A novel approach for optimal portfolio allocation: Feasible market factor estimation

This study gives the answer to the factor selection problem existing within the factor based covariance estimation framework. First, we illustrate the fact that the sample covariance estimation of the large scaled covariance matrices deteriorates the performance of the optimal asset allocation compared to naive diversification. Employing factors instead of assets into covariance estimation plays heroic role to lower estimation error and keep optimal asset allocation strategies alive for both asset management industry and the academic universe. However, choosing the optimal number of factors to reduce the estimation error to the rock bottom while not losing required information to estimate covariance profile of the asset universe, is still either sample size or time period dependent. To eliminate factor selection problem, we proposes a novel factor-based covariance estimation method, "Feasible Market Factor (FMF)" estimation, which generates a single artificial factor representing the sum of the individual variances of the observed factor universe. Under various constraint setting we compare the performance of the proposed approach to PCA, single and multi-factor and sample covariance estimation methods in terms of portfolio's performance ratios. Empirical results illustrate that FMF outperforms other factor based and sample covariance estimation methods in terms of Sharpe and Sortino ratios especially for the portfolios minimizing either variance or semi-variance. To also improve performance of the optimal portfolios maximizing return with given volatility, we implement the Autoregressive Markov Regime Switching model within FMF framework and enhance the performance of Maximum Sharpe Ratio (MSR) portfolio compared to CAPM.