Enhancing value-at-risk forecasting using adaptive models, deep learning, and text mining techniques on stock market indices

Financial risk can be measured through various methods, with Value at Risk (VaR) being a key metric to assess the likelihood of adverse events. The Basel Accords encourage prudent market risk predictions, avoiding overly conservative or permissive forecasts. Inaccurate models may lead to higher capital reserves, limiting investment or business expansion, and may incur penalties if they fail to capture market risks adequately. The challenges of financial risk measurement have grown since the turbulence of COVID-19, as conventional models struggle to reflect real risks. This research aims to improve VaR forecasting accuracy using parametric methods by proposing diverse models and approaches. It focuses on three critical factors, derived from financial asset characteristics, that influence VaR forecasting, explored through three empirical studies aimed at achieving more accurate and effective predictions. Empirical Study 1 evaluates the performance of the smooth transition exponential smoothing (STES) method, which employs adaptive, time-varying smoothing parameters, against conventional parametric, non-parametric, and semi-parametric methods across various confidence levels in the stock market. Empirical Study 2 investigates the integration of traditional time series models with deep learning (DL) techniques, specifically Long Short-Term Memory Network (LSTM), to enhance VaR forecasting by capturing complex financial data features. Empirical Study 3 focuses on improving VaR model precision through text mining techniques, such as sentiment analysis, leveraging textual information and emphasizing three key aspects. The key findings are as follows: 1) STES methods do not outperform traditional parametric, non-parametric, and semi-parametric approaches across confidence levels. 2) DL hybrid models, combining conventional techniques with deep learning, achieve superior performance compared to benchmark VaR models. 3) Incorporating textual data in DL hybrid models significantly improves prediction accuracy, with feature analysis highlighting the influence of topics such as "Earning Mix" (shaped by earnings, revenue, and trade activities) and "Market Movements" (driven by trends impacting close values, particularly in U.S. markets). Additionally, a longer time horizon is required for emotional polarity scores from news sentiment to affect VaR estimates, emphasizing the importance of long-memory characteristics in sentiment analysis.

Text SPE 2024 44 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) Text SPE 2024 44.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (5MB)

Actions (login required)

View Item