Overview
Half of the chain's snack-category sales come from items that live on the menu for a few weeks at a time. Every one of those items has to be ordered, staged, and staffed months before a customer ever sees it — and every one of them launches without a single receipt to forecast against. Too high means waste. Too low means empty trays on launch week.
Lower MAPE means a more accurate forecast. Industry baselines for LTO items with no direct history sit at 25–35%. We landed well under the target with a Bidirectional LSTM.
Challenge | Three things made the problem hard.
The chain set the bar below 20% mean error on a fourteen-day forecast for an item with no direct history — under what's generally considered realistic for new-product forecasting, where the absence of sales history usually pushes error far higher.
- No direct history. The product being forecast had never been sold. Any usable signal had to come from related items, comparable promotions, and the chain's broader seasonal patterns.
- Promotion changes the shape of demand. Customers don't buy a limited-time offer the way they buy a regular menu item. The lift from being new had to be modelled explicitly, not assumed away.
- One mistake is a lot of food. Supply commitments go out eight to twelve weeks before launch. A forecast too high becomes inventory the chain pays to destroy. Too low means lost sales and a bad opening week.
The work | A bake-off, then a single winning model.
We worked alongside the chain's demand planning, category, and supply teams. Because new items had no sales history, we built demand signals from comparable menu items, past promotions, seasonal patterns, and store-level performance. Then ran a structured bake-off across historical launches instead of choosing a model upfront. The scoring framework became part of the ongoing planning rhythm, letting the team monitor forecast accuracy launch over launch.
Trained BiLSTM
Sequence model artefacts, training pipeline, retraining notes.
Forecast window
14–28 day demand forecast for new LTO campaigns.
37-model evaluation
Competitive evaluation report across 3 paradigms.
“It's the first time we've taken a forecast for a new product into a planning meeting and stopped arguing about the number.”VP, Demand PlanningNational quick-service restaurant chain
Outcomes | A forecast the planning team can take into a meeting and defend.
The trained model, the bake-off, the eval framework, and the recommendations for inventory and LTO planning - all yours.
About the client
Half the category's sales come from items that have never sold a unit. A national quick-service restaurant chain partnered with us to forecast demand for new limited-time items 8 to 12 weeks before launch, and brought new-product forecast error in under 18%.