CAN MARKET STATE DRIVEN DYNAMIC REALLOCATION STRATEGIES GUIDED BY HIDDEN MARKOV MODELS OUTPERFORM THE MARKET?

Abstract

Dynamic portfolio allocation strategies seek to leverage cross sectional and longitudinal heterogeneities in assets performances to generate asset return. This paper examines the feasibility of outperforming the SPY by implementing the SPY-MM strategy, allocating between the SPY and the money market, and the sector rotation strategy, allocating between defensive and cyclical sectors of the S&P 500. I trained hidden Markov models (HMM) using the VIX and OAS (credit spread) and switched between different assets according to the models’ estimated one-step ahead probability of low or high state. I explored variations in the strategies by empirically testing sets of tuning parameters including retraining frequency, certainty threshold for switching, and the length of rolling window applied to retraining data. The results show that when training the HMMs semiannually using VIX, both the SPY-MM strategy and the sector rotation strategy can achieve Sharpe ratios between 0.56 and 0.60, which are higher than the benchmark of 0.55. The HMMs trained on OAS have mixed performances. When applied to the SPY-MM strategy, it was successful and achieved a maximum Sharp ratio of 0.66 with semi-annual retraining. However, the sector rotation strategy, under all combinations of tuning parameters, cannot outperform the benchmark when using the OAS-trained HMM. Finally, across the strategies, the optimal specifications of tuning parameters were semi-annual retraining, with either a 10% or 90% certainty threshold and either a 3 year or 10 year rolling window.