Demand forecasting with temporal fusion transformers
We used temporal fusion transformers to forecast item-level demand to under 10% MAPE - so stock and operations match real demand.
Time-series MLForecastingOperations
// The challenge
Where it started.
A US-based restaurant chain needed item-level demand forecasts accurate enough to plan stock and operational resources reliably.
// Our approach
- 01
Applied temporal fusion transformer (TFT) models to demand data.
- 02
Predicted demand at the item level.
- 03
Aligned stock and operational resources to the forecast.
- 04
Validated accuracy against held-out demand.
// The outcome
What it delivered.
<10%
MAPE achieved
item
level forecasts
TFT
transformer models
Item-level demand predicted to a high level of accuracy.
Achieved less than 10% MAPE for the restaurant chain.
Stock and resources planned to match predicted demand.
// More case studies