Online intra-day portfolio optimization using regime based models

Sammanfattning: In this thesis model predictive control (MPC) is used to dynamically optimize a portfolio where the data is sampled every 5 minutes. Previous research has shown how MPC optimization applied to daily sampled financial data can generate a portfolio that exceeds the value of standard portfolio strategies such as Strategic asset allocation. MPC has been found to have a computational advantage when return forecasts are updated every time a new observation becomes available. A two-state Hidden Markov Model with time varying parameters is used to forecast the financial return of a market index. The portfolio optimization is performed using both single period and multi-period forecasts where the only other asset is a zero interest rate cash account. Transaction costs are included to better reflect market conditions and to address estimation errors in the forecasts. The MPC portfolios are found to outperform a buy and hold strategy in the market index, displaying both higher returns and lower risk. The multi-period portfolios display lower returns and similar risk to the single period portfolio while having a smaller turnover. This led to the conclusion that the two-state Gaussian HMM provides sub par multi-period forecasts on the 5 minute sampled market index. The forecasting method is found to be very sensitive to the manual choice of hyperparameters.