Využití metod předvídání volatility při obchodování vybraných opčních strategií

This master's thesis examines the use of volatility forecasting models in the implementation of option strategies on the Canadian S&P/TSX 60 index. The main objective is to systematically compare the predictive performance of classical econometric models, machine learning methods and Bayesian approaches in both prediction accuracy and practical application in real option trading. The analysis follows a three-stage framework: model training on historical data (January 2010 – June 2022), out-of-sample evaluation with 30-day ahead forecasts (July 2022 – December 2022), and practical implementation of four option strategies during calendar year 2023. From a total of eleven models, six best-performing models were selected for application in option strategies - APARCH, HAR, feedforward neural network, LSTM, Bayesian GARCH, and Bayesian stochastic volatility. Trading signals were generated using the quintile approach, and all positions were held to expiry with both hedged and unhedged variants. A constant short-only straddle strategy served as benchmark for evaluating model-based approaches. Results show that all four strategies outperform the benchmark loss of -17.65 %. From all examined strategies, the hedged strangle achieves the highest return of 41.71 %, with the hedged straddle on the second place with a return of 31.78 %. Although the effect of delta hedging proves to be substantial, it is clearly dependent on the specific strategy. While hedging increases straddle performance by 16.40 percentage points, it leads to a reduction of 2.25 points in the case of the iron butterfly. In terms of model performance, classical econometric approaches, particularly APARCH and BGARCH, deliver the strongest trading results, which contrasts with the often assumed dominance of machine learning models. While the APARCH model shows a return of 14.74 % in the hedged strangle strategy, more complex approaches such as LSTM and Bayesian stochastic volatility show consistently weaker results across strategies. Overall, the results illustrate that higher prediction accuracy does not automatically mean higher trading profitability. Instead, the interaction between the chosen model, the trading strategy, and the hedging design plays an important role.