Skip to content
SQ

Forecasting2 July 20265 min

Ship distributions, not numbers

Why every forecast I put in front of a planner carries an interval, and what that does to the decision.

A demand forecast that returns one number invites a question it cannot answer: how wrong could this be? The planner's real decision is not 'how much will sell' but 'how much do I order given that being short is worse than being long'. That asymmetry only exists in a distribution.

What changes in the model

In the TCG forecasting system, per-SKU demand is modelled with gradient boosting over lag, rolling-window, trend, calendar and promotion features. The useful part is the quantile output: once you have the spread, stockout risk, safety stock and reorder quantity are derived quantities, not separate guesses.

  • Point forecast answers 'what'. Quantiles answer 'what do I do about it'.
  • Service level becomes an explicit input, chosen by the business, not implied by model error.
  • Back-testing gets honest: you evaluate calibration, not just mean absolute error.

Where the LLM belongs

The hybrid part of that system is often misread. The language model does not produce the forecast. It reasons over external context and explains drivers around a number that came from a back-testable statistical model. A forecast you cannot validate historically is not a forecast; it is a sentence.

Tabular data still favours boosting

With limited history per SKU and strong engineered features, LightGBM remained the right instrument — faster to train, easier to interrogate, and good enough that the marginal accuracy of a heavier sequence model would not have changed a single order quantity.