Let’s say I run a screen that chooses only stocks in the health care sector with a low price-to-book ratio, a low avgrec, and low short interest. The performance of these stocks, rebalanced every 4 weeks, is terrific over the last five years. Let’s say I run another screen that chooses only stocks with absolutely no sales, negative eps, negative cash flow, and return on investment and return on assets less than zero. The performance of these stocks, rebalanced every 4 weeks, is dismal. Yet almost every stock I’ve bought in the first basket is in the second basket too.
This makes me question the whole logic of screening stocks. Everything you’re looking for in a stock screener is what makes the stock look good. You’re not taking into account all the things that makes a stock tank.
So now I’m thinking: maybe I should abandon stock screening and just use P123 to EVALUATE stocks on the basis of, say, 20 different criteria: sector, momentum criteria, size criteria, value criteria, sentiment criteria, quality criteria, and growth criteria. To do this I’d have to run, say, five different screens with four variables each so that I can see how a particular stock with those specific values would have performed. An extremely time-consuming task indeed, though nobody ever said that wise investment was anything but.
On the other hand, a portfolio of stocks that I screened for using the health care criteria above would actually have performed well in the past, and putting, say, 20% of my money in it might not be such a bad idea.
Here’s an analogy. You have $240 to spend on 20 special lottery tickets for $12 each. There are 100 lottery tickets available, each one composed of four digits, and you can decide which ones you want. You are told that your winnings will be the sum of the four digits on your ticket except that 7s, 8s, and 9s count as -1, -2, and -3 respectively. So if you choose ticket #6666 you will win $24, if you choose ticket #0000 or #9999 you will get nothing, and if you choose #5678 you will get $8 back on your $12 investment.
The screener will choose all the tickets with the number 6 in them, no matter what. The presence of that number is better–a whole lot better–than the presence of any other number. It’s a winning strategy, but a lot of his choices will lose.
The evaluator will add up the numbers on every single damn ticket to see which ones make the most money. He will choose lots of tickets without 6s in them if they have 4s or 5s. He’ll even choose #3334 if it’s there. And if there are enough tickets in the 100 that add up to over 12, he’ll make money on every single one.
He couldn’t apply a ranking algorithm because then he’d end up with either lots of low numbers or lots of high numbers, both of which would lose. And unfortunately, that’s the major flaw of ranking systems in the stock market. Take the factor eps%changettm, for example. The highest rank should be given not to those stocks that with an EPS change of 80% or higher. Those stocks almost invariably lose money. The highest rank should be given to those whose growth is reasonable and sustainable. The same is true for a lot of other factors, including price to cash flow and EV/EBITDA. You want to avoid the extremes.
The flaw in your logic is your statement that “You’re not taking into account all the things that makes a stock tank.”
Nothing can take into account ALL of anything. But a good screen will (1) articulate a strategy and (2) guard against ways in which the strategy can go wrong.
A simple example would be a value screen. Articulating good value tests is easy. Anyone can do it; low P or Ev to whatever. But if the model stops there, you could very easily wind up with a lot of dogs. So it’s critical that you pair the value tests with others designed to filter out stocks that are cheap because they deserve to be cheap.
Another example is income. The yield threshold is easily expressed. The successor failure of the screen will depend on effective use of rules that weed out companies with a high probability of cutting their dividends. (And such a rule may even be in the form of Yield <x based on the notion that a too-high yield suggests the market sees a very high probability of a dividend cut and poor share performance.
Screening is incredibly powerful assuming it’s done thoughtfully. If it’s simply a grab bag of things that seem like they should work, then that’s a prescription for bad performance.
I’d even go so far as to say that screening without ranking has a better chance of succeeding than ranking without screening. The latter pushes the rules of probability very hard, perhaps too hard. It’s one thing to say stocks in the top decile for such-and-such characteristics are likely to be best. It’s quite another to move from that, in which a decile may be several hundred stocks, to say the top 10 stocks will be best. Settling on a top 10 can be dramatically enhanced by use of a screen to pre-qualify the universe that will be ranked, much the way a good sales person buys a list of qualified leads and prospects from that, rather than a phone book.
This is very helpful. Thank you. But how do you deal with the first contradiction in my query? The contradiction between the two screeners–one for health care stocks with low p/b ratios, low short interest, and low avgrec–that one performs superbly when backtested–and another with negative ROI, ROA, EPS, etc., which performs terribly, and contains almost all the stocks in the first screener?
What I’m saying is that theoretically every stock will turn up as a result on a large number of possible screens, and the performance of those screens will vary from terrible to terrific. So picking a stock based on its appearance in ONE screen alone is like betting on a sports team on the basis of one person’s recommendation.
Hard for me to reply precisely without seeing the screens, but I wonder if you relied too much on the backtests (WHAT worked or didn’t work) versus logic (WHY things work or don’t work). Low PB is not the best valuation metric around any more, and would seem especially inappropriate for healthcare, which includes biotech stocks that are less likely than others to trade on PB. And AvgRec is widely regarded asa a poor indicator of future share performance, almost to the point where many might use it in a contrarian way); change in AvgRec has more potential, but estimate change would probably be better. Low short interest, you can argue equally in both directions from that; you can treat it as bullish (evidence or sentiment), or you can say high short interest is bullish (possible squeeze; the beards have already sold). On the pother hand, really bad ROI and earnings trends are well recognized as likely to be associated with bad stock performance. So it looks to me like your bearish screen was more validly built than the bullish screen, and that you goit the results that were to have been expected.
You can’t discover a strategy through the backtest. The purpose of the backtest is to give you feedback on the effectiveness of the way you implemented a strategy you have reason (even without seeing a test) to expect to be effective. If you check Help >> Recommended Reading, you’ll find some suggestions re: books that can help you get a handle on strategy building.
Most of my screens are far more traditional than that, using price-to-sales, forward p/e, ev/ebitda, ROA, etc. But this one was eye-opening. There is, actually, a good reason for it to work. The avgrec field, in general, produces unreliable results, but in biotech, and health care in general, the results are excellent, even if there are only a few analysts covering the stock. I look for stocks with an avgrec of 1.3 or lower. Low p/b is basically the only way to obtain a low price for a company that has produced no earnings yet–I use 1.25 or 1.5 as my upper limit–that way you’re sure not to lose a bundle if the company goes under. And low short interest (under 2%) is just to make sure there’s not a significant number of doubters. There’s someone on Seeking Alpha who looks for high-quality biotech stocks with p/bs of 1 or thereabouts.
I’ll be sure to check out those books. Maybe they’ll change the way I develop screens.
Yes, that’s often a very effective choice. And lately, I’ve been working with EV/SalesTTM, which gets at the same thing but is better for dealing with comparisons among companies with differing debt burdens.
Sorry, I meant to say companies with no SALES. Most of these biotech companies are still in the R&D phase. The only way to get a value multiple on these kinds of firms is by using p/b.