Based on the results of a study conducted by an international team of researchers at the School of Business and Economics at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), AI-based algorithms can function as stock market traders. And they’re not just good at it. They’re actually much better than real live traders (man, that hurts). And they seemed to do particularly well during times of financial turmoil.
To arrive at these results, the team –headed by Dr. Christopher Krauss of the Chair for Statistics and Econometrics at FAU — studied the S&P 500 Index which basically consists of the top 500 US stocks. For the period from 1992 to 2015, they used different methods, specifically, ‘deep learning, gradient boosting, and random forests’, to generate daily predictions for each of the 500 stocks.
As different as the methods were, the models they trained learned one particular complex function — to define the relationship between features that were price-based and the future performance of the different stocks. Armed with this knowledge, the models were able to perform astonishingly well in the market.
From the year 2000, the returns earned were higher than actual market returns by 30% per annum. In the nineties, the returns were even higher. And notably, the models did extra well at times when the financial market was most unstable.
For instance, when the dot-com bubble collapsed in 2000, returns using the AI method were over 500%. And during the world financial crisis in 2008 – 2009, the AI systems showed returns of almost 700%. Well, this is not beating the market, it’s more like annihilating it.
As Dr. Krauss said in a statement they issued: “Our quantitative algorithms have turned out to be particularly effective at such times of high volatility, when emotions dominate the markets.”
In other words, when humans tend to lose focus and make decisions based on emotions rather than reason, AI remains on track because they only consider actual data and do not have any emotions that can get in the way of making sound judgements.
While there is much benefit that can be gained from using AI in the financial market sector, Dr. Krauss cautions that it is ‘not necessarily the ‘Holy Grail of capital market trading’. As he pointed out, earnings declined in the latter years of the study, and it’s probably because machine learning and computer-based trading were becoming more common then. Meaning, it was becoming harder to be at an advantage by using computational power because the technique has become more widespread and was no longer being used by a limited few.
Still, the outcome of the study proves without doubt how much wider the applications of deep learning are, especially in cases when human emotions tend to affect actions and decisions negatively.
Dr. Krauss and his team are now working on follow-up projects that involve ‘larger data sets and very deep network architectures’ that will hopefully result in even better forecasting models.
The results of the study have been published in the European Journal of Operational Research under the title “Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500”.