Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hello everyone,

If I were to use `tspreprocess` and `tsfeatures` through the CLI, does that mean I can only predict using `tsforecast`? Or is it possible to use those 2 preprocessing steps and then use your `statsforecast` / `hierarchicalforecast` library?

Thanks.

Hi <@U03S7N47XRC>,

We have used the <https://github.com/Nixtla/tsfeatures|tsfeatures> library to transform time series to serve as input in decision tree-based models like the one from the M5 competition and the M4 second place the Hyndman's FFORMA.

<https://www.sciencedirect.com/science/article/abs/pii/S0169207019300895|FFORMA>: Fits and predicts several simple models (like the ones from StatsForecast) and then ensembles them using features and an XGBoost Model.

<https://github.com/Nixtla/mlforecast|M5 Winner>: Transforms series and fits LGBM with Tweedie Loss specialized on sparse series (Keep in mind that M5 results are inconclusive due to the test set trend breaks).

As such, our experience with tsfeatures is to be a support library for a bit more complex decision tree models. Any of the models mentioned are compatible with the reconciliation strategies of the <https://github.com/Nixtla/hierarchical|HierarchicalForecast> library.

If you have other questions, please reach us.