Portfolio Risk-Return Decision Optimization Using AI

Abstract

Academia and businesses have extensively studied portfolio decision-making under uncertainty and risk. Prediction-based portfolio models usually forecast either the portfolio's future return or risk. On the other hand, portfolio optimization models are built upon statistically defined risk measures or returns using historical scenarios. However, multiple reasons are leading even advanced models away from good performance on both topics, mainly due to the high stochastic market behaviors. However, artificial intelligence and risk management models have the potential to interact with each other and achieve better outcomes in making better decisions. In our study, we assumed that if portfolios can be optimized upon profitable prediction results, the outcome will be beneficial from the combination of machine learning prediction and risk minimization portfolio optimization. In addition, it is believed that stock movements are predictable using various data sources. We use open-source API and public financial databases to generate prediction data. All data is processed through machine learning models, including SVM, Deep Neural Network, and Long Short Term Memory. Different metrics, including accuracy, Cross Entropy Loss, etc, are used to measure the prediction performance. The prediction results are selected and applied to the risk optimization model to obtain an optimal asset allocation strategy. The portfolio risk will be optimized over the conditional value at risk, which could be calculated using either historical scenarios or predictions-based scenarios. Sets of trading simulations have been devised to evaluate the performance of the proposed models under the condition of dynamic asset allocation. In addition, a Broker-Investor competition problem is formulated using bilevel architecture. The broker and investor are aligned by returns(lower level) while competing on risk(upper level). A bilevel optimization model fits well in simulating the situation and provides a novel view of risk and return management. From our case study, we achieve several conclusions: (1) the prediction accuracy has a significant impact on portfolio outcomes; (2) the prediction system, along with the optimization model, leads to better and more profitable portfolio outcomes. (3) bilevel portfolio risk-return framework shows its potential to model a realistic business situation and provide novel risk-return research paths in the future.