"One of the MOST IMPORTANT PAPERS IN OUR FIELD in a long time." - Cliff Asness

Research
1

I received Cliff Asness’ (AQR) letter today, titled, “The Replication Crisis That Wasn’t,” which brings to light an important new paper published on SSRN, and may be valuable to P123 members.

Cliff says:

"Factor investing has long grappled with criticisms of data mining and more recently another, more basic criticism – some backtests might never have been right to begin with. A “replication” crisis is most specifically about this – not being able to even replicate the original work. But, the literature and my colleagues’ new work cover both potential problems.

"What we haven’t had, until now, is a formal test. Well, along come [color=blue]Jensen, Kelly, and Pedersen[/color] to test, and test brilliantly, what we have argued, largely anecdotally, for many years. Their results are rather startlingly (even to me) positive for the field in general. To quote the paper’s abstract:

"The majority of asset pricing factors: (1) can be replicated, (2) can be clustered into 13 themes, the majority of which are significant parts of the tangency portfolio, (3) work out of sample in a new large data set covering 93 countries, and (4) have evidence that is strengthened (not weakened) by the large number of observed factors.

“I think this is one of the most important papers in our field in a long time. I am, of course, incredibly biased, both by my own interests and the esteem in which I hold my colleagues. So I do encourage you to read the [color=blue]actual paper[/color] (and/or Larry’s Swedroe’s [color=blue]excellent summary[/color] on Alpha Architect) and decide for yourself.”

Enjoy!

2

From the Paper:

“The factor research universe should not be viewed as hundreds of distinct factors. We show that factors cluster into a relatively small number of highly correlated themes,…”

So, I think there is more than one way to do things. And who is to say the authors got everything right. But let’s go with this for a moment, just for discussion.

If one accepts the quote as being true, one would HAVE TO think about trying factor analysis wouldn’t they? After all, the quote very effectively sums up the reasons we have factor analysis, doesn’t it?

P123 is ideally suited for factor analysis with its nodes. With the ability to place different weights on the nodes and on the factors within the nodes according to the correlation of the factors (and the results of a factor analysis). With the node itself being a “latent factor.”

It is probably just me but I find factor analysis infinitely easier to implement that the hierarchical Bayesian analysis discussed in the paper. And they should give similar results. If one wants the “shrinkage” that Bayesian analysis offers, ridge regression of the factors which can be proven to give the same results.

FWIW, if the authors are right about this.

Jim

3

In my opinion, the paper is great. I spent a lot of time last night and this morning reckoning with it.

It proves what P123 has been saying for a long time: there are a very large number of factors that work out-of-sample to generate excess returns.

But it also shines some light on the process of factor arbitrage, though not explicitly. That requires some reading between the lines and some close examination of the results.

4

Yuval - I concur. My thoughts exactly.

When I started on P123 back in 2004 we only had very basic factors to use, extremely limited technical (price-based) factors, and in retrospect, nowhere near enough data. The data ran started in 2001 — and in that year small-cap stocks were questionable for reasons that have been lost to time (or to me at least) — so we had just THREE reliable years upon which to base backtests, and unfortunately, 3/4 of those years were in a bear market. So it wasn’t much to go on, but since data-driven investing was still unknown in those days to 95% of investors, the models actually worked quite well. It was 'easy pickin’s back in the early days.

Since P123’s very primordial start-up years of 2003-2004, Marco has been on an admirable mission to expand the available data set, insure the highest quality PIT data, and expand the capabilities of the platform. The recent capability that allows users to design factors that work from imported external data is a game-changer that opens the P123 platform up to virtually any data set that a strategy designer can imagine.

While some of those most basic factors (PE Ratio, Price/Book, etc.) were long ago worn to a nub by use by everyone and their grandmother, there are virtually unlimited new possibilities from combinations of datasets and indicator configurations that virtually nobody has considered before, and which offer incredible opportunity.

Five years ago, I gravitated to ETFs from stocks because ETFs were not used as much by quantitative professionals as stocks. And for the last few years, I’ve been developing new, complex and sophisticated custom formulas—such as a Custom Series I call the Momentum of Fear Indicator © 2021 ETFOptimize.com, Optimized Investments, Inc.

The Momentum of Fear Indicator is made from a composite of multiple Rate of Change measures of the $VIX data series, i.e., the CBOE Volatility Index — measuring investor’s option bets on increases or decreases in near-term volatility. Here are the signals provided by one version of this Custom Series:

The signal is ‘Risk-Off’ when at 0 and ‘Risk-On’ when at 1. You can see how this one custom indicator provides very accurate signals for when to be long (Risk-On) the S&P 500 ETF (SPY) and when to be in cash or a defensive ETF (Risk-Off).

Moreover, this one indicator, i.e., 'Momentum of Fear’©, is used as part of a composite of multiple other indicators that when combined provide even more highly accurate signals. Other indicators I’ve created, and from which I build various composites for different uncorrelated models, include the 'Trend Channel Indicator’©, Stochastic Close Indicator, 'Trend Momentum Indicator’©, Share Accumulation/Distribution, Volume Accumulation/Distribution, Bollinger Band Signals, MACD Risk Indicator, Breadth Thrust Indicator, Net Advancing Volume Signal, and more than 50 other customized, carefully tested and robust indicators …

In this way, progressive thinkers in the investment community can always stay ahead of the masses if enough investors come along to use and ultimately deteriorate the performance of well-know investment indicators that are today’s equivalent of the P/E Ratio 15 years ago. If an indicator isn’t working well for you, the sky’s the limit for your potential to create a new one that will work.

