hey everyone, I'm trying to fit a LGBMRegressor with tweedie objective. The input data is of weekly sales in mlforecast format - unique_id, ds, y. Below are the features that I'm using but the prediction is always coming out to be the averaged value. Please suggest what set of features should I try here?

fcst = MLForecast(models=[LGBMRegressor(model params here)],

freq='W-Sun',

date_features=['week', 'month', 'quarter', 'year'],

lags=[1, 4, 8, 12, 52],

lag_transforms={

1: [(rolling_mean,4), (rolling_min, 4), (rolling_max, 4), (rolling_std , 4),

(rolling_mean, 8), (rolling_min,8) ,(rolling_max, 8), (rolling_std, 8),

(seasonal_rolling_mean, 12, 4),(seasonal_rolling_min, 12, 4),(seasonal_rolling_max, 12, 4),

],

},

)

Below is the plot which shows all the predictions are same. But it seems that there is a pattern in the data.

Try setting

dropna=False

in the fit call, those lag transformations will produce many NaNs and the model will be left too few samples (you can see that by running preprocess first)

Thanks José. Comparatively better than before but still the error is high. Based on my limited understanding, I provided week, moth, lag 52 & some rolling means. But the fit is still not that good. Any suggestions on the features side as what features should I use given the plot?

Is that your only serie or is that a sample? If you don't have that many observations try removing the features that require a lot of samples (seasonal rolling) and setting min_samples=1 for the rest.