Applying Machine Learning and Science to Trading Decisions

Sean Kruzel, founder and C.E.O of Astrocyte Research spoke at the most recent Boston ODSC Meetup about using Machine Learning in the investment industry. His first key point stressed that the type of data used in finance is different than that usually used in Machine Learning.

The latter uses 'stale' and static temporal data while the former uses evolving data with an adversarial component due to the nature of the markets. Thus the scientific process in finance is harder to implement, and relies on prior beliefs and flexibility.

A lot of Machine Learning investments come from Roboadvisors. While this product is very scalable, it is overly simplistic with the use of traditional techniques like linear regression to estimate factors like risk and return.

For one, using correlations only exposes superficial relationships within stock data. More importantly, using decades of historical stock data to make future predictions is flawed as the forces driving the market years ago may have no relationship to today's catalysts.

How then is Machine Learning used in finance? Three sections where it can be applied included the evaluation of portfolio managers, deciding when to enter or exit trades, and converting forecasts into investments. A Bayesian Machine Learning framework works well in tandem with a traditional investment metric like the Sharpe Ratio, the ratio of expected return to expected risk.

In evaluating portfolio managers the target Sharpe Ratio and frequency of evaluation depends on whether one is an asset or hedge fund manager. The key task is a classification problem to calculate the probability that a portfolio manager has a given Sharpe Ratio within a certain time interval.

Replacing the Sharpe Ratio with the Bayes Factor in trade entries and exits means that traditional methods like tracking losses over time and allocating a dollar value to risk can be enhanced by a Bayesian classification framework. The process in evaluating forecasts is similar.

There are more applications to speak of, and more will soon join the fold. In the future Machine Learning could even be applied to automated trade idea generation.

Find this article helpful?

Learn more about this topic and others just like it at ODSC's Machine Learning Conference in California this year.