Analytics · honest forecasting

demand-forecast-pipeline

A point forecast hides the decision risk. This pipeline reports what the backtest missed and turns those misses into empirical intervals.

Public · synthetic demo

The problem

A single accuracy score cannot show how uncertainty changes by horizon.

Choosing a model on one split can reward luck. Assuming Gaussian error can make an interval look cleaner than the backtest evidence supports.

Input

A synthetic daily series with rolling-origin evaluation.

Main series: 730 observations
Forecast horizon: 14
Seasonal period: 7
Minimum training window: 60
Backtest stride: 7
Nominal interval target: 80%

The money shot

Backtest evidence selects the model and sets the interval width.

Chosen model sMAPE

3.094

Fold/horizon points

1333

Main-series interval width

26.27

Short-series interval width

46.75

The selected additive Holt-Winters baseline had reference MAE 9.042. The short, erratic series had 21 observations and a wider average interval than the 730-observation main series.

How it's verified

Every candidate sees the same rolling origins.

Model	Backtest sMAPE	Backtest MAE
holt_winters_additive	3.09	9.042
seasonal_naive	3.45	10.012
moving_average	10.42	31.426

The first forecast is 313.03 with an empirical 80% interval of 301.97–323.49, built from 96 backtest residuals.

Honest limitations

Backtest-shaped uncertainty can still fail after a regime change.

The candidates are simple baselines with fixed smoothing constants.
The intervals reuse backtest residuals; coverage is a calibration check, not independent validation.
No promotions, weather, price, missing-data handling, or outlier treatment enter the model.
The data is synthetic, and the future can differ from every rolling origin in the fixture.

The interval is evidence from past misses, not a guarantee about the future. jigonyoo.com · Back to hub