| dc.description.abstract | This master's thesis, “Evaluating the impact of sampling frequency on volatility forecast accuracy” aims to answer the main problem statement of “how do varying sampling frequencies influence the accuracy of volatility forecasts?” The thesis is driven by the idea that better prediction accuracy could result from the increased data accessibility brought about by digital progress. Some research, like the one conducted by Chan et al. (2010), indicates that higher sampling frequencies may not considerably improve forecast precision. Ewald et al. (2023) discovered that increased sampling frequencies resulted in enhanced forecasting precision compared to others. The research goal is to determine if higher sampling rates truly improve the accuracy of forecasts and if the resulting time and computational demands are justified by the increase. Several RealGARCH models using various sampling frequencies are used to assess the relationship between sampling frequency and forecasting accuracy, with data from Brent crude oil futures.
Examination of the in-sample results revealed a link between increased sampling rates and better model alignment, as evidenced by reduced AIC and BIC values and elevated log-likelihood values as sampling frequency declined. This indicates that increasing sampling frequencies may boost the precision of the model. The out-of-sample assessment showed a different situation; the connection between sampling frequency and forecasting precision was not easy to understand. Analysis of visual and regression data indicated that increased sampling rates do not always lead to lower forecast errors. The findings indicated that errors decreased when the sampling frequency was lowered. This was not in line with the belief that increasing the frequency of sampling would lead to more precise predictions. Statistical regression models revealed that only a small percentage of the variations in forecast errors could be attributed to changes in sampling frequency. Both linear and polynomial regression models yielded comparable results, indicating that sampling frequency has minimal effects on forecasting error metrics based on adjusted R² values. Negative correlation coefficients between sampling frequency and error metrics (MSE and MAE) indicated a small enhancement in forecast accuracy with decreased sampling frequency, contradicting initial assumptions. This was backed by substantial p-values, suggesting an actual, albeit small, statistical correlation.
The results differ from the common view in the literature that increasing the frequency of sampling results in improved forecasts. On the contrary, the thesis proposes that there might be a case where lower frequency t enhances precision. The connection between sampling frequency and forecast accuracy seems intricate and is affected by a variety of factors, such as the model's nature and the data set's characteristics. The evaluation pointed out possible problems with the data's reliability and the model's suitability, potentially impacting the findings' generalizability. Certain data points did not align with anticipated price levels, and the model did not get better with higher complexity, possibly because of overfitting or inadequate model specification for dealing with the detailed data.
This thesis highlights the importance of carefully weighing the pros and cons of frequent data collection when predicting volatility. It paves the way for additional studies on the most effective sampling frequencies for diverse markets and asset categories, prompting more extensive testing in different contexts to gain a deeper insight into financial market dynamics by, for example, adding more variables to the models. The results could be particularly beneficial for non-professional traders and researchers who are dedicated to improving the precision of financial models. | |
