A1078
Title: Portfolio optimization with multiple covariance models: Forecast combination approaches
Authors: Alexandre Rubesam - IESEG School of Management (France) [presenting]
Andre Portela Santos - CUNEF Universidad (Spain)
Carlos Trucios - University of Campinas (Brazil)
Luiz Koodi Hotta - University of Campinas (Brazil)
Abstract: Although recent advances have made it feasible to estimate dynamic covariance matrices in high dimensions, portfolio managers face significant model risk when selecting a single forecast approach. The aim is to propose several forecast combination methods to mitigate model uncertainty and the risk of relative underperformance. These include distance-based approaches that penalize models producing forecasts that diverge excessively from others, as well as two optimization-based methods, which balance variance reduction with portfolio stability relative to individual models. Using daily U.S. stock data from 1980 to 2022, ten individual covariance forecasting models and several combination strategies are evaluated, constructing long-only minimum variance portfolios for universes of up to 1,000 stocks. Forecast combinations, particularly simple ones, reduce model risk and deliver realized volatility close to the best individual models. However, optimization-based combinations often introduce additional turnover, potentially offsetting their benefits after transaction costs. Introducing reasonable constraints in the portfolio optimization problem, it is found that forecast combinations deliver a performance on par with the best individual forecasts.
