Adaptive Statistical Evaluation Tools for Equity Ranking Models

Abstract

A major challenge in the investment management business is to identify which stocks are likely to outperform in the future, and which are likely to perform relatively poorly. To this end the strategy adopted by Genus is to identify factors (auxiliary information about the stock such as earnings-to-price ratio or dividend yield) that they believe are associated with future out-performance (i.e. factors that have predictive ability). The best of these factors are then combined (Genus use a weighted average) into a model which is used to rank the universe of stocks month-by-month. This ranking is then used to as the input to a trading strategy, resulting in a modified portfolio.

Genus had provided us with sample data, consisting of just over 12 years worth of monthly returns on a universe of 60 stocks, along with time series of 34 factors for each of the stocks. Using these data, the approach was to build software (MATLAB) models for:

1. ranking the stocks based on factor information;

2. implementing a trading strategy based on a stock ranking and assessing the performance of a given trading strategy by looking at measures such as hit ratio, information ratio and spread.

The IPSW team implemented a simplified trading strategy of selling the entire portfolio each month, and using the proceeds to invest equally in the top 20% of stocks as given by the computed ranking. They also implemented the following measures of portfolio performance: excess return, hit ratio and information ratio.