Machine learning in financial market

Abstract

This thesis will present 2 volatility forecasting models using deep neural network and a Hidden Markov Model (HMM) for market regime classification. Chapter 3 proposes a volatility index prediction using deep neural network, and Chapter 4 propose a jump detection model using Long-Short Term Memory (LSTM) neural network with a pattern recognition algorithm. In Chapter 5, a market regime classification model, which is aimed to detect intraday returns change using unsupervised machine learning model, is proposed.

In Chapter 3, titled Volatility Index Prediction using Deep Neural Network (DNN), a 3-layers DNN model is proposed. It is aimed to predict the direction of the next day US volatility index (VIX), given volatility index in different markets, i.e. Hong Kong, Japan and Europe. The experiment results show that the 3-layers DNN model outperforms classical financial market volatility time series forecasting models in terms of accuracy. The proposed DNN model, which takes historical data sequences into account, helps generating the outperformance.

In Chapter 4, we proposed a new hybrid method to detect jumps in the financial market time series. Jump detection is important in financial market, because it implies a volatility change or increasing risk. There are different risk management strategies, which uses volatility as an input parameter, such as tail protection strategy.

The key contribution of the hybrid jump detection model is that, the parameters of the jump magnitude threshold need not be predefined. Also, the model combines a Long-Short Term Memory (LSTM) neural network model and a machine learning pattern recognition model, which is aimed to capture the historical time series pattern.

The hybrid jump detection model is trained and evaluated using actual financial market data, using worldwide equity market indexes in major developed and emerging markets. The results of the experiment show that our proposed hybrid jump detection model is effective to detect jumps compared to the other classical jump detection algorithms, in terms of detection accuracy.

In Chapter 5, an intraday price returns regime classification using a 3-state first-order Hidden Markov Models (HMM) is proposed. Intraday financial market data is applied. the HMM model is tested using several equity indexes, volatility indexes and currencies markets. The HMM model classifies the market into 3 regimes, i.e. bull, neutral, and bear market. The experiment results indicate that implementing HMM as an intraday trading buy/sell signal is possible to outperform the benchmark in terms of returns.

In Chapter 6, it is the conclusion of this thesis and suggested further research.