This study proposes a novel loss function for neural network models that explores the topological structure of stock realized volatility (RV) data by adding Wasserstein Distance (WD). The study shows that the proposed loss statistically significantly improves the forecast accuracy of neural network models for magnitude-dependent error measures, for example, Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), but not necessarily for relative error measures, such as Quasi-likelihood (QLIKE). Additionally, this research provides user-friendly open-source code for researchers and practitioners to implement the proposed loss function efficiently and quickly.
