Sliding Window Backtest

[ajayganti3]

In Backtest we have 2 strategies named Expanding window and Sliding Window. As per my understanding currently Darts support Expanding window backtest. If Sliding window backtest is implemented as a new feature, then it would be really helpful because expanding window backtest over a 100K timeseries would be too much of compute intensive when compared to Sliding window backtest. Sliding window visual representation is given in the below image.

[pennfranc]

Hi @ajayganti3, thanks for your suggestion, this addition is definitely on our radar! For now, to speed-up your backtesting, you can also consider setting retrain=False in case you are using TorchForecastingModel instances. You can also increase the stride parameter for some linear speed-ups. But a sliding window would definitely be nice, we agree. Please stay tuned!

[pabuta88]

Any news here regarding this feature? It would be really nice if you could add this.

[shagn]

Hi @pabuta88, you can set the parameter train_length and then the moving window strategy is used. Quoting the documentation: "Default is set to train_length=None where it takes all available time steps up until prediction time, otherwise the moving window strategy is used" (source: StatsForecastAutoARIMA documentation)

[hrzn]

Closing as this is now implemented.