Monitoring Volatility Change for Time Series Based on Support Vector Regression.

Area of Science:

• Quantitative Finance
• Time Series Analysis
• Machine Learning

Background:

• Financial time series exhibit heteroscedastic conditional volatilities, posing challenges for anomaly detection.
• Existing monitoring methods may not effectively capture dynamic changes in volatility.
• Online monitoring is crucial for timely detection of significant shifts in financial markets.

Purpose of the Study:

• To propose an online monitoring process for detecting anomalies in time series with heteroscedastic volatilities.
• To combine the cumulative sum (CUSUM) method with support vector regression (SVR) for enhanced anomaly detection.
• To optimize the model's tuning parameters using particle swarm optimization (PSO).

Main Methods:

• Utilized a cumulative sum (CUSUM) approach integrated with support vector regression (SVR).
• Employed particle swarm optimization (PSO) for optimal selection of tuning parameters.
• Conducted Monte Carlo simulations to validate the proposed method's performance.

Main Results:

• The proposed CUSUM-SVR method effectively detects significant changes in financial time series volatility.
• Monte Carlo simulations demonstrated the validity and robustness of the developed monitoring process.
• The model showed versatility in real-world applications across different financial indices and stocks.

Conclusions:

• The integrated CUSUM-SVR approach provides a powerful tool for online anomaly detection in financial time series.
• Optimal parameter tuning via PSO enhances the method's accuracy and reliability.
• The model's applicability is confirmed through analyses of S&P 500, KOSPI, and Microsoft stock data.