Supply Chain & Sales Forecasting
Gradient-boosted demand forecasting for wholesale distribution — product-group forecasts on a monthly planning cycle, built so planners trust the number enough to order against it.
- Role
- Team lead & delivery
- Method
- Gradient-boosted trees
- Grain
- Product group · monthly
- Stage
- Delivered · in rollout
The problem
Wholesale distribution runs on lead times. Ordering against last period’s numbers works right up until demand turns, and by then the correction is weeks out — supplier lead times do not shorten because the forecast was wrong. The cost lands on both sides of the miss: overstock ties up working capital in a warehouse, understock hands the sale to whoever did have it on the shelf.
The approach
I led the team and owned delivery. Three decisions shaped the work:
- The grain — forecasts at product-group level on a monthly cycle, because that is the grain procurement decisions are actually made at. Forecasting finer than the decision buys precision nobody can act on, and pays for it in noisier, sparser series.
- The model — gradient-boosted trees over engineered temporal features (lags, rolling statistics, calendar and seasonality effects) rather than a deep sequence model. On tabular business data at this scale it fits comparably or better, trains in a fraction of the time, and stays interpretable enough that a planner can see why a forecast moved.
- Interpretability as a delivery requirement, not a modelling preference — a planner who cannot explain a number will override it, and an overridden forecast changes nothing.
That last one is what makes a model part of the planning process instead of a dashboard sitting next to it.
The outcome
Delivered and in rollout, with procurement shifting from reacting to last period toward planning against a forecast — at the grain, and on the cycle, that the planning already ran on.