In last week’s post, I made a strong assertion that has caused some great feedback and comments. When I first heard Mark Kritzman make a similar point at a UBS conference a few years ago, I had a similar reaction: “Hey, I’m a quant and I love my factors. They are definitely real!”. I still believe that using the ideas that are captured by what we call “factors” can be done successfully. But, as I look around at the current state of factor investing and the disappointing results that investors are facing, I am now understanding more about what Mark meant.

Here are some additional clarifying points.

When I say that factor investing and factors don’t exist, I mean it from the first moment perspective (the mean) and not the second moment (the covariance).

To build a good risk model that forecasts volatility or tracking error, using ‘factors’, i/e characteristics that capture correlation, is essential for model’s accuracy. There are many examples of such factor-characteristics like sectors, industries, currencies, and the cross-sectional ones like beta and size.

The confusion arises when the ‘factors-must-earn-a-risk-premium’ argument is introduced. According to the historically dominant academic financial theory, if something is a ‘risk factor’, it must get compensated with an extra return premium. This brings the conversation back to the first moment - the average return - and this is where things start to break down for me.

While in the ‘factors for risk models’

Factors like Price to Earnings and Price to FreeCashFlow might produce identical risk model forecasts because of the similar correlation they measure across stocks. But the returns to these two factors might be radically different to the point where one is negative and the other is positive. Because of these large return differences and the large number of permutations that any given factor can have, I don’t believe that assigning and relying on a required return premium to something like a broad value factor is accurate or wise. It diminishes the need to refine and innovate in order to actually capture the extra return from value investing.

We can look at lots of examples but even the basic one in the Fama-French data library uses a basic version of a dynamic contextual approach for the main factors, by ranking stocks within their size baskets. If you remove the size context, the results change. If you refine the context to, say, industries or company’s growth rates or small caps, the results change. If you start to tweak the ratio’s definition by using quarterly book values instead of annual, the results change. How can we reliably speak about factor-premia (i/e the positive return that we expect to earn from an abstract concept like value) when there is so much instability around the average each version generates?

Beyond the empirics, the theory behind factor premiums is still very confusing. First, I still often hear the argument that volatility is the same as risk, hence you should get compensated for holding something more volatile, which clearly is not what the original academic theory says. Only undiversifiable, systematic risk deserves a premium (like Beta, which does not earn one in practice). Second, most of the ‘factors’ in the popular factor premia baskets are clearly in opposition to the factor-premia academic theory, such as price and earnings momentum, accruals, profitability, low-beta, low-volatility etc. Perhaps only one factor, value, aligns in the right direction with theory, and even then, there has been a lot of debate about whether value stocks are actually ‘riskier’.

I have built, tested, and invented ‘factors’ for a long time, and anyone who has worked with the raw data as much can tell you how sensitive the results are to every line of code. Part of it is noise. The other part is a real opportunity to innovate and refine the models. If instead of blindly relying of some abstract and magical return that is supposed to arrive from factors, the researcher instead focused on innovating and developing characteristics that identify businesses that outperform in the future, the results would be much less disappointing in my view.