Models

Model types (Time series / Tabular)

Fold fundamentally supports both:

"Time series" models

The likes of ARIMA, RNNs, Exponential Smoothing, etc.

Their univariate variations only have access to y, and ignore all data in X. They're usually designed to be effective without additional feature engineering.

Examples:

• StatsForecast
• NeuralForecast
• Prophet

... provided in fold-wrappers.

Tabular ML models

The likes of Random Forests, Gradient Boosted Trees, Linear Regression, etc.

They depend on having X populated, and do not work as "univariate" models. Each row in X corresponds to a single dependent variable, in y.

Usually, you may want to add lagged values of y with the [AddLagsY][fold.transformations.lags.AddLagsY] class, or create other features for the tabular models with:

• AddLagsX: if you have exogenous data already.
• AddWindowFeatures: if you have exogenous data already, and you want to aggregate them across different windows.

Examples:

• Scikit-Learn
• XGBoost
• LightGBM

... provided in fold-wrappers.

Check out the Examples gallery to see how easy it is to engineer features with fold.

Online and Mini-batch Learning Modes

A mini-batch model is retrained for every split the Splitter returns. It can not update its state within a test window, but it may depend on lagged values of X or y.

An online model, on the other hand is updated after inference on each timestamp. Except for "in sample" predictions, which is done in a batch manner, with predict_in_sample()

We also give our "online" models a way to access the latest values and skip the step that'd update their parameters. This enables an efficient "quasi-online" behaviour, where the model is only re-trained (or, updated) once per fold, but can "follow" the time series data - which usually comes with signifcant increase in accuracy.

Baselines

As Time Series is a fundamentally hard problem, it's also important to use strong baselines, which we have our own, fast implementations, in fold-models.