Thank you for sharing. So for a given day let say day t, you are using lags from day t-1, day t-2,....., to day 1.

I also wrap around, so for day t I use the previous t-1 precipitations and the following 730 - t ones, as well as the precipitation of the day itself obviously

Very interesting, mine also works very well with lightgbm, in general I see classifiers work better than regressors. So far I am still trying to combine lightgbm and MLPClassifier. It seems that both models work quite well.

Sounds Great. Are you applying anything to the data, any tips?

I am using boosting models and I have tried placing lags as you said. I have tried winsorization, i have tried removing outliers, I have tried groupkfold, I have tried stratifiedkfolds, I did extensive feature engineering and still no significant boost. Am I missing something here?

Sorry to hear bro but the basic setup is really simple, just create lags for each day (I used cycling but also padding yields the same results) of each time series and binary classify each day, no particular preprocessing as tree models are not sensitive to data range.

As for the split, for each time series id, assign 1 if it contains a flood and 0 otherwise, then split the dataset so that both training and validation have the same percentage of time series with floods. Nothing else

Worked like magic ! I'm really grateful.