Deep learning for trading and hedging in financial markets

Deep learning has achieved remarkable results in many areas, from image classification, language translation to question answering. Deep neural network models have proved to be good at processing large amounts of data and capturing complex relationships embedded in the data. In this thesis, we use deep learning methods to solve trading and hedging problems in the financial markets. We show that our solutions, which consist of various deep neural network models, could achieve better accuracies and efficiencies than many conventional mathematical-based methods. We use Technical Analysis Neural Network (TANN) to process high-frequency tick data from the foreign exchange market. Various technical indicators are calculated from the market data and fed into the neural network model. The model generates a classification label, which indicates the future movement direction of the FX rate in the short term. Our solution can surpass many well-known machine learning algorithms on classification accuracies. Deep Hedging models the relationship between the underlying asset and the prices of option contracts. We upgrade the pipeline by removing the restriction on trading frequency. With different levels of risk tolerances, the modified deep hedging model can propose various hedging solutions. These solutions form the Efficient Hedging Frontier (EHF), where their associated risk levels and returns are directly observable. We also show that combining a Deep Hedging model with a prediction algorithm ultimately increases the hedging performances. Implied volatility is the critical parameter for evaluating many financial derivatives. We propose a novel PCA Variational Auto-Enocder model to encode three independent features of implied volatility surfaces from the European stock markets. This novel encoding brings various benefits to generating and extrapolating implied volatility surfaces. It also enables the transformation of implied volatility surfaces from a stock index to a single stock, significantly improving the efficiency of derivatives pricing.