A retail analytics platform replaced a legacy statistical pipeline with a gradient-boosted forecasting system built by Toptal AI experts.
Strata’s weekly ARIMA-based forecasts degraded badly during promotions and new-product launches, forcing category managers to override half of the outputs by hand.
The data team was fully committed to reporting workloads and had no bandwidth to design, train, and productionize a modern forecasting stack.
Toptal assembled an ML engineer, a data engineer, and a forecasting specialist who shipped a gradient-boosted model with hierarchical reconciliation on Strata’s existing warehouse.
Every model change ran through an automated backtest harness covering 104 weeks of history, so accuracy claims were verified before deployment.
Weighted MAPE dropped from 24% to 15% across the top 500 SKUs, and manual overrides fell below 10% of forecasts.
The new pipeline retrains nightly and publishes forecasts before the merchandising team starts the day.
The backtest harness alone changed how we talk about accuracy internally. We finally trust our own numbers.
