Taken from J. Scott Armstrong, 1985, Long-Range Forecasting, 2nd ed., p. 487.
Some people are impressed by a high R2. There are simple things that can be done to raise R2; other than that, they are of no value. The most important thing is not to use these rules (please don't), but to be aware that others use them.
- Discard outliers after you examine the regression results.
- Aggregate the data, especially when it reduces sample size significantly.
- Experiment by trying lots of variables.
- Try different functional forms.
- Use stepwise regression and retain all coefficients with t statistics greater than 1.0 (Haitovsky, 1969).
- Include a lot of variables in the final equation.
- Use R2 rather than .
These rules should yield R2 values of over 99% for time series data and about 90% for cross-sectional data. My advice is that you should not use R2 for time series data. The dangers outweigh any potential benefits.