Intermittent Demand and Cold-Start Forecasting for Grocery Replenishment
A regional grocery chain operates 240 stores and forecasts 18,000 store-SKU combinations weekly. About 60% of series have zero sales in more than half of weeks, and 1,200 new items launch each year. The current XGBoost model uses 4-week lag features and achieves WAPE 0.41 across all series, but WAPE is 0.78 for slow movers and bias is -0.22 for new items. Automated replenishment uses the forecasts: perishables have a 2-day lead time and dry goods 14 days. Last quarter, new items had 30% stockouts in their first four weeks, while slow movers had 19% excess inventory. The planning team wants one forecast per store-SKU per week, and the CIO asks whether a single global model can handle both fast and intermittent items.
1. What modeling approach would you recommend for intermittent and cold-start series, and why?
Read the worked answer
Separate the problem by demand pattern but keep a shared global backbone. For intermittent series, use a global model with covariates and a loss suited to zeros, such as Tweedie or a zero-inflated negative binomial, or a TSB/Croston-style method as a baseline. Do not rely on 4-week lags alone; add calendar, price, promotion, store traffic, and item hierarchy features. For cold-start items, borrow from parent category and similar items, and use a hierarchical prior or transfer learning; a new item's forecast should start at the category-seasonal profile, not at zero. For fast movers, the existing gradient-boosted model can remain. A single global model can work if it includes segment interactions and item/store embeddings, but it must be evaluated by segment. The design should output a full distribution or at least quantiles for replenishment, because lead-time variability and service-level targets need more than a point forecast.
2. Which evaluation metrics and validation splits should govern model selection?
Read the worked answer
Use rolling-origin evaluation, not a random split: train through week t, forecast t+1 to t+4, roll forward weekly. Report MASE and pinball loss for scale-free and quantile accuracy, plus bias by segment. WAPE is useful for total volume but hides over- and under-forecasting, so pair it with signed bias. Evaluate fast movers, slow movers, and new items separately; an overall WAPE of 0.41 can hide 0.78 on slow movers.
Add business metrics: stockout rate, excess inventory, and service level. For model selection, require no segment to degrade beyond an agreed threshold, and use the lower quantiles for safety stock. Validate that the forecast horizon matches the 2-day and 14-day lead times; a one-week-ahead model cannot directly drive a 14-day dry-goods order without a multi-horizon forecast. Compare against a seasonal naive baseline per segment so gains are not overstated by aggregation.
3. How would you reconcile forecasts across store, region, and category for replenishment?
Read the worked answer
Forecast at the store-SKU level, then reconcile to store, region, and category totals. Bottom-up preserves local detail but can drift from category plans; top-down is stable but loses store variation. Use MinT or a simpler weighted reconciliation that accounts for forecast error covariance, then apply non-negativity and capacity constraints. For perishables, reconcile at the store-day level with a short horizon; for dry goods, reconcile weekly at store-SKU and check against warehouse capacity. Publish both the pre-reconciliation and post-reconciliation forecasts so planners can see adjustments. Reconcile before replenishment, not after, and recompute safety stock from the reconciled error distribution. The CIO's question is really about governance: one global model can produce the base forecasts, but the hierarchy and reconciliation layer must be explicit and owned by planning.