Brain Machine Learning proprietary platform is exploited to generate a daily stock ranking based on the predicted future returns of a universe of 1000 stocks on four time horizons: 2,3, 5 and 10 days (other time horizons could be developed and tested upon request).
The model implements a voting scheme of machine learning classifiers that non linearly combine a variety of features with a series of techniques aimed at mitigating the well-known overfitting problem for financial data with a low signal to noise ratio.
Some example of features are:
- Time varying stock specific features like price and volume related metrics or fundamentals
- Time fixed stock specific features like the sector and other database information
- Market regime features such as volatility and other financial stress indicators
- Calendar features representing possible anomalies, for example the month of the year
The stock universe is represented by the 1000 US stocks with the largest market cap and it is updated every year.
The model is trained and tested with a walking forward approach.
An history of approximately 10 year is available as Free Trial for testing in the time interval 2010/01/01- 2019/02/28.
In the following graph we report the cumulative returns of a long-short portfolio with weekly rebalancing based on the out-of-sample predicted ranking related to the future stock returns for next 5 days.
A one page summary with more details is available at this link