Forecasting Skills for Supply Chain

2 min read

Supply-chain forecasting is not a contest to build the most complex model. The objective is to create a useful demand estimate, understand its error and bias, and improve downstream decisions.

Forecast problem setup

Define:

  • what is forecast;
  • at what level;
  • horizon;
  • time bucket;
  • lead time;
  • decision supported.

A monthly family-level forecast and a daily SKU-location forecast are different problems.

Baseline methods

Always create a simple baseline.

Examples:

  • last period;
  • seasonal naïve;
  • moving average;
  • simple exponential smoothing.

A complex model should beat a sensible baseline.

Seasonality and trend

Check whether demand shows:

  • trend;
  • weekly/monthly seasonality;
  • promotion spikes;
  • intermittent behavior;
  • lifecycle effects.

Do not force seasonality into sparse data.

MAE, MAPE and WAPE context

MAE

Easy to interpret in original units.

MAPE

Intuitive percentage but problematic when actual demand is near zero.

WAPE

Useful for aggregate percentage error across a group, but still needs careful interpretation.

No single metric should replace business judgment.

Bias

Bias indicates systematic over- or under-forecasting.

Persistent positive or negative bias can drive:

  • excess inventory;
  • shortages;
  • poor capacity decisions.

Track bias alongside absolute error.

Forecast value add

Ask whether manual overrides improve the baseline.

If sales overrides repeatedly worsen accuracy, the planning process should learn from that.

Overrides

A justified override should have:

  • reason;
  • owner;
  • size;
  • time horizon;
  • expected event.

Track the outcome.

Validation

Use out-of-sample or rolling validation where possible.

Avoid selecting a method based only on the same history used to fit it.

Project

Build a forecast comparison across product types.

Include:

  • baseline;
  • two or three methods;
  • error;
  • bias;
  • segmentation;
  • override scenario;
  • inventory/service implications.

The strongest forecaster knows when a simple method is enough and when the process, not the model, is the real problem.

Related content

Sources