Change Point Test for the Conditional Mean of Time Series of Counts Based on Support Vector Regression.

Area of Science:

• Statistics
• Machine Learning
• Financial Econometrics

Background:

• Time series analysis of count data presents unique challenges.
• Accurate prediction and change point detection are crucial in financial modeling.
• Existing methods may not fully capture the complexities of count-based financial time series.

Purpose of the Study:

• To apply and evaluate Support Vector Regression (SVR) and Twin SVR (TSVR) for time series of counts.
• To optimize hyper parameters of SVR and TSVR using Particle Swarm Optimization (PSO).
• To utilize these optimized models for change point detection in financial data.

Main Methods:

• Hyperparameter tuning of SVR and TSVR using Particle Swarm Optimization (PSO).
• Forecasting using Integer-Valued Generalized Autoregressive Conditional Heteroskedasticity (INGARCH) models.
• Change point detection via the Cumulative Sum (CUSUM) test on model residuals.

Main Results:

• Demonstrated validity of PSO-SVR and PSO-TSVR with linear and nonlinear INGARCH models through Monte Carlo simulations.
• Successful application to real-world financial data, specifically return times of extreme events in stock prices.
• The CUSUM test effectively identified change points using residuals from the proposed methods.

Conclusions:

• PSO-SVR and PSO-TSVR offer a robust framework for time series count data analysis.
• These methods provide a valuable tool for change point detection in financial econometrics.
• The study highlights the practical applicability of advanced machine learning techniques in finance.