Skip to content
Time Series & Forecasting

Classical Forecasting

Naive Baselines: Beat Them or Go Home

7 min+65 XPStep 1/10

The dumbest forecast that isn't dumb

Four forecasts cost nothing and are embarrassingly hard to beat:

  • Mean: the next value is the historical average — decent for flat, noisy series
  • Naive: the next value equals the last observed value — tomorrow's temperature is today's
  • Seasonal naive: the next value copies one season ago — next Monday looks like last Monday
  • Drift: naive plus the average trend — draw a line from first point to last, then extend it

No training, no hyperparameters, no GPU. And yet on real demand data, seasonal naive routinely lands within striking distance of models that took weeks to build.

insight

Every forecasting method must first beat seasonal naive on held-out data. If your neural network loses to 'copy last Tuesday,' you don't have a model — you have a science project.