{"id":193130,"date":"2023-07-07T11:00:00","date_gmt":"2023-07-07T15:00:00","guid":{"rendered":"https:\/\/ibkrcampus.com\/?p=193130"},"modified":"2023-07-07T10:55:55","modified_gmt":"2023-07-07T14:55:55","slug":"mill-street-research-maer-stock-ranking-model-do-you-see-what-ic","status":"publish","type":"post","link":"https:\/\/www.interactivebrokers.com\/campus\/traders-insight\/securities\/stocks\/mill-street-research-maer-stock-ranking-model-do-you-see-what-ic\/","title":{"rendered":"Mill Street Research MAER Stock Ranking Model: Do You See What IC?"},"content":{"rendered":"\n

A few months ago, I wrote a blog post<\/a> giving a brief history of the development of Mill Street\u2019s MAER stock selection model and comparing the in-sample and out-of-sample test results now that it has been just over 10 years since the current model was originally launched.<\/p>\n\n\n\n

The key metric I showed was the Information Coefficient (IC), which is the correlation between the decile ranks of the stocks and the decile rank of their subsequent 1-month returns (it can be done on other return horizons as well). <\/p>\n\n\n\n

To visualize it, this is a stylized hypothetical example of what a 100% information coefficient could look like: perfect linear relationship between the decile rank and the return in the subsequent month.  An IC of -100% would be the exact opposite: top decile ranked stocks would have the lowest next-month return, etc.<\/p>\n\n\n

\n
\"information<\/figure>\n<\/div>\n\n\n

Source: Mill Street Research<\/em><\/p>\n\n\n\n

This is an example of what a 50% IC could look like: higher ranks have better next-month returns than lower ranks on average, but the relationship is more mixed. <\/p>\n\n\n

\n
\"information<\/figure>\n<\/div>\n\n\n

Source: Mill Street Research<\/em><\/p>\n\n\n\n

A 0% IC would show no relationship between decile ranks and returns at all (just random results). If a stock ranking has no predictive ability, we would expect the average IC over a longer period to be near zero.<\/p>\n\n\n\n

Why focus on the IC rather than a traditional portfolio-based backtest? Well, I have done MANY portfolio-based backtests and they are indeed much of the focus when discussing MAER with clients \u2014 the results of the standard decile-based backtest are included at the end of this post.<\/p>\n\n\n\n

But the key issue is that one must make many decisions about<\/strong> how to create portfolios based on an underlying ranking model<\/strong> (or alpha score), and any of those decisions can produce seemingly different results in the test.<\/p>\n\n\n\n

Portfolio returns = stock ranking + portfolio construction constraints<\/strong><\/p>\n\n\n\n

The differences in backtest results caused by differing portfolio construction<\/em> decisions will in many cases simply reflect the constraints imposed in the portfolio construction process, rather than the underlying ranking itself.<\/p>\n\n\n\n

That is, a single ranking of stocks can produce an almost infinite array of possible portfolio results depending on the portfolio construction rules used<\/strong>. Varying the rules can thus cause random variability in the results that is not directly related to the ranking itself.<\/p>\n\n\n\n

What kinds of portfolio construction choices can make notable differences? Big ones include:<\/p>\n\n\n\n