1. Introduction
The U.S. 10-year Treasury bond yield plays a pivotal role in financial markets as a benchmark for long-term interest rates and a barometer for macroeconomic expectations. It affects a wide array of financial instruments, including mortgage rates, corporate bonds, and equities, and serves as a critical input in monetary policy assessments, asset pricing models, and risk management strategies [1,2]. Because of its wide-reaching implications, accurately forecasting the 10-year yield is of significant interest to investors, policymakers, and financial analysts.
Changes in the 10-year Treasury yield are influenced by a variety of macroeconomic factors, most notably the federal funds rate, inflation (commonly measured via the Core Consumer Price Index), GDP growth, and government fiscal conditions such as the federal debt growth rate [3,4]. These variables collectively reflect the interplay between monetary policy, economic activity, and fiscal sustainability, making them valuable predictors in yield modeling.
Traditional approaches to modeling bond yields have relied heavily on econometric models, such as the Nelson–Siegel framework [5] and vector autoregression (VAR) models. While these models offer interpretability and theoretical grounding, they often struggle to capture the nonlinear and high-dimensional nature of modern financial data. In recent years, machine learning (ML) methods have emerged as powerful tools in financial forecasting due to their ability to model complex, nonlinear relationships and adapt to evolving patterns in large datasets [6,7].
This study aims to investigate the short-term forecasting performance of four different models, linear regression (LR), decision tree (DT), random forest (RF), and multilayer perceptron (MLP) neural networks, on the U.S. 10-year Treasury yield. By incorporating key macroeconomic indicators as input variables and comparing model performance based on statistical accuracy metrics, this research seeks to determine which modeling approach provides the most reliable forecasts. The emphasis on short-term prediction (1 to 3 months ahead) is aligned with practical decision-making needs in trading and risk management settings, where timing and precision are crucial.
Additionally, this research contributes to the growing body of literature on applying machine learning in macro-finance by focusing specifically on a single, influential yield rather than the entire term structure. Through this approach, this study offers actionable insights for both academic researchers and financial practitioners aiming to enhance predictive capabilities using data-driven methods.
2. Literature Review
Forecasting government bond yields, particularly the U.S. 10-year Treasury yield, has been a long-standing topic of interest in both macroeconomics and financial research. Numerous models and approaches have been proposed, ranging from traditional econometric frameworks to recent applications of machine learning (ML) and artificial intelligence.
2.1. Traditional Econometric Models
The Nelson–Siegel (NS) model, introduced by Nelson and Siegel [5], has been one of the most influential tools for modeling the term structure of interest rates. Its extension by Diebold and Li [1] introduced a dynamic component that improved its forecasting capabilities. These models typically represent the entire yield curve using a small number of parameters that capture the level, slope, and curvature of interest rates. While effective for curve fitting and macro-finance interpretation, these models are less suited for short-term prediction of a single yield like the 10-year bond.
Linear regression (LR) models have also been widely used to understand the relationship between macroeconomic indicators and bond yields. Studies such as [4,8] examined the predictive power of variables such as inflation, short-term interest rates, and GDP growth on long-term yields. Although LR models are interpretable and statistically tractable, they are limited in capturing nonlinear dynamics present in real-world financial data.
2.2. Machine Learning Approaches
Recent advances in ML have enabled researchers to explore more flexible, data-driven methods for financial forecasting. Decision trees and ensemble methods like random forests (RFs) have been successfully applied to bond and stock market prediction due to their ability to model complex, nonlinear relationships and interactions among variables [6,9]. RF, in particular, reduces overfitting risks common in single decision trees by averaging multiple models trained on different data subsets.
Neural networks, especially multilayer perceptron (MLP) models, have also demonstrated promise in financial forecasting tasks, including yield prediction [10,11]. These models can capture highly nonlinear and high-dimensional relationships but require careful tuning and large datasets. While they are powerful, their lack of interpretability, often referred to as the “black box” problem, can be a limitation for economic policy applications.
Several studies have compared traditional statistical models with ML models in forecasting yields or macroeconomic indicators. For instance, Ghosh and Chaudhuri [12] found that ML methods consistently outperformed traditional regression models in terms of predictive accuracy for interest rate forecasting. Similarly, Zhang [13] demonstrated that ensemble learning models provided more robust yield predictions, particularly during volatile economic conditions.
