Hi there, I'm stacking two models to make my forecast. For example the first model only forecast from T to T+5 and the second one for T+6 to T+24. I have y that's available up to T-1. The reason I'm splitting them is I have some features that are very useful only up to T+5 but not beyond that and from T+6 to T+24 I have another group of features which is more stable, essentially two groups of features. So what I'm trying to figure out is I want to make the gap between T+5 and T+6 as smooth as possible, thus I update the second model with prediction generated from the first model, specifically T to T+5, but I'm still seeing "gaps", such as huge drops or increase in scale, from time to time. I look into the source code of

MLForecast.update

and it seems like all it does is update the stored ts. I'm wondering how does update even help in this case except I have

RollingMean(window_size=3)

in the feature frame, which I assume it got calculated based on T+3 to T+5 I fed in for T+6, and so on. Could you please confirm I'm understanding correctly? I would really appreciate if you have any other insights about how to "smoothly connect" two forecast.

hey. can you provide an example of what you're doing? do you have two mlforecast instances trained on the same data?

yes I have two different models trained on the same target but different feature frame. here's a pseudo code

fcst1 = MLForecast(models = {'modelA':lgb.LGBMRegressor(**lgb_params)})
fcst2 = MLForecast(models = {'modelB':lgb.LGBMRegressor(**lgb_params)},
            lag_transforms={  
                3: [RollingMean(window_size=3)],
            },)
fcst1.fit(df1)
fcst2.fit(df2)#y in df1 and df2 are the same
pred = fcst1.predict(h=5, X_df=cov1)
fcst2.update(pred)
fcst2.predict(h=19, X_df=cov2)

yeah that looks ok. so the problem is that the second model produces weird predictions?

yes especially the first index. It looks very different in scale whereas I am expecting it to be roughly the same as previous index, i.e. the last index in update.

the update method appends the new target values to the stored ones and updates the last training dates. the rolling mean looks at the last window_size samples of each serie, which after updating should be your predictions

yes that sounds aligning with my understanding.

are you using target transformations?

no. I'm not applying any transformation on target or features

can you try without exogenous features? just to see if they look better

sure