. Wembley Stadium, Stade de France, Olympiastadion Berlin and Helsinki Olympic Stadium illustrate why venue identity alone is insufficient: event format, local travel and ticketing arrangements vary. These are contextual examples, not sources of shared performance data.
| Input | Working definition | Failure to check |
|---|---|---|
| Booking time | Timestamp relative to on-sale and event time | Different sales windows look comparable |
| Available stock | Sellable seats at the observation time | Stock limits are mistaken for low interest |
| Ticket value | Consistent treatment of fees and refunds | Price effects are misread |
03 / Identify missing demand
Treat sellouts as censored observations
Once a section sells out, subsequent sales cannot reveal its unconstrained demand. A flat curve might mean interest stopped, but it might simply mean customers had nothing left to buy. Record the stock boundary. Waiting-list registrations, unsuccessful searches and abandoned baskets may add information, although none is a clean count of unique people willing to pay.
Do not manufacture certainty by filling those gaps with a single estimate. Compare plausible assumptions about lost demand and show how the forecast changes. Customer journeys may move across sections or channels, so summing unsuccessful searches can double-count the same person. If the evidence is insufficient, describe the result as a forecast of observed sales under the existing inventory policy.
04 / Test forward
Ask whether complexity earns its place
Start with a seasonal average, a comparable-event booking curve or a simple regression. Evaluate more complex methods against the same baseline and information cutoff. Gradient-boosted trees can capture nonlinear associations; hierarchical models can share information across related event groups. Neither removes the need for meaningful labels, stable inputs and sufficient relevant history.
Use time-ordered validation: train on earlier events and evaluate on later ones. Randomly splitting individual orders can put nearly identical event conditions in both training and evaluation sets. Report errors by forecast horizon and event group, not only as a single overall score. Examine whether uncertainty intervals have reasonable coverage and whether errors systematically change around campaigns, release dates or unusually popular fixtures.
05 / Hand over responsibly
Deliver a decision-ready forecast
The handover should state the target, current offer assumptions, expected range, known data gaps and conditions that would trigger revision. Avoid precision unsupported by the sample. A forecast for a new competition or substantially changed venue configuration should carry an explicit out-of-scope warning rather than borrowing confidence from unrelated historical events.
Finally, separate prediction from intervention. A model trained on existing prices cannot reliably say what would happen after a price change merely by replacing one input value. Price may have been chosen because staff already expected strong interest. Use the forecast to identify an operational question, then evaluate the intervention through an appropriate test and approval process.
Limitation: search activity is not willingness to pay
Searches can reflect repeat visits, bots, curiosity, unavailable products or customer-service journeys. Use them only with clear definitions, proportionate privacy safeguards and evidence that they improve out-of-time evaluation.