2.3. Gaps in the Literature
Despite the progress in modeling the yield curve, there is limited research focused specifically on short-term forecasting of the 10-year Treasury yield using multiple ML techniques in direct comparison. Most ML applications tend to focus on forecasting the full-term structure or use highly complex deep learning models such as LSTM (Long Short-Term Memory), which require extensive computational resources and are difficult to interpret [14]. This research seeks to fill that gap by evaluating and comparing the predictive performance of interpretable and widely used ML models, LR, DT, RF, and MLP, on a key individual yield over a short time horizon.
2.4. Theoretical Framework Linking Macroeconomic Variables to Treasury Yields
The selection of macroeconomic variables in this study is grounded in well-established economic theories that explain the behavior of interest rates and bond yields in response to changes in monetary and fiscal policy, inflation, and economic growth.
Federal Funds Rate (DFF):
According to the Expectations Theory of the Term Structure of Interest Rates, long-term interest rates (such as the 10-year Treasury yield) reflect expectations of future short-term rates. Therefore, changes in the effective federal funds rate, a primary policy tool of the Federal Reserve, signal anticipated shifts in future monetary policy and directly influence the yield curve [3].
2. Core Consumer Price Index (CPILFESL):
Inflation expectations are central to interest rate determination under the Fisher Effect, which posits that nominal yields adjust in response to expected inflation to preserve real returns. Core CPI, by excluding volatile food and energy prices, provides a stable measure of underlying inflation and is closely monitored by central banks in policy decisions [15].
3. Real GDP Growth (GDP_growth):
Strong economic growth typically increases the demand for credit and raises expectations of inflationary pressure, both of which contribute to higher bond yields. This relationship is supported by IS-LM and Taylor Rule models, which show that higher output levels lead to upward pressure on interest rates as central banks react to overheating economies [2].
4. Federal Government Debt Growth (GFDEBTN_growth):
The impact of fiscal policy on interest rates can be interpreted through the Ricardian Equivalence and Crowding-Out Hypothesis. While Ricardian theory suggests debt issuance has neutral long-term effects if consumers anticipate future taxation, the crowding-out effect argues that higher government borrowing increases competition for loanable funds, pushing yields higher [4,16].
Together, these variables capture the core channels through which monetary, fiscal, and macroeconomic conditions influence bond markets. Including them in a machine learning framework allows for empirical testing of these theoretical linkages in a high-dimensional, nonlinear setting.
3. Methodology
3.1. Research Design
This study adopts a quantitative modeling approach to predict the short-term movement of the U.S. 10-year Treasury bond yield using multiple machine learning (ML) techniques. The research compares four predictive models, linear regression (LR), decision tree (DT), random forest (RF), and multilayer perceptron (MLP), based on their performance using historical macroeconomic indicators as predictors. The primary goal is to identify the model that offers the highest predictive accuracy for short-term yield forecasting, defined as a 1- to 3-month horizon.
3.2. Data Collection
Monthly data were collected from the Federal Reserve Economic Data (FRED) database, covering the period from January 2000 to December 2023. The dependent variable is the 10-year U.S. Treasury bond yield (FRED code: DGS10). The independent variables include the following:
Effective Federal Funds Rate (DFF): A short-term interest rate target set by the Federal Reserve, used as a proxy for U.S. monetary policy [3].
Core Consumer Price Index (CPILFESL): A version of the Consumer Price Index (CPI) that excludes food and energy prices to measure underlying inflation trends [15].
Real Gross Domestic Product Growth Rate (GDP_growth): The inflation-adjusted rate of change in the value of all goods and services produced in the U.S. economy, indicating overall economic output trends [2].
Federal Government Debt Growth Rate (GFDEBTN_growth): The percentage change in the total outstanding debt issued by the U.S. federal government, reflecting fiscal sustainability and long-term borrowing concerns [4].
The independent variables include the following:
Effective Federal Funds Rate (DFF): A short-term interest rate determined by the Federal Reserve, representing U.S. monetary policy stance.
Core Consumer Price Index (CPILFESL): A measure of core inflation that excludes food and energy prices to reduce short-term volatility.
Real Gross Domestic Product Growth Rate (GDP_growth): Calculated as the year-over-year percentage change in real GDP using the formula:
GDP_growtht = (GDPt − GDPt−12) × 100/GDPt−12
where GDPt is the real GDP in month t.Federal Government Debt Growth Rate (GFDEBTN_growth): Calculated as the year-over-year percentage change in the total federal debt outstanding:
GDPEBTN_growtht = (GDPEBTNt − GDPEBTPt−12 ) × 100/GDPEBTNt−12
where GFDEBTNt is the federal debt level at time t.These transformations ensure consistency in scale and comparability across variables and account for seasonality and long-term trends, both of which are critical for meaningful macroeconomic analysis.
