The core of VisitorCast's prediction engine is a Machine-Learning time series model — a well-established statistical forecasting approach, used and trusted far beyond museums. Because the model is statistical and fully transparent, every forecast is repeatable, can be validated against your own history, and can be broken down so you can see exactly what is driving the numbers.

The five layers of the model

In a simplified explanation, the model builds every forecast from five layers:

  • Trend — the long-term, non-seasonal direction of your visitor numbers. Is attendance gradually growing, stable, or declining? The trend is allowed to change direction over time, so a museum that turned things around two years ago is not held back by older history.
  • Seasonality — recurring, calendar-driven cycles. The model learns two: your yearly rhythm (summer peaks, winter lulls) and your weekly rhythm (how a Saturday differs from a Tuesday).
  • System drivers — demand drivers that VisitorCast maintains for you: public holidays, school vacations, paydays, and weather. Every museum is different, so each museum analyses which of these actually influence its visitor numbers and enables only the drivers with real impact.
  • Museum drivers — your own exhibitions and events, added in VisitorCast by your team. These work like system drivers, with two refinements: they are weighted by popularity, and exhibitions can have separate opening-week and closing-week effects, because a blockbuster's first and last weeks often behave differently from the middle of its run.
  • The unexplained rest — no model captures everything. Whatever the layers above cannot explain — one-off surprises, random day-to-day variation — is treated as noise rather than forced into a pattern. This honesty is also why every forecast comes with a confidence band instead of a single false-precision number.

An additive model — every driver counts in visitors

The model is additive: the forecast for a given day is the sum of the layers. Trend plus seasonality, plus what each enabled driver adds or subtracts.

This is what makes VisitorCast's forecasts explainable. Because each layer contributes a concrete number of visitors, VisitorCast can show you that, for example, a school vacation is lifting the coming month by thousands of visitors while a closing exhibition is pulling it slightly down. The driver cards on the Forecast page and the "What explains the difference" card on Analyse trends are read directly from this decomposition — they are the model's actual arithmetic, not an interpretation of it. The Drivers page answers a different question and uses a different method: it measures each signal's observed lift by comparing days when it was active to otherwise similar days, so its percentages will not match the visitor figures shown by the model.

What the model needs from you

To forecast your museum, the model needs at least 180 open days of daily visitor history. Closed days do not count toward this minimum — the model trains only on days you were actually open, so closed days never drag your forecast down. Two or more full years of history give the best results, because the model can then learn your yearly seasonal pattern reliably.

The quality of the forecast follows the quality of the data behind it: clean daily numbers, correct closed days, and well-maintained exhibition and event records make a real difference.

Uncertainty

No forecast is exact, and VisitorCast does not pretend otherwise. Every prediction comes with a confidence band showing the range your visitor numbers are expected to fall within on 8 out of 10 days. A narrow band means the model finds your museum predictable; a wide band is a signal to plan with more buffer.

You never have to take the model's word for it. Run an accuracy check, and VisitorCast will predict a period that has already happened and show you the forecast next to your actual numbers.

Where AI fits in

At the core of VisitorCast is a statistical time-series forecasting model that uses machine-learning techniques to learn from historical visitor data. Machine learning is a branch of artificial intelligence.

VisitorCast also uses AI-assisted analysis to validate data, identify patterns and evaluate potential forecasting factors—such as weather, holidays, exhibitions and events - to determine whether they improve forecast accuracy.