The problem
Planning runs on last quarter’s spreadsheet plus intuition. Vendor "AI forecasting" is a black box nobody trusts, so it decorates a slide and the spreadsheet still decides.
The system
A forecasting pipeline that starts with strong statistical baselines, adds ML only where backtests prove lift, quantifies uncertainty as intervals rather than point promises, and publishes its own error metrics every cycle — into the sheet or tool planners already use.
How it's built
- Feature store over your sales/ops history with holiday/seasonality handling
- Rolling-origin backtests; champion/challenger promotion only on measured lift
- Prediction intervals surfaced, not hidden; error dashboards per segment
- Scheduled retraining with drift alerts
Delivery
Sprint backtests on your history and reports achievable accuracy before you commit; Build automates the pipeline.
What to expect
- Forecast error known per horizon and segment — before you rely on it
- Planners keep their workflow; the numbers in it get better
- A model that loses to the baseline never ships
Documented results in the wild
Independent, published deployments of this class of system — cited as market evidence that it works at scale. These are not our clients.
- OTTO Runs 30 billion demand forecasts a month across 2.5M SKUs up to 450 days ahead; 35% of ranges reorder fully automatically. OTTO / The Economist ↗
- Instacart Tiered real-time availability models across ~100,000 stores cut prediction compute costs by ~80%. Instacart Engineering, 2023 ↗