That’s a great point. In my original work I used GAMs for energy consumption modeling and I also applied Fourier encoding to represent calendar features as cyclical variables. However, when I transitioned to LightGBM, I found that incorporating cyclical encoding didn’t significantly improve the model’s accuracy. It seems that GBMs handle these patterns effectively on their own, although I guess more tests would be required to substantiate this claim.
Great article Benedetto.
When using time features I tend to use cyclical encoding instead of raw features ("hour" and "dayofweek). Maybe this could improve the model.
Thanks, Johann!
That’s a great point. In my original work I used GAMs for energy consumption modeling and I also applied Fourier encoding to represent calendar features as cyclical variables. However, when I transitioned to LightGBM, I found that incorporating cyclical encoding didn’t significantly improve the model’s accuracy. It seems that GBMs handle these patterns effectively on their own, although I guess more tests would be required to substantiate this claim.