search

mthorrell / gbnet

Gradient Boosting Modules for pytorch
MIT License
11 stars 0 forks source link

added forecasting #20

Closed mthorrell closed 2 months ago

mthorrell commented 2 months ago

I figured that a simple method of forecasting was to just fit trend and seasonality (not possible with base xgb, but possible with xgb and backfitting if you really didn't want to use torch). I started with that and got pretty far in comparison to Prophet on Prophet's example datasets. I added some additional things to accommodate changepoints in a pretty ad-hoc way and now this seems to out-perform Prophet on all of the example datasets in the Prophet repo.

This PR also introduces the models/ directory which will contain different models or applications for gboost_module.

mthorrell commented 2 months ago

Putting some numbers here for posterity:

image

Long range forecast is train on X% of the data, predict on the rest. For each dataset we test X = [10,20,30,40,50,60,70,80,90]. These results are the gboost_module win-rate across all comparisons.

mthorrell commented 2 months ago

Worth also noting that in the example_pedestrians_covid case, one of Prophet's "wins" looks like this. IF we call this a win for gboost_module, then 9/10 datasets favor this gboost_module forecasting on the long-range forecast comparison. image

mthorrell commented 2 months ago

Results for short-range forecasts (forecast forward 10% of the data) all favor gbmodule forecasting:

image