Forecasting

Predicting future values from past patterns, with honest uncertainty.

Also known as: forecasting, demand forecasting, sales forecast

Forecasting predicts future values of a time series from its history: next quarter’s revenue, tomorrow’s server load, next week’s signups. Every method — from eyeballing the trend to ARIMA — is a bet that past patterns persist, plus machinery for turning that bet into numbers.

history ──▶ model (trend + season + recent pace) ──▶ future values + uncertainty

The workflow matters more than the method. Start with a naive forecast as the baseline, add one model at a time, compare on held-out trailing data with a real accuracy measure (forecast accuracy), and ship intervals, not just points. Most business forecasting failures are process failures — no baseline, no holdout, no uncertainty — not algorithm failures.

The classic mistakes:

Forecasting what you should scenario-plan. A forecast answers “what will happen if patterns persist”. A product launch, a price change, a competitor move breaks persistence — that needs scenarios, not extrapolation.
Complexity before baseline. A neural net that barely beats naive on a holdout is a maintenance burden with no payoff. Earn complexity with measured gains.
Point forecasts as commitments. A naked number becomes a promise the moment it leaves your laptop. Always attach the interval.
One model forever. Series change shape — growth saturates, seasons shift, regimes break. Re-validate on recent data; retire models whose edge over naive vanished.

When not to forecast: when the decision does not need it (a range or a trigger level suffices), when history is too short or the world too changed, or when the honest answer is “we cannot know yet”. A refused forecast beats a precise-looking fiction.