Machine Learning vs Econometrics in Yield Curve Forecasting
A recent preprint on arXiv analyzes the forecasting capabilities of various methods on U.S. Treasury yield curve data, utilizing daily records from the past 47 years. This research investigates techniques that have not been previously applied to yield curve forecasting, such as ARIMA and its variations, naive benchmarks, ensemble techniques, RNNs, and several forecasting-specific transformers. Surprisingly, ARIMA and naive econometric models surpassed more sophisticated methods, raising questions about the assumption that machine learning always enhances time-series forecasting in finance. The Treasury yield curve is a crucial tool for bond market participants, who dominate over equity markets. This paper adds to the ongoing discussion regarding the influence of machine learning on financial time-series forecasting.
Key facts
- arXiv:2605.09842v1
- U.S. Treasury yield curve data
- 47 years of daily data
- Compares econometrics, classical ML, and deep learning
- Methods include ARIMA, naive benchmarks, ensemble methods, RNNs, transformers
- ARIMA and naive models outperformed complex methods
- Bond markets larger than equity markets
- Machine learning impact on time-series forecasting disputed
Entities
Institutions
- arXiv
Locations
- United States