JIm, Steve,
Be careful about trying to force a framework into an area to which it doesn’t apply.
In an ideal world with good data that always means exactyl what we think it means and one in which we can always define our models in the exact correct way, nobody would need, for example, more than one ratio to define value.
But that’s not even close to the world in which we operate. Our data is stupendously noisy so the chances of getting misleading numbers are extremely high. (What the data is saying: A P/E of 45.7, wow, don’t touch that overpriced piece of $@#!. What’s real: The company finally ditched a money-losing subsidiary, booked a one-time loss on the sale, and ex that, the P/E would be 11 vs. an industry median of 14.2, so that stock is potentially a good value play.)
Issues like this that hit a particular factor are analogous to “unsystematic (company-specific) stock risk.” You diversify. Portfolio to try to mitigate, as best one can the impact of unsystematic risk. Similarly, you diversify expressions of your factor ti mute the unsystematic risks associated with each item. If you have 15 value factors that all correlate with one another in the .97-1.00 range, so what! No problem. It simply means that your value model is telling you what you think it’s telling you. It means you will not, at least in this factor, get burned by having a mis-specified model. That’s way more important than allegiance to truisms from a discipline unrelated to investing.
The same idea applies to the mixing of ratios, PE, PS, EVS, P/FCF EV/EBITDA, etc., etc. etc. We don’t actually care about any single ratio. What we care about is that the stock be priced well relative to the merits of the company, and each ratio tells the same story but in a different way. The more broadly diversified a portfolio of value ratios, the better the chances that if you;re looking for value stocks, your ranking system will correctly point them out to you.
You’re still not out of the woods here. You also have to make sure that the value you’re seeing relates constructively to expected future growth and business risk, and each of those other two considerations involve their own sets of ratio portfolio construction issues, and actually, specifying those aspects of the model can be even more challenging.
P123 gives you a great set of tools you can use to address these challenges. Use the ones that feel right. Nobody has to use everything. But don’t handicap yourself at the outset by opening the toolbox dumping out most of the tools even before you get started, just because somebody else applying statistical solutions to different kinds of problems has no need for the kinds of tool sets that can and do help us (and might even find our toolsets overblown and inappropriate if forced into their own disciplines.
As to number of Buy/Screening rules, there’s no need for dogma. There can be a lot or a few depending on how strict the rules are. If the end result of the filtering gives you what you believe is a reasonable number of stocks to rank, then you’re fine. If you have too few, then loosen up your rules or eliminate some of them.
And please read the presentation that was cited at the top of this thread. I’ve had more time with it. I still reserve judgment on the machine learning topic, but what he says about econometric methods is spot on.