Loading...

Hybrid LSTM-KAN variants with Quantile Loss for Value-at-Risk Forecasting

Javadi Rad, Abolfazl | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58463 (44)
  4. University: Sharif University of Technology
  5. Department: Management and Economics
  6. Advisor(s): Zamani, Shiva; Arian, Hamid Reza
  7. Abstract:
  8. This study proposes a novel framework for Value-at-Risk (VaR) forecasting by integrating Long Short-Term Memory (LSTM) networks with recently developed Kolmogorov-Arnold Network (KAN) variants, trained directly on a quantile loss objective. We benchmark the proposed models, including an enhanced LSTM-MultKAN variant, against a standard LSTM-Multi-Layer Perceptron (MLP) and a suite of ARMA-GARCH models. Unlike traditional approaches that rely on information criteria, we optimize the specifications of the GARCH benchmarks using a machine-learning-based validation based on the out-of-sample performance as measured by quantile loss, ensuring a fair comparison. Furthermore, to address the interpretability challenge of neural networks, a deep interpretability analysis is conducted, employing SHapley Additive exPlanations (SHAP) alongside the inherent transparency of the KAN architecture. The LSTM-KAN variants show statistically significant improvements in predictive accuracy and robustness, establishing a clear performance advantage over traditional GARCH models. Within the network architectures, the LSTM-MultKAN variant demonstrates the strongest performance, outperforming the standard LSTM-MLP benchmark. Our interpretability analysis reveals that this superior performance is not arbitrary; the network models learn to identify well-documented financial phenomena, such as volatility clustering and the leverage effect. By achieving both accuracy and interpretability, this study presents a powerful and trustworthy VaR forecasting tool for researchers and risk managers
  9. Keywords:
  10. Value at Risk ; Kolmogorov-Arnold Networks ; Financial Risk Management ; Long Short Term Memory (LSTM) ; Interpretability ; Autoregressive-Moving Average Models ; General Autoregressive Conditional Heteroskedastic (GARCH) ; Autoregressive Moving Average-Generalized Autoregressive Conditional Heteroskedasticity (ARMA-GARCH) ; Hybrid Deep Learning Models

 Digital Object List

 Bookmark

...see more