In this article, classical machine learning methods, econometric methods, and deep machine learning methods were used as a way of determining if an out of the box multivariate time series transformer would be more effective at predicting the stock volatility index. Since closely and consistently predicting the stock volatility index is more beneficial for forecasting the stock volatility index the models will be evaluated on the RMSE of each model. The MTST model performed better than the AR-GARCH and LGBM models even without the use of a hybridized model that have been generally used by other state of the art stock volatility index models. (Hochreiter & Schmidhube, 1997; Kim & Won, 2018; Ramos-Pérez et al., 2021)

Due to the time dependent nature of the data the LGBM was at an inherent disadvantage but due to its previous success in volatility forecasting the lackluster performance indicates that more preprocessing and feature selection is required to make the LSTM a viable.

It is worth noting that although the MTST model performed better than the LGBM and AR-GARCH models it required a much longer period to train compared to the other two models. Additionally, given how much effort generally goes into refining Transformer models it may be more difficult to generalize the MTST model compared to AR-GARCH and the LGBM.

Despite the limitation of the MTST the success of the out of the box MTST suggests the MTST may be viable as a general tool for future stock volatility forecasting without requiring the extensive resources for model refinement. Since current models are highly technical to work with and often lack documentation, creating software packages that make MTST more readily available for use as an unboxed model would make deep learning for stock volatility index predictions a viable option for individuals who are not machine learning engineers.