Financial markets are becoming a new proving ground for deep learning. The AI technique has already achieved remarkable success in image recognition, speech detection and sentiment analysis, and is believed to be well-suited for dealing with financial data.
At last month’s RE•WORK Deep Learning in Finance Summit in London, leading AI industry practitioners and academics from prestigious universities discussed their research, provided insights on business trends and real-life AI applications, and addressed current challenges facing the AI industry as a whole.
Synced visited the summit to explore how deep learning techniques, such as neural networks and LSTM, can be applied to proprietary trading. This article focuses on Dr. Filippo Scopel’s presentation Learning to Trade and Dr. Luigi Troiano’s Supporting Trading in Financial Markets using Deep Learning Tools.
Dr. Scopel is a machine intelligence engineer at Merantix, a Berlin research lab building AI ventures. Dr. Scopel’s research speciality is mathematical modelling and algorithm development with applications in finance and other fields. For the past year, he has been training deep learning models for time series prediction on financial microstructure data.
Dr. Scopel discussed Merantix’s use of deep neural networks to predict price movements with short-term samples of under five minutes, and how these predictions can be transferred into trading strategies and then trading algorithms.
Deep learning models are trained to make predictions by absorbing raw data such as past prices, conducting data computation, and generating results. Due to the complexity of the hidden layers within such neural networks, it is impossible to make an exact interpretation of how those predictions were generated.
Predictions will be turned into specific trading strategies regarding bid-asks and particular time horizons. A decision to buy, hold, or sell is then generated and executed by the algorithm.
Although the Merantix predictions can be more accurate than traditional methodologies, Dr. Scopel pointed out the system is far from infallible. Even with the right predictions, Merantix can still record losses due to the characteristics of high-trading frequency. What he means is transaction costs, and full bid-ask spreads can erode the potential profitability, sometimes resulting in losses.
Dr. Troiano is an Assistant Professor of AI, Machine Learning and Data Science at the University of Sannio. He began his presentation by identifying areas within finance where AI and particularly deep learning could be applied, including reshaping the analysis space, searching complex patterns in data, and trading robotisation. Data filtering removes noise to enable better quantitative analyses, and complex correlations and co-occurrences in large datasets can be identified to generate improved trading strategies. Automation reduces costs and improves the long-term effects of hedge funds and proprietary trading firms.
In forecasting the variability of prices, also referred to as volatility, Dr. Troiano said long short-term memory (LSTM) networks — a type of Recurrent Neural Network (RNN) capable of learning long-term dependencies — work best when volatility is extremely high, which is a beneficial finding since such periods also provide high-profit opportunities.
Dr. Troiano then discussed his research into algorithmic trading using technical indicators such as the moving average convergence divergence (MACD) with deep learning techniques. A deep learning system can construct trading strategies after observing historical prices and indicator values instead of being explicitly programmed to execute those strategies. What this means is that AI has the potential to learn, or more appropriately, invent trading effective strategies.
However, Dr. Troiano stressed that his research is still at its preliminary stage, and only certain deep learning techniques have been tried. More research is needed to determine how neural networks can improve on old-fashioned methods.
Overall, it is clear that deep learning can accelerate the performance of tasks such as short-term future price predictions and proprietary trading strategies, but the AI technique has not yet reached maturity for full application in the trading market. The industry needs to be patient before widely employing neural networks to make market predictions.
Author: Guangyao Ma | Editor: Tony Peng, Michael Sarazen