Yield Curve Forecasting using Machine Learning and Econometrics: A Comparative Analysis

A comprehensive study comparing machine learning, deep learning, and traditional econometric methods for forecasting U.S. Treasury yield curves reveals that classical ARIMA models and naive benchmarks generally outperform advanced algorithms, though TimeGPT and RNNs show promise among machine learning approaches. The research challenges assumptions about deep learning's universal superiority in financial forecasting.

This research addresses a fundamental question in quantitative finance: whether sophisticated machine learning and deep learning models can improve upon established econometric methods for predicting Treasury yield curves. The study's scope is substantial, analyzing 47 years of daily U.S. Treasury data across multiple algorithmic families, providing empirical evidence that contradicts the widespread assumption that newer technologies automatically outperform traditional approaches. The finding that ARIMA and simple econometric models maintain competitive or superior performance is particularly significant given the massive investments in AI infrastructure by financial institutions.

The Treasury yield curve serves as a critical indicator across global financial markets, influencing bond pricing, mortgage rates, and broader economic expectations. Its predictability directly affects trillions of dollars in fixed-income markets and shapes monetary policy interpretations. The dominance of traditional methods suggests that the yield curve's dynamics may be fundamentally driven by structural economic factors that ARIMA-type models capture effectively, rather than complex non-linear patterns that deep learning excels at discovering.

For institutional investors and quantitative traders, this research validates continued reliance on econometric frameworks while suggesting selective application of modern techniques. The strong performance of TimeGPT and RNNs in specific contexts indicates that hybrid approaches combining classical and machine learning methods warrant further exploration. Market participants should reconsider costly deep learning implementations that fail to deliver measurable forecasting improvements over simpler alternatives, particularly when transaction costs and model complexity are factored into real-world deployment decisions.

• →Traditional ARIMA models outperform most machine learning and deep learning approaches for Treasury yield curve forecasting
• →Among advanced methods, TimeGPT, LGBM, and RNNs demonstrate the strongest performance relative to other ML algorithms
• →The choice between stationary and nonstationary data inputs significantly affects deep learning model effectiveness
• →Empirical evidence challenges the assumption that newer AI technologies universally improve financial time-series forecasting
• →Bond market participants may achieve superior cost-adjusted returns by prioritizing econometric methods over expensive deep learning solutions