Here at Version 1, we are passionate about advances in technology, specifically how new technologies and models can be employed to benefit our customers.

Machine learning is one of those technologies, so we spoke with Richard Faloon, our Capital Markets Enterprise Architect, to find out more about how machine learning models are benefiting our automated trading clients.

Tell us about the Machine Learning models that you are using with our capital markets clients?

There are many machine learning models – Support Vector Regression (SVR) and Long Short-Term Memory Networks (LSTM) are just two that can bring real benefits to the automated trading community.

The Efficient Market Hypothesis states that theoretically, neither technical nor fundamental analysis can produce risk-adjusted excess returns, or alpha, consistently and only inside information can result in outsized risk-adjusted returns.

So what does that actually mean?

In essence, the market is volatile and random, and that it should be impossible to predict future behaviour on past results.

So why bother with automated trading?

For starters, we know in practice that this is not totally accurate. There are both micro and macro-level data points that can be used to inform non directional percentage movement of tradable instruments, such as past order book activity, sentiment analysis, growth, earnings and dividends.

The general trend of quantitative analysis was to build a model of the tradable instrument and use that to generate a theoretical fair price. These mathematical models often include past data points to generate parameters such as historical volatility; they also use multiple correlating factors of current market conditions to assist in generating the end theoretical price. These automated trading algorithms then look for a market where the theoretical price compared to the model price is overvalued or undervalued which is called edge, the greater the edge the greater the potential profit. This all makes sense from a mathematical model standpoint however actually converting the edge to profit is far from guaranteed.

With machine learning we are using the SVR or LSTM to generate the behavioural model for the instrument and using that to generate our theoretical price for T+1 time to generate edge.

Why use ML over quantitative models?

As mentioned before, there are lots of highly correlated inputs that can be used to generate a single output. This is ideal for training neural networks. The neutral networks can be retrained often therefore making them much more dynamic than quantitative models. For instance if market dynamics change such that the quantitative model no longer applies a new model must be constructed, however for ML this is a retraining exercise with a newer data set, no development time is required.

Why not just always use ML in automated trading?

At this stage, it is still relatively new compared to traditional methods and anecdotal reports of success vary. This is partly due to numerous pitfalls such as overfitting the model, incorrect model selection or incorrect data selection. There is also the speed to consider – ML can be slower to execute than many other statistical arbitrage algorithms – meaning an opportunity can be easily missed.

At Version 1, the current focus is on ML for crypto currencies but we are also applying it effectively to other businesses and industries. Given the high volatility in the market we believe it to be an excellent candidate for prediction of non-bias directional movement and the opportunity that trading its derivatives presents.

If you want to find out more about Machine Learning and how it can help your organisation, get in touch with our team.