Interactive technical demonstration
A forecasting model should have to beat an honest baseline.
This browser-based lab creates deterministic synthetic series and backtests three simple methods. It demonstrates evaluation mechanics; it is not a claim about client data or production accuracy.
Naive: repeat the last observed value.
Seasonal naive: repeat the value from 12 periods earlier.
Moving average: average the previous 3 values.
Ready to run
Select a pattern and compare the baselines.
The lab backtests each method on the selected final horizon and reports mean absolute error. Lower MAE is better for this demonstration, but real projects must choose metrics based on the decision and cost of errors.
Interpretation
The winning baseline depends on the structure of the series.
A seasonal baseline may outperform a recent-value baseline when the cycle is stable. A moving average may reduce noise but lag a trend. Intermittent demand can make all three weak. That is the point: complexity earns attention only after the evaluation reflects the real operating decision.
What a real study adds
Context the synthetic lab cannot supply.
- Data rights and lineage
- Decision horizon
- Error costs
- External drivers
- Segment behavior
- Operational adoption
The generator is deliberately simple, contains no probabilistic intervals, and uses one deterministic pattern per option. MAE alone does not capture bias, service levels, asymmetric costs, structural breaks, or whether a forecast improves a real decision.