A live AI system that forecasts where each trade show will finish — updated every day.
Trade show revenue builds slowly and unevenly across a long booking cycle. A running total tells an event organiser where they are — not where they will end up. Judging whether the current pace is good or bad against prior years is, at best, informed guesswork.
By the time a shortfall becomes obvious from the numbers, the window to push sales and marketing has almost closed. The gap between sensing something is wrong and knowing something is wrong costs real revenue.
The client wanted to close that gap entirely. Not a periodic report or a static model — a system that re-forecasts every event's revenue and space curve each day, as new bookings arrive, so the answer to "where will we end up?" is always current. And accurate enough to act on.
Ten million warehouse records across shows, booths, orders, exhibitors and customers — cleaned, unit-normalised, joined, and reshaped into a 12M-row daily pacing dataset purpose-built for time-series machine learning.
A single model can't span the full forecasting problem. An annual anchor needs to know how the event typically performs; a daily curve model needs to know where it's heading right now. Getting them to agree — with reconciled revenue and space targets — requires careful design.
A forecast nobody believes gets ignored. It had to be accurate enough, early enough in the booking cycle, that commercial teams would route decisions through it.
Two-stage forecasting architecture
An annual event-level anchor model — 1,500+ features reduced to a few hundred via rigorous selection — pins the trajectory. A daily curve model then forecasts forward recursively, updating as each day's bookings arrive and tightening its estimate as the event approaches.
Dual-target reconciliation
Revenue and space are forecast together and blended via a tuned integrity weighting — so the two curves stay consistent. Macro context (GDP, inflation, unemployment, equities, holidays) is folded into the feature set, grounding each forecast in the market conditions the event is actually selling into.
Fully automated, end to end
The entire pipeline — warehouse extraction, feature engineering, model inference, output generation — runs automatically on the cloud and refreshes daily. Raw data in, updated predictions out, feeding a live customer-facing dashboard. It shipped as a running product.
The full data-to-product arc, delivered by a lean two-person team: data engineering, time-series modelling, feature selection, dual-target reconciliation, macro integration, and end-to-end automation. From raw warehouse records to a live daily product in a single engagement.
Back-tested across ~120 events, predictions are accurate to within 3–4% of actual revenue at the median for a full booking cycle — with the best third within 1–2%, and forecasts tightening materially as each event approaches. The pipeline runs fully automatically, refreshed daily, and ships into a live product. What had been a question nobody could cleanly answer now has a number, updated every morning.