Scenario generation for financial data: a machine learning dynamic copula approach based on realized volatility and correlation

Abstract

Portfolio optimization is a fundamental issue in quantitative finance, and scenario generation techniques play a vital role in simulating the future behavior of assets for use in allocation strategies. In the literature, various approaches exist for generating scenarios, ranging from historical observations to models predicting asset volatility. In this dissertation, we propose a novel methodology for generating discrete scenarios one day ahead, which are then used as input for portfolio allocation. Our approach employs machine learning upervised algorithms as forecasting models to predict the realized variance and intraday Kendall correlation of assets. Using these predictions, we apply a copula approach with extreme value distributions to simulate the multivariate probability distribution of the assets. Our computational experiments indicate that our approach may yield more accurate volatility and correlation forecasts, as well as better risk-reward portfolios compared to traditional literature baselines.