RankNet with LightGBM

P123 is all about rank and for good reason. Even when we use ML we sort the predictions to create ranks. We can do that directly, perhaps. Bypassing MSE and a bunch of metrics we don't really care about.

Has anyone looked into RankNet using LightGBM?

Edit: Thank you for the suggestion ZGWZ. I will definitely look at that.

from lightgbm import LGBMRanker
ranker = LGBMRanker(objective="lambdarank")

I think I will look at this. Claude 3: "generally, it's considered relatively efficient compared to many other machine learning algorithms, especially for ranking tasks. "

Edit: Thank you ZGWZ. I will definitely look into that.

Update. With ChatGPT and Claude 3 I was able to try the LightGBM RankNet version with a walk-forward backtest.

Short story: not an improvement. But amazing how quick the turnaround on an idea can be now with LLMs helping with the coding.

2 Likes