3.3. Data Preprocessing
All time series were transformed into percentage changes or growth rates, where appropriate, to ensure stationarity and comparability. Missing values were handled using mean imputation, where the missing entries were replaced with the mean of the respective variable over the training period. The proportion of missing data was below 2% for all variables, and no extended sequences of missing values were observed. Since the dataset is monthly and time-ordered, temporal consistency was taken into account by ensuring that imputation was performed only within the training set to avoid data leakage. Additionally, standardization was applied to all input variables to normalize feature scales, which is essential for optimizing multilayer perceptron (MLP) performance [7].
The dataset was split into a training set (80%) and a test set (20%), preserving temporal order to avoid data leakage. This split was chosen to ensure that the models were trained on a sufficiently large portion of the historical data while retaining enough out-of-sample data for robust performance evaluation. The 80/20 division aligns with common practice in the time-series forecasting literature and balances the need for learning temporal patterns with the ability to assess generalization across future periods.
3.4. Model Specification
3.4.1. Linear Regression (LR)
A multiple linear regression model was estimated using the ordinary least squares (OLS) method, which minimizes the sum of squared residuals to obtain coefficient estimates. This approach is standard in econometric modeling and does not require hyperparameter tuning [17]. The regression was implemented using the standard LinearRegression class from the Scikit-learn library in Python 3.13.2, which computes the weights analytically based on OLS. No regularization techniques were applied in this baseline model.
3.4.2. Decision Tree (DT)
The DT model recursively partitions the input space by choosing features and thresholds that minimize the mean squared error (MSE) within each split [18]. The maximum tree depth and minimum samples per leaf node were optimized via 5-fold cross-validation to prevent overfitting. Grid search was used to test the following ranges:
max_depth: [3, 5, 10, 15]
min_samples_leaf: [1, 2, 4, 6]
The best-performing configuration was: max_depth = 10, min_samples_leaf = 2.
3.4.3. Random Forest (RF)
The RF model is an ensemble of decision trees trained on bootstrap samples with random feature selection [9]. We used 100 trees and tuned the number of features considered at each split. RF is particularly suited for capturing nonlinearities and interactions among variables. Hyperparameters were tuned over the following ranges:
n_estimators: [50, 100, 200]
max_features: [‘auto’, ‘sqrt’, ‘log2’]
max_depth: [10, 15, 20, None]
The final selected parameters were: n_estimators = 100, max_features = ‘sqrt’, max_depth = 15.
3.4.4. Multilayer Perceptron (MLP)
The MLP is a feedforward neural network consisting of one input layer, one hidden layer (with 100 neurons), and one output layer. The ReLU (Rectified Linear Unit) activation function was used, and the model was trained using the Adam optimizer with a learning rate of 0.001 [7]. Early stopping was employed to avoid overfitting. The tuning space included:
hidden_layer_sizes: [(50,), (100,), (100, 50)]
activation: [‘relu’, ‘tanh’]
learning_rate_init: [0.001, 0.01]
The final model used hidden_layer_sizes = (100,), activation = ‘relu’, and learning_rate_init = 0.001.
Early stopping and a maximum of 200 iterations were applied to mitigate overfitting.
3.5. Performance Evaluation
The models were evaluated based on three standard metrics:
Mean Squared Error (MSE): Measures the average squared difference between predicted and actual yields.
Mean Absolute Error (MAE): Captures average prediction error magnitude without emphasizing large deviations.
R-squared (R2): Indicates the proportion of variance in the target variable explained by the model.
These metrics were computed using the test set to assess each model’s out-of-sample forecasting ability [8].
4. Results
4.1. Overview
This section presents the empirical results of applying four different forecasting models, linear regression (LR), decision tree (DT), random forest (RF), and multilayer perceptron (MLP), to predict the U.S. 10-year Treasury bond yield over a short-term horizon. The models were trained using historical macroeconomic data and evaluated on their ability to accurately predict bond yields on unseen data.
4.2. Model Performance on the Training Set and Testing Set
To evaluate the predictive performance of each model, we computed three widely used error metrics on the test dataset: mean squared error (MSE), mean absolute error (MAE), and R-squared (R2). Table 1 summarizes the performance of each model on the training dataset and testing set.
Linear regression (LR) typically shows similar performance on training and test data unless it underfits the data. In contrast, decision tree (DT) and random forest (RF) models often achieve very low training error and high R2 scores, indicating potential overfitting. Meanwhile, MLP (Neural Network) models tend to perform better on training data as well, though the extent of overfitting is usually less severe and depends on factors like regularization and the training setup.
