Subject: Re: LargeCapCash Questions
Two months later, sell all the picks you've held for four months
...
What is the magic of using 2 months and 4 months? Did someone also do a backtest to show that 2 months/4 months showed the best results? Maybe 2.3 months and 4.6 months would be better?


Two different types of answer

In general, the traditional MI screens have holding periods ranging from a month to a year. I've tested and used strategies down to 3 days. Many strategies test well at hold periods of a month or less, but a lot of the gain is either spurious luck in the backtest or disappears in the additional trading costs. Personally I don't trade more frequently than once every two months, for a variety of reasons. First, it's just too much work. Second, in real world results I don't think it hurts, because (secondarily) I don't have as much confidence in backtests with higher frequency trading. In particular, since most quant screens do not add value after they are created/tuned/published, frequent trading is just adding drag from trading costs. Even if a screen is totally random, you can generally expect to track an equally weighted portfolio within a statistical window...minus the trading costs. That's why we try to keep the total costs under 1%/year.

Another general answer: many (most?) screens work better as "dozens". For example, as here, divided up into N sub-portfolios each with hold period X months, with trading dates for each portfolio staggered by X/N months, generally ensuring that no stock is held in more than one sub-portfolio. This has fairly low turnover, while ensuring that each purchase you're making is done with the most recent available data. It seems to offer an advantage in backtests, I think for two reasons: keeping the recentness of the data for each new pick, and the fact that you're never doing a whole lot of trades all at once, which can be an issue if it happens to be right at a big turning point int he market (in direction, or leadership). Returns seem to be quite a lot smoother.

In the case of this specific screen, the rate of return seems to roll off for hold lengths longer than about four months, likely simply because of stale data (maybe something had a great ROE when you bought it but the business has deteriorated in the last 6 months...) So the likely suspects for trading would be holds of 1-4 months. I'm too lazy for monthly, so the choices would be 2-4 months. As mentioned above, two staggered portfolios of four month holds is generally preferable to a four month cycle. In backtest, you get even better performance with two sub-portfolios each with two month holds, with a trading day for alternating sub-portfolios each month, maybe 0.2%/year higher returns if the backtest is representative. And simple "top 40 every month" tests at another 0.25%/year better. Since this is really more of a "skewed indexing" investment approach, I don't think chasing the last fraction of performance is likely to be fruitful, and it involves a bunch of typing every single month. If the general strategy beats the S&P by 5%/year over time in real life (highly unlikely, but that's what the backtests show for all these variants with holds up to 4 months) then it is doing more than well enough.

As for whether 2.3 months or 4.6 months might be better, I don't know. First, the difference almost certainly be smaller than the already-large error bars on the tests. There are only so many picks to be had--for this screen, they change very slowly. And second, I don't currently use a backtester with the appropriate data that tests anything other than hold periods that are a multiple of a month : )
As mentioned above, I find anything under 2 months to be too much of a bother for no convincing additional upside, so I'm not really missing much. I can test 2 and 3, and they are extremely similar (difference smaller than the error bars), so 2.3 would be the same.

Jim