Learn
The model reads booking curves, seasonality, route behaviour and more than one hundred derived features.
Case study · Demand forecasting
9.7% median forecast error, down from ~42%.
A demand model gives planners an early forecast and an operational recommendation for every service.
Operating proof
9.7%Median errorFourteen days before departure
366MHistorical recordsUsed across the modelling programme
26Months of historyWith more than 100 engineered features
02 / Problem
Demand becomes visible after the capacity decision has already been made.
Sixty-six percent of tickets were bought in the final three days before departure, while train composition had to be planned much earlier. The reference model missed final demand by roughly 42%, leaving planners without a dependable early signal.
03 / System
A forecast that includes uncertainty and ends in a decision.
Eight architectures were tested before a hybrid model was selected. It produces P50, P85 and P95 demand ranges, is validated on periods it has not seen, and feeds an operational panel rather than a static report.
Forecast
Every service receives a demand range instead of one falsely precise number.
Recommend
The panel translates the forecast and its uncertainty into a planning suggestion.
04 / Outcome
A planning signal for every departure, delivered to the operator.
The system reached 9.7% median error fourteen days before departure. Its 90% prediction interval covered 88.6% of observations, and the operational panel was handed over with code and documentation.