Hey Rudiger,

Great question and interesting paper. It's a well known issue that, often, seemingly new factors will just be re-discovering existing and known anomalies. Of course the way that this is usually expressed is through a factor model

F_new = alpha + beta_1 * F_1 + beta_2 * F_2 + ... + beta_n * F_n

If a linear regression estimates low alpha when existing factors F_1, ... , F_n are used then you have reason to suspect that there isn't much new content beyond what is already explained by F_1, ... , F_n. Keep in mind that 'alpha' is often used to mean generally new information, however it will depend on what factors are being used as your existing factors in this model. So a factor can appear to have high alpha but just because a similar known factor isn't being used as one of the independent variables.

Having completely independent factors in a trading model is the ideal situation, as even when some aren't working others will, and each will bring something new to the table. You're completely correct that if you use a factor that is explained by other existing factors, you aren't really providing alpha as much as increased exposure to these existing factors and therefore risk.

It's incredibly important to measure factor risk exposure, not only when evaluating new alpha factors, but also when considering an overall model or algorithm return stream. Quantopian doesn't currently offer a user-facing API to measure factor or algorithm risk exposure, but it is something we're currently thinking about. You can also manually compute it yourself as is shown here:

https://www.quantopian.com/lectures#Factor-Risk-Exposure

When we evaluate algorithms for allocation, we use a wide variety of tools. Our research team is constantly building new infrastructure to do it better, and we certainly look at risk exposure of our algorithms during the process. Our goal is to turn as much of what the research team produces into front-facing tools and APIs so users can benefit. That has already been done with PyFolio and Alphalens, our goal would be to do the same for a risk modeling API.

Hi Delaney -

Having completely independent factors in a trading model is the ideal situation

My understanding is that at present, you have ~15 algos trading an average of $100K real money each, and many others in the queue. You have the advantage of selecting from presumably a relatively large, diverse pool of ideas. To what extent have you been able to achieve the ideal of having independent factors?

I realize that you may not be at liberty to share much information, but I feel compelled to ask the question, since you might be starting to get some sense if the theoretical is realizable. Have you been able to apply statistical tests to show that at a reasonable confidence level that there are N independent factors in the pool of algos trading real money, and the ones you are considering adding?

Best regards,

Grant