Hi everyone,
I have a dataset. I want to predict a value (a regression problem)
As clear from the correlation matrix, the target (the variable I want to predict) has a strong relationship with independent variables (features). Thus, I think an easy model can predict the correct value.
But I tried a lot of regression models but it fails always.
The problem is: I want to predict a value, but the MSE, MAE, or RMSE are huge some times in thousands or millions.
Can anyone recommend me a model to run or discover the problem?
I can share the models for help.
Thanks for your time.
Hi!
Even if there is no standard size of error coefficients since they are highly dependent on the scale of your variable, it seems to be quite high.
Are your errors higher in your test data than in your train set? If yes, it looks like your model is severly overfitted.
May I ask how many observations you have for your regression?
Did you tried a penalizing regression? E.g. ridge regression if it's neccessary to take all independent variables into account or lasso regression if you can kick some variables out of your model.
You can try transforming both the independant variables and dependant variable, e.g log() or ln() transformation, and then run regression analysis in terms of the transformed variables.
Hi, Ahmed,
JIC you didn't check, perhaps a nonlinear model will fit better?
- https://odajournal.com/2019/01/19/when-to-evaluate-a-nonlinear-model/
Unfortunately, regression is well-known to be inaccurate, even including models which have a high associated r-squared value (this phenomenon is attributable to regression toward the mean). However, correlation and regression model accuracy can sometimes be improved a bit:
- https://odajournal.com/2013/09/20/maximizing-the-accuracy-of-multiple-regression-models-using-unioda-regression-away-from-the-mean/
Nevertheless, if you are able to, and would like to try a modern approach, there is a hands-down superior alternative (which literally maximizes predictive accuracy for the sample) compared to regression:
- https://odajournal.com/2016/09/19/novometric-analysis-with-ordered-class-variables-the-optimal-alternative-to-linear-regression-analysis/
Finally, here is a short review of novometric theory--which some (I) have described (just now) as being an intellectual tsunami which is heading toward academic beaches across the globe (all joking aside):
- https://odajournal.com/2020/06/28/what-is-novometric-data-analysis/
I feel that the huge RMSE, MSE are a result of a number of reasons like high correlation among the independent variables. I would suggest you go for step wise regression that is adding one variable at a time, and studying the regression. This will also remove the problem of multicollinearity.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression