It is no secret that nowadays most transactions on many financial markets are being done by algorithms: computer programs that buy and sell according to certain pre-specified rules. This development has brought about a monumental shift in the micro-structure of financial markets: computers taking the place of what used to be human traders.
And now we have entered the era of machine learning: computer programs that buy and sell based on rules they have inferred themselves from the data, instead of rules being hard-coded by humans. Surely that will impact the micro-structure of financial markets as well.
I want to zoom in on one class of machine learning algorithms that might be particularly interesting for the financial markets: recurrent neural networks. A recurrent neural network takes into account the sequence (or “development”) of a variable over time in making its predictions. For example: an algorithm that tries to find a pattern between the end-of-day prices of a stock over the past 90 days. Assuming the algorithm succeeds, he can know predict tomorrow’s stock price with an accuracy that allows it to make an above-average risk-adjusted :profit.
Such developments are not only interesting from a trading or computer science perspective. It might also touch upon our fundamental theories of how financial markets work. And one could say that there are few economic theories more fundamental than the Efficient Market Hypothesis.
The Efficient Market Hypothesis comes in many forms. A minimal version (the “weak version”) of it implies that:
Traders who engage in technical analysis might beg to differ. Such traders believe that there exist patterns in the prices of financial products that repeat themselves over time, and that being able to spot such a pattern before it has fully completed, can allow you to make an above average risk-adjusted profit (see picture for well-known patterns).
Technical analysis is generally deemed more an art than a science: sometimes it works and sometimes it doesn’t. Especially in retrospect it is very easy to detect the pattern that has occurred. But this doesn’t imply that there are no patterns.
Maybe the patterns are just too complicated for humans to grasp. Maybe there are certain non-linear dependencies between prices at subsequent points in time that do in fact repeat themselves, and that would allow a trader to make above-average risk-adjusted returns, if only they would know them. They are simply to complex to be caught by a set of lines.
Which brings us back to Recurrent Neural Networks. Maybe such algorithms could succeed in capturing complex time-series dependencies in stock prices, thereby succeeding (assuming a reasonable accuracy) in predicting future prices based on past price-information (possibly in a limited segment of the financial markets). Call it technical analysis 2.0.
If that would be the case (and who knows: maybe this is already being applied): what does that imply for the thesis that “future prices of financial products cannot be predicted by analyzing prices from the past”? I leave the answer up to you.
Moreover, maybe such algorithms can be used in economics to test parts of The Efficient Market Hypothesis. If researchers could find certain parts of the financial markets that allow recurrent neural networks to make above average risk-adjusted profits consistently, then that might be an indication that past prices can be predicted based on prices from the past.
I find such ideas very interesting. I am curious what you think.