sktime_lgbm.py

from sktime.forecasting.compose import make_reduction, TransformedTargetForecaster
from sktime.forecasting.model_selection import ExpandingWindowSplitter, ForecastingGridSearchCV
from sktime.performance_metrics.forecasting import MeanAbsolutePercentageError
import lightgbm as lgb

def create_forecaster():

    # creating forecaster with LightGBM
    regressor = lgb.LGBMRegressor()
    forecaster = make_reduction(regressor, window_length=5, strategy="recursive")

    return forecaster


def grid_serch_forecaster(train, test, forecaster, param_grid):

    # Grid search on window_length
    cv = ExpandingWindowSplitter(initial_window=int(len(train) * 0.7))
    gscv = ForecastingGridSearchCV(
        forecaster, strategy="refit", cv=cv, param_grid=param_grid, 
        scoring=MeanAbsolutePercentageError(symmetric=True)
    )
    gscv.fit(train)
    print(f"best params: {gscv.best_params_}")

    # forecasting
    fh = np.arange(len(test)) + 1
    y_pred = gscv.predict(fh=fh)
    mae, mape = plot_forecast(train, test, y_pred)

    return mae, mape


param_grid = {
    "window_length": [5, 10, 15, 20, 25, 30] # parameter set to be grid searched
}
forecaster = create_forecaster()
sun_lgb_mae, sun_lgb_mape = grid_serch_forecaster(
    sun_train, sun_test, forecaster, param_grid
)

@hiteshgupta2507

I have been following your online material but couldn't find anywhere freq being explicitly being fed into input series.

Only the WPI data have that conversion in that blog post.

I was looking at the code nile data set code.

Nile dataset is not indexed with datetime values. It just has numeric values of years, hence it does not require frequency on its index.