Creativity and innovation will always be able to provide Portfolio123 users with a competitive advantage if they put innovative ideas to work. And with the capability to use imported external data series that P123 implemented last summer, the potential to always stay ahead of the masses is only limited by the strategy designer’s (YOUR) creativity!

5

So, I wonder if those reading this paper and commenting on it believe in the method? It is a paper about the Bayesian method, I think. Or, at least, there is a lot about it and that is what they use use to prove their conclusions.

It would be okay, I think, if one did not completely buy into its usefulness. For example, what about the problem of non-stationarity of the data? Is that addressed?

But it seem that maybe some do believe in the method. At least for as long as it supports what they were already doing. And surely no one would go around citing studies they do not believe in to support what they are doing. That would be a pretty bad example of confirmation bias.

If members read this and end up believing in it then P123 might consider using it.

Whether I am fully on board with this or not, I would like to see the hierarchical Bayesian approach outlined here used for the designer models. This seems like an obvious use and I think anyone believing in hierarchical Bayesian methods would have to agree.

NOTE: While I would not mind seeing this used for the Designer Models this is not a feature request. P123 might have more important items on its to-do-list (truly).

On a practical note for anyone interested in this, the method most likely to succeed for finding a first practical use for this would be to look at JASP I think–at least to start with.

JASP is a free download that emphasizes the Bayesian approach and provides frequentist methods alongside the Bayesian methods. One could move to some good programs in R and Python if they are really motivated. But as I said in the above post there may be some simpler methods that accomplish same thing.

JASP download: JASP. JASP is good if you are interested.

Jim

6

All,

Low Risk is one of the “clusters” mentioned in this paper. And low return volatility is a factor within this cluster.

This is fairly well know. Indeed, P123 has a core ranking system called" Core:Low Volatility" and it is a node in the “Core: Combination” ranking system.

I had been looking at this separately even before this paper was posted in an attempt to understand why Minimum Variance Portfolios CAN outperform. And I mean outperform as far as absolute returns and not just risk adjusted returns.

Don’t get me wrong. I am not saying some of my backtests may not be misleading, that anyone could or should even begin to look at just the variance when developing a ranking system or deciding what equites (or ETFs) to invest in. Or that anyone else should try this at home.

But for those interested—especially in light of the results of this paper—here is a good link that expands on this idea and provides some possible theoretical insights:Finding opportunities through the low-volatility anomaly

Here is an article by Alpha Architect. I do not find it to be particularly good but it supports the idea that it is a serious subject: Low Volatility Factor Investing: Risk-Based or Behavioral-Based or Both?

BTW, it is easy to show that volatility-drag can have an effect on returns within the SPDR sectors. The sectors with the greatest average returns are not necessarily the sectors with the best CAGRs (average returns vs geometric mean). For example, XLV has a low-ranking mean weekly return (among the sectors) using the adjusted close starting January 2000 but is one of he better sector ETFs with regard to the geometric mean.

This anomaly appears to go beyond just benefiting from the reduced volatility-drag, however.

Jim

7

This paper confirms the findings in the paper we discussed.

https://www.sciencedirect.com/science/article/pii/S0378426621000819

Illiquidity is one of the best-known predictors of stock returns. However, it works only in microcaps (and probably in nanocaps). Outside this market segment, no illiquidity premium exists. New evidence from 45 countries in our recent JBF paper.

8

I just published a review of this paper. You can read it here: https://blog.portfolio123.com/thoughts-on-is-there-a-replication-crisis-in-finance/

9

What do people think of the magnitude of the alpha in this paper?

The paper says: “The posterior mean alpha is relatively stable around 0.3% to 0.4% per month during our sample.”

The upper confidence interval is below 0.5% for all factors in the paper (single factors to be sure).

Unlike, say the Sharpe ratio that uses the square root (of 12) to annualize a monthly number, I think you just multiply a monthly alpha by 12 to annualize it. So, 6% annualized alpha is the best case for a single factor used in the paper.

Marc Gerstein’s Designer Model–Underestimated Blue Chips–has an annualized alpha of 3.77% after slippage so maybe this is about right for larger-caps. This is Marc’s best model with respect to alpha. Maybe you can do better than Marc has done here.

Not bad really. Is 6% a realistic upper limit for large-caps (for most people)?

Let me be the first to say (or agree with previous posts even) that these factors may work better with micro-caps that have less analyst coverage. Micro-caps may not be an “efficient market” or may be less efficient. And that there is much more that can be said on the subject.

Full disclosure: I do use some of these factors myself.

Jim

10

Great summary, Yuval. Appreciate the insights.

11

Thank you Yual and all, Great Paper!!!

I get an error in running the Factor Zoo (after copying it)…

Best Regards

Andreas

12

You need to put in two custom formulas, which are given in the blog post.

13
14

Thanks Yuval for taking the time to test these factors. Iwanted to fool with the ranking system. Here is the error i get, after i copy the custom formulae

In ‘highest 5 days of return scaled by volatility’-> In $21medmed: Error near ‘�Eval’: Invalid command ‘�Eval’

15

Make sure your quotation marks aren’t smart quotes. If they’re kind of curly, they won’t work. You have to change them to straight quotes.