Most demand forecasting in marketing was built for retail, and retail is a forgiving problem. Inventory can be replenished. The customer comes back. A missed sale today is a sale next week. Live entertainment breaks all three assumptions at once, which is why generic forecasting tools produce confident numbers that fail quietly.

Four structural differences

Capacity is fixed. A show cannot sell more than the room holds. That sounds obvious until you notice what it does to optimisation: past a certain point, additional spend cannot produce additional revenue, only additional cost. A retail model has no concept of a ceiling and will keep recommending scale.

The deadline is absolute. An unsold seat at showtime is worth nothing, permanently. There is no clearance, no next season, no markdown that recovers part of it. This makes the value of a ticket time-dependent in a way no retail SKU is, and it means the cost you should accept per ticket changes as the date approaches.

The purchase is a one-off. Lifetime value logic, the engine behind most acquisition models, barely applies. Someone who buys for a Hans Zimmer night is not reliably a buyer for the next one. Frequency exists at the fanbase level and at the venue level, rarely at the show level.

Demand is lumpy and event-driven. An announcement, a press hit, a festival line-up drop, a competitor on-sale in the same city. Demand arrives in bursts triggered by external events, not as a smooth baseline with seasonality on top.

What a ticketing-native forecast needs

A forecast that is actually useful for a run has to answer a narrower and harder question than retail forecasting does: given what has sold so far, what is the probable final sell-through of this date, and what does that imply for spend today?

• Pace, not volume. The input is percentage of capacity sold against time remaining, per date, not cumulative revenue.
• Comparable curves. The reference set is previous runs of similar genre, capacity, market and lead time, because the shape of a sell-through curve is more transferable than its level.
• Price band structure. A run selling out its cheapest band while the top band sits still has a different future than the aggregate suggests.
• The calendar around it. Other on-sales in the same market and week are part of the demand environment, not noise.

The decision the forecast is actually for

Forecasting in live entertainment is not a reporting exercise. It exists to answer one recurring question: should the next euro go to this date, another date, or nowhere. A forecast that cannot be turned into that allocation, per date, per day, is an interesting chart.

This is the part that requires the sales data rather than platform data. The signal that a date will finish at eighty percent instead of sixty is in the ticketing ledger, days before it is visible anywhere else. A model that reads only campaign metrics is forecasting its own activity.

Confidence, and knowing when to ignore it

Every forecast on a live run carries wide uncertainty early and narrows fast. In the first days after an on-sale the honest answer is a range. By the second week the range is usually tight enough to act on, and by the final fortnight the forecast is mostly arithmetic. The practical discipline is to size decisions to the confidence available: small, reversible moves early, decisive reallocation once the curve has settled.

The NYBA model was trained on ticketing signal from more than 1,200 events a year, which is what makes the comparable-curve step possible. It is a narrow model for a narrow problem, and the narrowness is the point.