Forecasts drive planning for sales, staffing, inventory and budgets.
Start With Baselines
- Naive: next value equals the last value.
- Seasonal naive: equals the same period last year or week.
- Moving average.
Sophisticated methods must beat these to be worth using.
Statistical Methods
- Exponential smoothing: weights recent observations more; handles trend and seasonality.
- ARIMA: models autocorrelation in the series.
- Regression with external variables: holidays, prices, promotions.
Machine Learning Methods
Gradient-boosted trees and neural networks can use many features and learn across many related series. Pretrained forecasting foundation models are also emerging.
Evaluation
Use time-based backtesting: train on the past, test on later periods, repeated across several points in time.
Communicate Uncertainty
Provide ranges, not just point forecasts, and explain assumptions.
Combine Judgement
Planners know about events models can't see. Structured adjustments with tracking of their accuracy work better than ad hoc overrides.