Factor Models: Can There Be Too Many Factors in a Model?

With the impending transition off of Barclays POINT, questioning the importance of specific factors will be especially important for fixed-income managers faced with making decisions around which new factor model to utilize that will likely not have the same factor decomposition as POINT.

With the impending retirement of the popular fixed income risk and performance analytics system Barclays POINT, I thought it would be a good time to post a series of blogs on some interesting insights I’ve gleaned from recent conversations relating to investment risk decomposition. These conversations focused on the evolution of factor models and the inherent flaws associated with them. At the most basic level, every conversation reflected the theme that each factor model—no matter how simple or complex—is just a statistical observation of broad risk sources that may influence a stock’s performance.

Picking these risk sources can be tricky and the identification of these sources has dated as far back as Markowitz’s mean variance model in 1952. Later, in 1964, the CAPM essentially gave us our first factor model by indicating that a security’s exposure to market risk utilizing beta will affect a security’s return. Then, in 1976, Stephen Ross developed the Arbitrage Pricing Theory, introducing the concept that assets should be priced as a linear function to their sensitivities to multiple factors. The problem with this was that the factors weren’t defined, leaving questions around which factors investors should care about or how many factors there are. It wasn’t until 1993 that Fama and French published a three-factor model that expands on CAPM indicating that in addition to market risk, company size and company book-to-market ratio should be used to measure and predict returns. It is said that this three-factor model explains over 90% of a diversified portfolio’s returns.

Even with this, Fama and French decided to publish a five-factor model that adds a profitability factor, which simply subtracts the returns spread for the most profitable firms from the least profitable firms, and an investment factor which takes the return spread for the firms that invest conservatively vs. the firms that invest aggressively. Adding these two factors are estimated to have added 3–4% additional explanatory power to the three-factor model. When asked about adding factors to this model, Fama will indicate that he is not fully sold on the investment factor yet which presumably will be his focus prior to adding factors. Fama’s factor models are still used today by very successful firms such as AQR and Dimensional Fund Advisors. This factor model, while more robust, did not include the momentum factor which was incorporated in an extension of the three-factor model by Carhart in 1997 and has proven to be effective ever since.

With that said, when I think about how factor models have evolved today to include—in some cases—over 100 factors and sub-factors, I wonder how much incremental value could be added by immensely increasing the number of factors in the model. This leads me to question if adding these factors could be impairing results to distract from the “true” market factors that explain risk. A simplification of the factors as they exist in today’s complex factor models are just statistical observations of how a universe of securities react to factor changes historically. If we are just looking at historical statistical observations of how a return reacts to some other variable that is changing over that same period of time, where do we draw the line for what we are calling a market factor?

To use a dramatic example, if one were to use Super Bowl winners to predict an up or down market, they could prove statistically that an NFC winner would predict an up market and an AFC winner would predict a down market and they would be right approximately 80% of the time. Now, simple logic tells us that using Super Bowl winners to predict how the market will perform is a really bad idea because there is no meaningful connection whatsoever between the two variables. I use this example simply to prove a point.

When thinking about factor models, we have been provided with a solid framework by pioneers of our industry that have proven to explain much of a portfolio’s returns and risk. This isn’t to say that we can’t improve on these concepts, but with the surge in the publication of factor models and the increased use in the market, it is important to ask if the model being used truly provides the best decomposition of investment risk. It is also important to ask if modern factor models are truly advancements of risk decomposition or if adding new factors to a model is becoming an exercise in event studies and data snooping.

With the impending transition off of Barclays POINT, questioning the importance of specific factors will be especially important for fixed-income managers faced with making decisions around which new factor model to utilize that will likely not have the same factor decomposition as POINT.