Much attention is now paid to what models to use for time series forecasting, but not to how exactly to perform training and inference.
Tsururu is a Python-based library which aims at overcoming the aforementioned problems and provides a large number of multi-series and multi-point-ahead prediction strategies that can be used with any underlying model, including neural networks.
Also tsururu provides various preprocessing techniques.
- Local-modelling:
- Individual model for each time series.
- Global-modelling:
- One model for all time series;
- features made up of individual series do not overlap.
- Multivariate-modelling:
- One model for all time series;
- features made up of individual series corresponding to the same time point are concatenated for all time series.
- Recursive:
- one model for all points of the forecast horizon;
- training: the model is trained to predict one point ahead;
- prediction: a prediction is iteratively made one point ahead, and then this prediction is used to further shape the features in the test data.
- Recursive-reduced:
- one model for all points in the prediction horizon;
- training: the model is trained to predict one point ahead;
- prediction: features are generated for all test observations at once, unavailable values are replaced by NaN.
- Direct:
- individual models for each point in the prediction horizon.
- MultiOutput (MIMO - Multi-input-multi-output):
- one model that learns to predict the entire prediction horizon.
- Also, this strategy supports the presence of exogenous features (only for local- or global-modelling).
- FlatWideMIMO:.
- mixture of Direct and MIMO, fit one model, but uses deployed over horizon Direct's features.
To install tsururu via pip you can use:
pip install -U tsururu
For usage example please see:
- Tutorial_1_Quick_Start for simple usage examples
More examples are coming soon.