A Return Prediction-based Investment with Particle Filtering and Anomaly Detection
This paper proposes a new stochastic volatility model with time-varying expected return, which enables us to predict returns based on exponential moving averages of the past returns frequently used in practice. Particularly, exploiting a particle filter in a self-organizing state space framework, we demonstrate that a simple return prediction- based strategy is superior to well-known strategies such as equally-weighted, minimum-variance and risk parity portfolios, which do not depend on return prediction. In addition, we develop three types of anomaly detectors that are easily implemented in the algorithm of the particle filter and apply them to investment decision. As a result, our model robustly outperforms the exponential moving average. Our dataset is monthly total returns of global assets such as stocks, bonds and REITs, and investment performances are evaluated with various statistics such as compound returns, Sharpe ratios, Sortino ratios or drawdowns.