Factor Momentum

[yuvaltaylor]

If “factor momentum” means “invest using the strategy that has worked the best over the last 12 months” then the concept is not only counterintuitive but dangerous. Any study of mutual funds will show that strategies tend to mean revert over such a short period, and any backtest of a variety of different strategies using Portfolio123 will show the same.

…

But a long-term study of such a strategy would show that it would normally have been a disaster. And then, if you switch your factor directions and weights every month depending on what has worked for the twelve previous months, you’ll never hold onto a stock for more than a month. Is that what you want?

I have friends who have made similar arguments against factor investing. It can’t work, they say. How can you do better than Morgan Stanley, GS and the tens of thousands of big brains trying to beat the market unsuccessfully?

It’s like the old joke. An economist is walking down the street and sees a $100 bill on the floor. He doesn’t pick it up. Because it can’t be there. Someone else would have gotten it first.

Well, it’s a good question, but this stuff is working–at least for some of us.

There was clearly some factor momentum over the past twelve months, as high volatility, cheap value, and high PSR continued to outperform after the covid crash. And there was some factor momentum from the 2018 correction until the covid crash, with high quality large growth outperforming consistently. I actually predicted both periods in advance. At this point in the market cycle however, I expect the first bounce might be almost done. After that, the market will probably soon start to switch to a new set of factors (or type of stock) that it favors. This for me is the toughest part of the market cycle to predict which factors will start outperforming.

In general, how can you tell which factor is set to outperform? There are numerous indicators. Some are based on where you are in the market cycle. Some are based on factor mean reversion. Some are based on long term factor success. There are also valuation indicators, macroeconomic indicators etc. Some are more predictive and some are less predictive.

There is literature about this, for example, predicting momentum crashes.

But as always, the best way is to create hypothesis and set out to prove or disprove it. To make this kind of testing feasible I highly suggest that you think about allowing dynamic asset weights in books, and dynamic formula weights in ranking systems. I also don’t expect any system to be correct 100% of the time. We don’t need 100% accuracy. Did you know that if a ranking system misranks 99% of stocks, it can still earn 50% a year?

It’s a matter of back-testing to find the best indicators. But back-testing factors is very, very, hard using the current tools. Steve mentioned needing over a thousand custom formulas. That sounds about right. That’s one reason why I am advocating for rules in books. That should allow some simple factor switching. After that I would like to see dynamic asset ranking in books. Eventually I would like to see tools to chart correlations between various indicators and factor out-performance.

Some have reported no luck with dynamic factor weights. But some have reported double the excess returns. Who is right? They both are. It depends on which indicators you use to pick factors. But I have seen suggestions that in the right hands, factor switching can double the excess returns of a strategy.

I think the most basic backtest should be to see if factor momentum even exists. To do that–and I’ve done it–you backtest 50 different strategies–some emphasizing value, some growth, some high volatility, some low volatility, some small caps, some large caps, etc.–over the past 22 years. Then you look at the correlation between returns over different periods. So, for example, I looked at the correlation between the returns of my 50 strategies over the nine months following a three-month gap (just to make sure I have new stocks) after 1 year, 2 years, 3 years, 4 years, 5 years, and so on. Here are my results, which reflect the average correlation measured on eight different starting dates. The X axis is the number of years looking back; the Y axis is the correlation with the returns of the nine-month out-of-sample period.

Of course, with correlations this low, there’s a lot of noise in the data. But the trend seems clear enough. You’re more likely to have factors work like they did in the past if you use a 10-year lookback period than a 1-year. If you use a 1-year period you’ll do better with a reversal than a continuation, since the correlation is quite negative. I’ve done this exercise with tons of different types of strategies, with tons of different OOS periods (I normally use 3 years instead of 9 months), over the last five years. The results are very consistent: anything less than eight or ten years has lower correlation.

Now, technically, using a 10-year lookback period is just as much “factor momentum” as a 1-year lookback period, and if so, I use factor momentum myself and strongly believe in it! It’s just the 12-month part that I find alarming.

All this presupposes that I’m understanding what is meant by “factor momentum.” But it might mean something entirely different, in which case all of this is entirely irrelevant. I have looked at all the papers that have been posted on this thread, but am no closer to really understanding what is meant by “factor momentum” than before.

By the way, our screens and simulations now allow you to use ten different ranking systems with the Rating and RatingPos commands. So backtesting a rotation strategy has never been easier. Not to say it couldn’t be even easier than that if we were to develop something along the lines you suggest, but choosing one out of ten ranking systems depending on certain conditions will give you a good sense of whether factor momentum (as I understand it) can work better than sticking with one ranking system. If your choice is purely portfolio-based, I suggest the following workaround: test 50 different portfolios with 50 different ranking systems, download the results into Excel, and choose the one with the best performance over the last X months for the next X months. That shouldn’t be too hard.

We have a ton of development work taking place over the next few months (which I’ll announce shortly). But after that, maybe we can look into rules-based books.

Lastly, our AI feature will be very adaptive and will change factor weights over time. It should be ready sometime this year.

---