Comparative Analysis of Machine Learning and Econometrics for Yield Curve Forecasting
A new academic paper published on arXiv presents a comprehensive comparative analysis of forecasting methods for the U.S. Treasury yield curve, utilizing daily data spanning 47 years. The study addresses the ongoing debate regarding the efficacy of machine learning in financial time-series forecasting by evaluating econometric models, classical machine learning algorithms, and deep learning techniques. Tested methods include Autoregressive Integrated Moving Average (ARIMA) models, naive benchmarks, ensemble methods, Recurrent Neural Networks (RNNs), and transformer-based architectures like TimeGPT. The results indicate that traditional ARIMA and naive econometric models generally outperform more complex machine learning approaches, with the exception of one specific time block. Among the machine learning techniques evaluated, TimeGPT, LightGBM (LGBM), and RNNs demonstrated the strongest performance. Additionally, the research investigates the impact of data stationarity on deep learning model inputs. This study provides critical insights for bond market participants, highlighting that while advanced AI tools show promise, established econometric methods remain highly competitive in predicting yield curves, which are fundamental to the broader fixed-income market.
Wire timeline
Comparative Analysis of Machine Learning and Econometrics for Yield Curve Forecasting
A new academic paper published on arXiv presents a comprehensive comparative analysis of forecasting methods for the U.S. Treasury yield curve, utilizing daily data spanning 47 years. The study addresses the ongoing debate regarding the efficacy of machine learning in financial time-series forecasting by evaluating econometric models, classical machine learning algorithms, and deep learning techniques. Tested methods include Autoregressive Integrated Moving Average (ARIMA) models, naive benchmarks, ensemble methods, Recurrent Neural Networks (RNNs), and transformer-based architectures like TimeGPT. The results indicate that traditional ARIMA and naive econometric models generally outperform more complex machine learning approaches, with the exception of one specific time block. Among the machine learning techniques evaluated, TimeGPT, LightGBM (LGBM), and RNNs demonstrated the strongest performance. Additionally, the research investigates the impact of data stationarity on deep learning model inputs. This study provides critical insights for bond market participants, highlighting that while advanced AI tools show promise, established econometric methods remain highly competitive in predicting yield curves, which are fundamental to the broader fixed-income market.
cs.AI updates on arXiv.org