Finite Order Universal Portfolio with Moment-Generating Function and Machine Learning Model

This paper investigates the creation of universal portfolios generated by recursive calculation of moment-generating functions for continuous variables using data from Yahoo Finance and Refinitiv. Three portfolios were constructed, each comprising three companies, with stock selection aided by machine learning models (Ridge, Lasso, ENET, BRT, LSTM). These models predicted company performance and classified top-9 and bottom-9 performers, benchmarking against KLSE. All the portfolios constructed by machine learning models outperformed the KLSE returns. Bayesian optimization technique was adopted for the LSTM and BRT model to determine the best parameters that minimize the MSE. Additionally, Bayesian optimization was also adopted to determine the best parameters for different orders, portfolios, and distributions that maximize the portfolio's return. The performance of the universal portfolios was compared to the Best Constant Rebalanced Portfolio (BCRP) and other strategies. Results showed that the best-performing universal portfolios closely matched BCRP, surpassing all other strategies. Gamma and Generalized Pareto distributions performed best, aligning closely with BCRP for all orders. The company selected is based on the data cleansing criterion, which does not represent the whole stock market in Malaysia. Further studies can be carried out by increasing the number of orders in the universal portfolio, or by including 4 or more companies in a single portfolio. Besides, only 4 distributions (Gamma, Lognormal, Weibull, and GPD) moment-generating functions used in the universal portfolio construction, an exploration into other continuous random variables can be conducted. The universal portfolios were constructed to find the best parameters for all trading periods, hence varying parameters of distribution on certain intervals can be conducted in the future. A wider range of distribution parameters can be determined in the future as this paper studied only a finite range of parameters.