Authors: Oksana Anatolyevna Malysheva
Abstract: The new financial market environment after 2020, due to the COVID-19 shock, unprecedented monetary interventions, and increased macroeconomic uncertainty has cast new doubt on the reliability and persistence of the traditional models of asset pricing factors. Although classical factor models have traditionally been the basis of portfolio construction and the management of risk, there has been mounting evidence that factor stability can be lost in the face of structural regime changes. The paper reexamines previous post-2020 period factor models with specific focus on occupation of factor permanency and incremental importance of machine learning methods in explaining and predicting investment outcomes. The main aim of this paper is to evaluate whether the traditional risk factors hold their values and are economically significant beyond 2020 and to test the possibility of machine learning-based methods to predict better than traditional linear models to forecast the portfolio performance and results. The study builds standard factor portfolios with a wide equity universe over the post-2020 time sample, and it compares their performance to machine learning-based models that help to identify the nonlinear links and time-varying interactions between firm characteristics. The analysis methodology will be a combination of benchmark linear factor regressions with supervised machine learning algorithms, such as ensemble-based algorithms, applying consistent training and validation to these algorithms. Factor stability is determined with the help of rolling-window estimation and structural change analysis and investment performance with the help of risk-adjusted returns measures and transaction cost-adjusted portfolio performance. The results show that the stability and persistence of a number of conventional factors significantly decreased after 2020 and became more sensitive to market regimes. The models of machine learning have shown greater out-of-sample, and risk-adjusted returns, and the returns are not uniform across factors. The research provides empirical data on post-crisis factor behavior and provides a practical direction of applying machine learning in the integration of factored investment strategies in changing market conditions.
