This thesis presents a comprehensive quantitative analysis of traditional yield curve models and advanced deep learning architectures, with a special focus on the Nelson-Siegel-Svensson model and Long Short-Term Memory networks. The objective is to rigorously evaluate and benchmark these models with both in-sample and out-of-sample forecasts, placing a strong emphasis on their applicability in simulating different interest rate evolution scenarios. This research is propelled by the primary aim of developing a robust Earn At Risk tool for forecasting the evolution of Corporate Bank portfolios of Non-maturity deposits,notably susceptible to interest rate trends. The Nelson-Siegel model, a cornerstone of this research, is well-known for its efficacy in smoothly fitting yield curves at a fixed time, employing parameters that capture the curve’s level, slope, and curvature dynamics. The Dynamic Nelson-Siegel model extends this by incorporating time variability, making it more suitable for dynamic financial environments. In contrast, the LSTM network, represents the forefront of deep learning technologies, is explored for its capability in capturing complex, long-term dependencies in multivariate time series data, which is crucial for accurate yield curves prediction. A crucial aspect of the research involves rigorous calibration of these models, aiming to minimize the discrepancies between theoretical yield curves and those observed in the market, ensuring that the models reflect real-world dynamics accurately. This indispensable process is key for the precise prediction of future values and constitutes the groundwork for the Earn At Risk tool's development. The tool aims to provide a reliable estimate or a range of future values for interest rates, enabling effective hedging strategies and optimization of the bank's Net Interest Income. This study also delves into various other relevant topics, such as Vector Autoregressive models, Ornstein-Uhlenbeck processes, and stochastic volatility implemented in simulations. The study investigates the concept of time-varying factor loadings for Dynamic Nelson Siegel model and employs advanced techniques like Kalman filtering and Differential Evolution for model parameters calibration, while Markov Chain Monte Carlo and Hamiltonian Monte Carlo have been used to calibrate the parameters of Bayesian VAR simulations. The empirical analysis entails a comprehensive comparative assessment of the Nelson-Siegel-Svensson model and LSTM networks, contrasted against traditional benchmarks RW and the static NS. This comparison is critical in determining the relative efficacy of these advanced models in real-world financial forecasting. The study evaluates the models' performance in various forecasting horizons, focusing on their accuracy and computational efficiency. In conclusion, the research paves the way for future development of sophisticated tools for risk management and economic forecasting in the banking sector.
This thesis presents a comprehensive quantitative analysis of traditional yield curve models and advanced deep learning architectures, with a special focus on the Nelson-Siegel-Svensson model and Long Short-Term Memory networks. The objective is to rigorously evaluate and benchmark these models with both in-sample and out-of-sample forecasts, placing a strong emphasis on their applicability in simulating different interest rate evolution scenarios. This research is propelled by the primary aim of developing a robust Earn At Risk tool for forecasting the evolution of Corporate Bank portfolios of Non-maturity deposits,notably susceptible to interest rate trends. The Nelson-Siegel model, a cornerstone of this research, is well-known for its efficacy in smoothly fitting yield curves at a fixed time, employing parameters that capture the curve’s level, slope, and curvature dynamics. The Dynamic Nelson-Siegel model extends this by incorporating time variability, making it more suitable for dynamic financial environments. In contrast, the LSTM network, represents the forefront of deep learning technologies, is explored for its capability in capturing complex, long-term dependencies in multivariate time series data, which is crucial for accurate yield curves prediction. A crucial aspect of the research involves rigorous calibration of these models, aiming to minimize the discrepancies between theoretical yield curves and those observed in the market, ensuring that the models reflect real-world dynamics accurately. This indispensable process is key for the precise prediction of future values and constitutes the groundwork for the Earn At Risk tool's development. The tool aims to provide a reliable estimate or a range of future values for interest rates, enabling effective hedging strategies and optimization of the bank's Net Interest Income. This study also delves into various other relevant topics, such as Vector Autoregressive models, Ornstein-Uhlenbeck processes, and stochastic volatility implemented in simulations. The study investigates the concept of time-varying factor loadings for Dynamic Nelson Siegel model and employs advanced techniques like Kalman filtering and Differential Evolution for model parameters calibration, while Markov Chain Monte Carlo and Hamiltonian Monte Carlo have been used to calibrate the parameters of Bayesian VAR simulations. The empirical analysis entails a comprehensive comparative assessment of the Nelson-Siegel-Svensson model and LSTM networks, contrasted against traditional benchmarks RW and the static NS. This comparison is critical in determining the relative efficacy of these advanced models in real-world financial forecasting. The study evaluates the models' performance in various forecasting horizons, focusing on their accuracy and computational efficiency. In conclusion, the research paves the way for future development of sophisticated tools for risk management and economic forecasting in the banking sector.
Deep Learning in Banking Risk Management: A Comparative Assessment of LSTM Network and Nelson-Siegel Stochastic Modeling
MARTEMUCCI, WALTER
2022/2023
Abstract
This thesis presents a comprehensive quantitative analysis of traditional yield curve models and advanced deep learning architectures, with a special focus on the Nelson-Siegel-Svensson model and Long Short-Term Memory networks. The objective is to rigorously evaluate and benchmark these models with both in-sample and out-of-sample forecasts, placing a strong emphasis on their applicability in simulating different interest rate evolution scenarios. This research is propelled by the primary aim of developing a robust Earn At Risk tool for forecasting the evolution of Corporate Bank portfolios of Non-maturity deposits,notably susceptible to interest rate trends. The Nelson-Siegel model, a cornerstone of this research, is well-known for its efficacy in smoothly fitting yield curves at a fixed time, employing parameters that capture the curve’s level, slope, and curvature dynamics. The Dynamic Nelson-Siegel model extends this by incorporating time variability, making it more suitable for dynamic financial environments. In contrast, the LSTM network, represents the forefront of deep learning technologies, is explored for its capability in capturing complex, long-term dependencies in multivariate time series data, which is crucial for accurate yield curves prediction. A crucial aspect of the research involves rigorous calibration of these models, aiming to minimize the discrepancies between theoretical yield curves and those observed in the market, ensuring that the models reflect real-world dynamics accurately. This indispensable process is key for the precise prediction of future values and constitutes the groundwork for the Earn At Risk tool's development. The tool aims to provide a reliable estimate or a range of future values for interest rates, enabling effective hedging strategies and optimization of the bank's Net Interest Income. This study also delves into various other relevant topics, such as Vector Autoregressive models, Ornstein-Uhlenbeck processes, and stochastic volatility implemented in simulations. The study investigates the concept of time-varying factor loadings for Dynamic Nelson Siegel model and employs advanced techniques like Kalman filtering and Differential Evolution for model parameters calibration, while Markov Chain Monte Carlo and Hamiltonian Monte Carlo have been used to calibrate the parameters of Bayesian VAR simulations. The empirical analysis entails a comprehensive comparative assessment of the Nelson-Siegel-Svensson model and LSTM networks, contrasted against traditional benchmarks RW and the static NS. This comparison is critical in determining the relative efficacy of these advanced models in real-world financial forecasting. The study evaluates the models' performance in various forecasting horizons, focusing on their accuracy and computational efficiency. In conclusion, the research paves the way for future development of sophisticated tools for risk management and economic forecasting in the banking sector.File | Dimensione | Formato | |
---|---|---|---|
martemucci_walter.pdf
accesso riservato
Dimensione
8.82 MB
Formato
Adobe PDF
|
8.82 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/59325