2

I understand the answer to this question is that it entirely depends on the data set.

However, this does not help people understand if their model is suitable or whether they should explore other variables.

So I am attempting to getting this question answered by people with experience in different fields where statistics is applied.

Assuming the model was not overfit and no other problems existed with the model, what I would like to know from someone experienced is the vastly different projects they worked on from very different industries and R^2 they obtained from their models so that people may gain a 'feel' whether they are heading in the right direction.

This question is of a highly atypical format so let me know if this is inappropriate for this website and I will remove it. Any suggestions on where to post this will also be appreciated.

Ali Turab Lotia
  • 589
  • 1
  • 6
  • 10
  • As you already noticed "it depends on data" (see similar case http://stats.stackexchange.com/questions/132536/how-to-choose-a-confidence-level/132538#132538). – Tim Aug 30 '16 at 12:45
  • In time-series, I get 0.90. In panel data, I get 0.20. $R^2$ is a terrible metric and really should be eliminated from software, as it only mislead people. – luchonacho Aug 30 '16 at 13:29
  • 5
    @luchonacho If you eliminate anything from statistical software that could seriously mislead the naive or the ignorant, there would be nothing left. Essentially all inferential results and even means and SDs would be high on the list. $R^2$ makes good sense for simple models provided you look at data too. It is just not a universal, omnibus, factotum, portmanteau statistic that encapsulates model merit, but nothing else is either. – Nick Cox Aug 30 '16 at 13:56

1 Answers1

0

I recently created a model for the amount of health insurance claims made by an individual on the basis of age, gender etc. and used a Gaussian family distribution.

The $R^2$ I got was 30.58% which I believe to be good considering how random the amount a person spends (given the person has no pre-existing condition, since those are not covered) on health insurance is.

I recently read a post where an $R^2$ of 7% on stock market prediction resulted in millions of dollars of additional profit. The stock market is truly volatile and most affected by current events that typically happen randomly. Therefore, an $R^2$ for a statistical model based on the stock market would be low while still being useful.

Nick Cox
  • 48,377
  • 8
  • 110
  • 156
Ali Turab Lotia
  • 589
  • 1
  • 6
  • 10