<![CDATA[The highly volatile nature of cryptocurrency prices often causes conventional predictive models to fail in capturing complex nonlinear patterns. This study integrates the Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS) deep learning model with nonparametric Historical Simulation Value-at-Risk (VaR) method for price forecasting and risk analysis. Using univariate data on daily Ethereum closing prices from January 1, 2021, to January 31, 2025 (N = 1,491 observations), the out-of-sample evaluation was executed using a rolling cross-validation scheme initiated testing from a cut-off point in April 2024 through December 2024, where each evaluation window was set for the next 30 days. The research results show that the N-HiTS model can predict price dynamics with high accuracy, achieving an MAPE of 3.25%, an MAE of 107.825, an RMSE of 136.83, and directional accuracy of 48.28%. Risk analysis using historical simulation yielded a VaR of -6.23% at a 95% confidence level.]]>
Copyrights © 2026
