Intelligent Financial Portfolio Construction: Machine Learning Based Optimization

Abstract

Investors have always been highly interested about stock price forecasting. It’s observed that lower middle-income people and middle-income people can contribute only 10-15% of their wages in investment. Machine Learning based stock price forecasting is proven to be most efficient for price prediction according to the conventional research processes. The proposed research is conducted in order to derive a quality stock price forecasting technique for lower middle income or middle-income people so that their financial distress could be relieved. Here, the research introduces chi-square test for finding the differences between observed and predicted prices. For the price prediction the machine learning (ML) tools such as SVR, LogR, XGBOOST, DTR, RFR, and LSTM are introduced, where input features for the ML are obtained from principal component analysis (PCA) and statistical averaging method. Statistical averaging method calculates a new feature from the stock price features open, low, high, adj close, and close, and finally obtains a new feature vector for a ML algorithm combining the stock price features and the new feature. Moreover, Portfolio is constructed observing the higher trend of predicted prices for the different stocks of an investor to reduce the risk of investment. From the experiments it is observed that the proposed average feature-based method using the chi-square test (confidence 10%) achieved a feature dependency score of 26%, whereas PCA-based features did not achieve minimum benchmark of 10%. Besides, LSTM is found to be the best forecasting method and provides the highest accuracy of predicted prices which are 90.11% and 88.15% for the proposed feature set using the statistical averaging method and conventional price features (open, low, high, adj close, and close), respectively. Moreover, the portfolio created based on statistical average price feature provides a return on investment of 23.32% and reduces the risk by a Sharpe Ratio of 50.32 (standard value should be greater than 1.0).