Manual training of models for multi-step direct forecasting

[laxhammar]

Hi!

I'm experimenting with multi-step direct forecasting. MLForecast.fit supports this through setting max_horizon, which will train multiple models under the hood. However I'm trying to figure out how I can do something similar in case of manual feature engineering/training, ie when using MLForecast.preprocess? There is MLForecast.fit_models but it does not seem to support multi-step direct forecasting models by max_horizon or similar? Is there some other function I can use? Could anyone point me to an example/tutorial where this is done?

[jmoralez]

Hey @laxhammar, thanks for using mlforecast. It's possible to use MLForecast.fit_models, that method checks if the target is a matrix and if it is it trains one model for each column. Here's a minimal example:

from mlforecast import MLForecast
from mlforecast.utils import generate_series
from sklearn.linear_model import LinearRegression

series = generate_series(5)
fcst = MLForecast(
    models=[LinearRegression()],
    freq='D',
    lags=[1],
)
X, y = fcst.preprocess(series, max_horizon=4, return_X_y=True)
fcst.fit_models(X, y)
assert len(fcst.models_['LinearRegression']) == 4

[laxhammar]

Hi José, thanks a lot! This solved my problem!