Applications Of Machine Learning In Finance

Accurate and efficient financial data analysis are crucial to investors that can help them to avoid the potential risks and to make profitable strategies.Therefore,financial data analysis have important research significance.However,the financial market is a complex nonlinear dynamic system which affected by many factors,so it is a challeng-ing task to analyze the financial data according to the acquired information.Due to complexities and long-term dependencies of financial time series forecast-ing,this paper proposes a forecasting model with Long Short-Term Memory(LSTM)neural network based on deep learning technique.Firstly,stacked denoising autoen-coder architectures are applied for feature extraction from the basic market data and the technical indicators of financial time series.Then,LSTM neural network uses the ex-tracted features as inputs for financial time series forecasting.The accuracy of financial time series forecasting is improved by long-term dependencies characteristics of LSTM neural network.Compared with the traditional neural network,the experimental results show that the LSTM neural network has high forecasting accuracy when combined with deep learning technique by using the stock index data.In addition,this paper explores an adaptive trading system for financial asset trad-ing that combines deep feature learning techniques with direct reinforcement learning.Deep denoising autoencoder architectures are applied for market environment explo-ration and feature learning,and direct reinforcement learning is used for realtime finan-cial asset trading.We evaluate the deep trading system on real financial time series with different performance functions(i.e.,the total profit,the differential Sharpe ratio,and the differential Sortino ratio),and the experimental results show that the reinforcement trading systems are efficient when combined with deep structures.When optimizing a deep trading system,it is suggested to use the risk-adjusted return as the reward evalua-tion for reinforcement learning which balances risk and profit.In addition,we increase transaction costs to approximate the slippage phenomenon that occurs in real trading,which is due to the data set consisting of indicative quotes that are not actually tradeable in real-time.