I'm building a new multi-formula Ranking System and am curious how to get the most out of each formula, and how to discern which formulas
I should keep and which to discard. I'm wondering if testing each formula (across multiple timeframes) for Alpha, Sharpe, Spearman/Pearson
correlations is beneficial. It's a lot of effort to do so. If you believe there is sufficient value in doing so, what are the best methods of testing for Alpha?
I've been using the Screener->Backtest and the Ranking->Performance test. Using only P123 functionality, are there more productive and accurate approaches?
What I have done is create a separate ranking system for each factor (one for higher is better and lower is better). I arranged them according to style. Then I create a simple simulation where, for example, I keep the top 5 or 10% of the parent universe (top 50 stocks in S&P 500).
From here, I run optimizer. This allows you to change one variable at a time. So I then load in 50 different ranking systems and let it go for 20 minutes. I take note of any factors that perform well, write them down, and load it with the next 50 factors. Typically I start with the stock P123 style factors so I know quickly which styles to focus on before getting into individual factors.
After I have maybe 12 - 15 factors that work well, I then open the individual ranking system and test them one at a time to see if the factor works across all the sub-portfolios or just in the highest ranked portfolio. I really want to see it work in all of them and the sub-ports look like an ascending staircase. I adjust the universe and holding periods to test it out.
From here, you can start to mix factors together to see which ones have synergy. Does value/momentum work? Or sentiment and FCF yield?
There is more but this is how I sometimes start to get the gears turning in my head.
Kurtis Hemmerling
1 Like
Thanks, Kurtis, for your reply. Am I to understand correctly that if you have, for example, 100 factors to test, you'll create 200 ranking systems (one each for a factor ranked higher and one for ranked lower). And that's your basis for the Optimizer analysis? And then onto refinement of the factor list.
Yes. That's just some initial work for factor discovery. It took a long time to make all those ranking systems initially but it really helps when I have a simulation with rules I like and I can test a variety of single-factor ranks on it.
2 Likes