How can I use MAPIE with statsmodels?

[ilkot]

MAPIE currently works with scikit-learn-compatible estimators, which is great, but this also makes it difficult to use directly with statsmodels forecasting models such as the ARIMA family or ExponentialSmoothing.

It would be very helpful if MAPIE could include an example demonstrating how to use it with a statsmodels wrapper.

If there is already a recommended workaround for making MAPIE compatible with statsmodels, I would be very interested to hear about it.

[GBrelurut]

Hello @ilkot

Thank you for your feedback. We don't currently have functionnalities to handle statsmodels built into Mapie. The best way to use such models is to implement a wrapper before providing it to Mapie. It may look like this :

# can inherit scikit learn BaseModel
class StatsmodelsWrapper:
    
    def __init__(self, model):
        self.model = model
        self.results = None

    def fit(self, X, y, **kwargs):
         self.results = self.model.fit(X, y, **kwargs)

    def predict(self, X, **predict_params):
         if self.results is None : 
             raise AttributeError('Model is not fitted.')
        return self.results.predict(X, **predict_params)

We are currently framing a rework of Mapie backend, and will do our best to ensure this workaround is not necessary in future versions of the package.