Analysis
Predict
Jalaran Predict forecasts a time series with linear, Holt-Winters, ARIMA and naive models compared on a held-out test set — with the naive baseline always shown.
Who Predict is for
For anyone forecasting demand, traffic, cost or any other series over time, and particularly for anyone who has presented a sophisticated forecast without checking whether it beat repeating the last value.
What Predict does
Predict fits several model families and compares them on data held out from fitting, which is the only comparison that means anything — a model evaluated on data it was fitted to will always look good. The naive baseline is always displayed, never hidden, because a forecast that cannot beat "last value repeated" is important to know about and is exactly what an impressive-looking chart conceals. Forecast intervals are shown, since a point forecast without uncertainty is a guess in a suit.
- Linear, Holt-Winters, ARIMA and naive models
- Held-out test comparison
- Naive baseline always visible
- Forecast interval visualisation
How Predict works
Load the series
Your time series, with its time index.
Fit the model families
Linear, Holt-Winters, ARIMA and naive, all fitted together for comparison.
Compare on held-out data
Evaluation on data the models never saw. Anything else measures memorisation.
Check against the naive baseline
Always shown. If nothing beats it, that is the finding, and it is worth having.
What Predict does not do
Predict forecasts from history and cannot know about anything that has not happened before — a regime change, a new competitor, a policy shift. Longer horizons degrade quickly and the intervals widen honestly to show it. It handles univariate series: no external regressors, no multivariate forecasting, no causal inference. A good fit on history is not a guarantee about the future, and no forecasting tool can make it one.
Common questions
Why is the naive baseline always shown?
Because a lot of sophisticated forecasts do not beat repeating the last value, and an impressive chart hides that completely. Knowing your model adds nothing is more useful than not knowing.
How far ahead can it forecast?
Accuracy degrades quickly with horizon, and the intervals widen to show it. Forecasting from history cannot anticipate anything unprecedented, however good the fit looks.
Can I include external variables?
No. It handles univariate series only — no external regressors, multivariate forecasting or causal inference.
Explore the workspace
Jalaran is one workspace of 85 modules. Browse the rest of the arsenal: