In R I created a gradient boosted random forest from 100,000 records with 10 cross validation folds using the gbm library. I want to communicate the strength and accuracy of the model, because it worked very well (far surpassing the random forest, kNN or SVM).

There are plenty of strength and accuracy tests (RMS, MAE, etc.) but the one I favor in the work place is correlation, cor(), because it has proven to be the most intuitive for non-stats people.

The correlation between the actual and predicted values was 0.80. I understand that correlation is a measure of linear association, but in this example, it is also a measure of predictive ability. If cor = 1, we are 100% accurate. But is it fair to say we are 80% accurate if cor = 0.80? Or if cor = -1, could we say it is -100% accurate; that doesn't seem right.

So my question is:

  • For predictive regression models, how are you communicating the strength and accuracy of the model to non-stats people?

The squared correlation $R^2$, which in your case would be $0.8^2 = 0.64$, has the interpretation of the proportion of variance in the response that is explained by the model. Note that what counts as a "good" R-square is context-dependent. A physicist might be unhappy with an $R^2$ of 98%, while someone working with social science, business or econometric data might be happy with 50%.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.