/sci/ - Science & Math » Thread #14191740

36KiB, 1172x847, file.png

View Same Google iqdb SauceNAO

Regression with weird distribution

Anonymous Sat 05 Feb 23:13:45 2022 No.14191740 View Reply Original Report

Quoted By: >>14191758 >>14191851

Hi /sci/, I'm trying to predict a variable that has a weird distribution and when I run OLS I get a set of significant predictors but low R^2. (picrel, it's returns for a highly volatile asset).

I've tried doing a log-log regression but that didn't work. My independent variables are mainly either categorical or highly skewed (exponential with many observations being 0). What should I do?

Anonymous

Anonymous Sat 05 Feb 2022 23:17:49 No.14191758 Report

Quoted By: >>14191767

>>14191740
Who even thinks about regression in this day and age? Just train a neural net on it lmao.

Anonymous

Anonymous Sat 05 Feb 2022 23:20:22 No.14191767 Report

Quoted By: >>14191795

>>14191758
I mean I can go train something on it in sklearn but the dinosaurs at the No-Fun-In-Econ committee won't like that, right?

Anonymous

Anonymous Sat 05 Feb 2022 23:26:26 No.14191795 Report

Quoted By: >>14191865

>>14191767
But it's no different from any other parametric model you would fit to the data, and, in my experience at least, a lazy NN with hyperparameters pulled out of my ass tends to beat all other methods of regression/classification.

On a more serious note, what exactly are you trying to achieve here? If the distribution is really complicated and you don't have enough data, it's a GIGO situation.

Anonymous

Anonymous Sat 05 Feb 2022 23:42:00 No.14191851 Report

Quoted By: >>14191865

>>14191740

random forest that shit

Anonymous

Anonymous Sat 05 Feb 2022 23:45:38 No.14191865 Report

Quoted By: >>14192100

>>14191795
I appreciate the help.

I'm basically trying to see whether company fundamentals predict excess returns over a certain period. This is stuff like profitability (binary) and number of employees (exponential/power law). So when I run the regressions I get that, for instance, profitability has a significant and large impact on excess returns. The model has a low R^2 . Now ultimately I'm not surprised that's the case and don't really care all that much because, well, excess returns are very hard to predict and their variance is unlikely to be tied to fundamentals like that. But I was wondering how I could juice the regression as far as I can without overfitting to patterns that aren't there.

>>14191851
thanks for the hint frien

Anonymous

Anonymous Sun 06 Feb 2022 01:00:14 No.14192100 Report

Quoted By:

>>14191865
If the simple question is how much your independent variables provide systemic information (i.e. not noise) about excess returns (how well can you make future predictions using the same variables) then it seems to me that requires cross-validation.

I'm not sure what you mean when you say log-log regression didn't work? Why didn't it work? But if the distributions are really giving you trouble then I agree with the above post that random forests are a good shout as they don't require scaling of any form. They will also provide a good level of interpretability of predictor importance as with OLS. You might even try a single cross-validated decision tree if you need to pitch to non-technical audience as an RF might not be necessary for the data (although I don't know).

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches