Sentiment-aware realized volatility forecasting in financial markets
This study evaluates the effects of -news sentiment data on volatility prediction, and how further boost the capabilities of the garch model. Also taken into account are the effect of the attention (news flow), and a macroeconomic variable (EPU). We develop a parsimonious GARCH(1,1)-X framework with...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/445183 |
| Acceso en línea: | https://hdl.handle.net/2117/445183 |
| Access Level: | acceso abierto |
| Palabra clave: | Econometrics Emotions Artificial Intelligence Volatility prediction Sentiment Econometria Emocions Intel·ligència artificial Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial |
| Sumario: | This study evaluates the effects of -news sentiment data on volatility prediction, and how further boost the capabilities of the garch model. Also taken into account are the effect of the attention (news flow), and a macroeconomic variable (EPU). We develop a parsimonious GARCH(1,1)-X framework with sign-separated sen- timent and optional attention/macro terms, and we compare hundreds of specifi- cations across four U.S. large companies in different sectors. Main findings. 1. Sentiment materially improves volatility forecasts. Across firms, adding polarity shows improved models over a price-only GARCH baseline. One thing of note to notice was how for out data the positive-polarity term add more importance than the negative one, even though the papers men- tion the contrary. This difference could be due to the difference datasets used and the methodology followed. For a general prediction using the ('unique','WeightedMean') with the polarityRob with different EMA can wield good results for most companies. 2. Volume of news Attention features (volume/RCV) yield small average gains, without the sentiment they are not worth adding, and even with the polarity the model did not show a significant improvement, which again does not coincide with what the papers mentioned. 3. Macro is stock-specific. EPU shows the expected correlation with realized volatility but delivers limited incremental predictive power. 4. Diagnostics. Standardized residuals exhibit no residual autocorrelation, indicating that mean/variance dynamics are well captured. QQ-plots show fat tails, so Gaussian errors understate extreme moves. This means sentiment alone is not enough to have completely reliable models 56 Limitations. We focus on a daily horizon, Gaussian innovations, and an additive variance link (GARCH-X). These choices emphasize interpretability, testing and tractable comparison across many aggregation/EMA variants, but might miss the possible details that could be gained from testing other models. Future work. (i) Test with other garch model variants. (ii) Test with different news sources to see the effect of the volume of news. (iii) Test with other macroe- conomic variables. (iv) Test the effect of the sentiment for smaller companies, which present higher volatilities. |
|---|