Portfolio Optimization under Ambiguity Aversion Using Deep Learning, Machine Learning, and Time Series

After modern portfolio theory and the Markowitz model, which considered risk and risk aversion in order to optimize investment portfolios and caused a great revolution in the field of economics, now behavioral finance topics such as ambiguity and ambiguity aversion have also been considered. Given the positive results achieved by machine learning algorithms, deep learning, and time series in predicting stock returns, this article used three methods that performed better than in most studies in the field of stock return prediction. Long short-term memory models are used in deep learning, random forests in machine learning, and autoregressive integrated moving average (ARIMA) models in time series. This article compares forecasting models in order to determine which model is more accurate and then predicts the returns based on that model. This article combines their predictive results with risk and ambiguity metrics in order to develop optimization models. An evaluation was conducted based on historical data for the six-year period from 2017 to 2022 for the top five industries in the Iranian stock market. According to the experiments, the deep learning model performed better than other models when it came to long-term predictions. Hence, this article presents a model for optimizing stock portfolios by integrating investor unknowns with forecasting models for the first time.