FlyingCircus’s reply is pretty much the way I see it. Back in the early 2000’s simple screen with 2 or 3 fundamental factors with a momentum kicker worked reasonably well. For the last 10 years very, few screens have performed well with enough consistency for one to remain confident enough to continue.

However way back in 2005 Dan Parquette posted an Excell spread sheet of with a 20-bin histogram of both fundamental and factor performance. I borrowed his concept and made my own versions for both P123’s data and Stock Investor Pro data. The performance of many individual factors still work over a long period but not consistently over time. Simple one less/greater than one branch of a decision tree stopped working as the competition started using ranking systems. You can build ranking systems on GTR1 but it was easier on P123’s site. Ranking systems give you the advantage of combining a number of individual factors that all perform a little better than average increasing your probability of success.

Unfortunately, ranking systems that actually work and continue to work are very difficult to put together. If one kept track of all the individuals on the P123 site that offered users access to their great backtested outperforming models you would observe that very few outperformed post discoveries. The core group of individuals that have continually outperformed as far as I can tell are not weekend warriors they are very analytical statistically focused individuals.

The more current trend is using machine learning. Which in essence is just an automated way of making a more sophisticated ranking system. I’ve spent the much more than time than I thought it would take going down the ML path. At this point the results seem very promising but I can already see the professional systems I have become aware of will soon arbitrage my temporary advantages.