Are you really interested in ranking different indicators, or do you just want to know how you should combine them to make the best predictor possible? Is there any reason you can't use several together?
Correlation coefficients would certianly be a reasonable starting point for this. You have two obvious problems that come up if you do it this way:
- A correlation coefficient doesn't tell you how economically significant the relationship is, just how statistically significant. In other words, a certain variable might correlate very closely with returns, but the difference in the returns might be very small for different values of the variable. However, if you are comparing different predictors for the same population, there is a total amount of variability, and something with a higher correlation coefficient will explain more of it. But it might not be possible for you to realise all the information you have as gains. For example, being able to identify big losers and big winners could be much more valuable than sorting out the mass of small winners and small losers from each other.
- You will find that your indicators are correlated with each other. How do you measure the contribution each makes to determining valuation? Typically if you have two variables, both together will predict better than either one on its own. You might want to do various linear regressions with different sets of parameters and use some way of deciding which is the best. Correlation coefficients are closely related to single-variable regression, so this approach is kind of a generalization of using corrlation coefficients.
Calibrating in time t-1 and using the results to predict time t is also a reasonable approach. You should be careful of overfitting. This is when you develop a model which doesn't have any real content, but just happens by coincidence to match exactly how well something has done in t-1 and so doesn't work in any other period. You can take further periods and look for something which works pretty well over all of them to try and get round this. Or you could use various techniques like log-likelihood to measure how complex your model should optimally be.
If you really want to compare different indicators taken in isolation, paper trading could be a good way, although it makes more financial sense than statistical sense. Formulate a similar strategy for each indicator, eg. buy the best 10% and short the worst 10%, and see which would have made the most money after a certain time. This will force you to decide what time scale you are going to look for returns over (or to develop some other way of choosing when to close out positions), which is critical for statistical strategies.