I am currently working on a personal project for predicting financial equities movements into 3 classes. The three classes are 1.) above a certain percentage move within a certain timeframe, 2.) below a certain percentage move within a certain timeframe, or 3.) a move that is in between the two previous classes (AKA don't care). For my purposes I only care about the first 2 classes, for the time frames that I have calculated on my dataset the ratio of class 3 to each class 1/2 is about 10 to 1. I have tried using class weights and it gave me strange results in that there were more predictions in classes 1&2 but they were much less accurate, versus with no class_weights I got more accurate results, but with much more false positives in class 3 and less inferences in classes 1&2. What other methods should I use such that the true ratio of the minority classes is represented in the model and accurately? Thanks!
Asked
Active
Viewed 46 times
0
-
1Obligatory reference Why is accuracy not the best measure for assessing classification models?. – user2974951 Oct 05 '23 at 05:29
-
Thank you, yes I am using precision as my metric for my purposes. – Victor Minin Oct 05 '23 at 22:06
-
The point of my post was to try and dissuade you from using such metrics, because of reasons described in the post. – user2974951 Oct 06 '23 at 06:30
-
Second important reference https://stats.stackexchange.com/questions/247871/what-is-the-root-cause-of-the-class-imbalance-problem – picky_porpoise Oct 07 '23 at 06:38