The random forest (RF) model consistently outperformed the other models across all three evaluation metrics, with the lowest MSE and MAE and the highest R2 value. These results are consistent with findings from previous studies that highlight the strength of ensemble learning methods in forecasting financial time series [6,9].
The decision tree (DT) model also showed solid performance but exhibited a higher variance compared to RF, suggesting susceptibility to overfitting. MLP achieved moderately good results but not better than RF, possibly due to limitations in training data size or insufficient hyperparameter optimization. As expected, LR performed the weakest due to its linear constraints, which limits its ability to capture the complex relationships between macroeconomic variables and bond yields [8].
4.3. Forecast Visualization
Each model employs a unique approach to learning patterns from the data. Linear regression relies on a simple linear relationship between input features and the target variable. Decision tree uses a hierarchical, rule-based structure to split data based on feature values. Random forest builds on the decision tree concept by aggregating multiple trees to improve prediction accuracy and reduce overfitting. Lastly, the MLP (multilayer perceptron) is a type of neural network that captures complex, nonlinear relationships through layers of interconnected nodes. Figure 1, Figure 2, Figure 3 and Figure 4 illustrate the structures of these models in detail. Figure 5 is a training loss trend chart of the four models in 50 training cycles. The training progress chart shows that the MLP model achieved the lowest training loss and converged the fastest among all models, indicating strong learning efficiency. Linear regression also demonstrated steady and consistent improvement, ending with relatively low training loss. The decision tree model showed moderate performance, with a gradual decrease in loss over time. In contrast, the random forest maintained higher training loss throughout, likely due to its ensemble structure, which emphasizes generalization rather than closely fitting the training data. Overall, MLP outperformed the others during training, followed by linear regression, decision tree, and random forest.
4.4. Feature Importance
For the tree-based models (DT and RF), we assessed the relative importance of input variables. In RF, the most influential predictors were as follows:
Effective Federal Funds Rate (DFF)
Core CPI (CPILFESL)
Federal Debt Growth (GFDEBTN_growth)
GDP Growth (GDP_growth)
These findings align with the prior literature, which has shown strong empirical links between monetary policy, inflation expectations, and long-term interest rates [1,4]. The prominence of the federal funds rate supports the view that expectations about short-term monetary policy strongly influence the long end of the yield curve.
4.5. Temporal Robustness Analysis
To evaluate the robustness of model performance across varying macroeconomic conditions, we conducted a time-segmented analysis by dividing the test dataset into three distinct economic regimes:
Pre-Pandemic Period (January 2014–December 2019): Characterized by stable economic growth and moderate interest rates.
Pandemic Shock Period (January 2020–December 2021): Defined by heightened volatility due to COVID-19 disruptions, aggressive monetary policy interventions, and sharp yield movements.
Post-Pandemic/Inflation Shock Period (January 2022–December 2023): Marked by rising inflation, supply chain disruptions, and a tightening monetary policy cycle by the Federal Reserve.
For each period, model performance was re-evaluated using MSE, MAE, and R2. The random forest (RF) model maintained the highest accuracy and stability across all three regimes. Notably, RF exhibited strong predictive ability during the volatile pandemic period, suggesting resilience to structural shifts and nonlinearities in the data. In contrast, linear regression (LR) and decision tree (DT) models experienced performance degradation during the pandemic phase, likely due to their sensitivity to abrupt regime changes and limited capacity to capture complex interactions.
This segmented evaluation confirms the temporal robustness of ensemble-based machine learning models and highlights their suitability for real-world applications where economic conditions are dynamic and often nonstationary.
5. Discussion
5.1. Summary of Key Findings
The results of this study demonstrate that machine learning (ML) models, particularly the random forest (RF) algorithm, significantly outperform traditional linear regression (LR) and standalone decision tree (DT) methods in predicting the short-term movement of the U.S. 10-year Treasury bond yield. Among the models evaluated, RF achieved the highest accuracy (R2 = 0.5760), followed by DT and MLP, while LR yielded the lowest performance. These findings are consistent with prior research that has shown ensemble methods like RF to be highly effective in capturing nonlinear and high-dimensional relationships in financial data [6,9].
The analysis further confirms that key macroeconomic indicators, namely the federal funds rate, core inflation (CPI), GDP growth, and the federal debt growth rate, have significant explanatory power in bond yield forecasting. The importance rankings derived from the RF model suggest that monetary policy and inflation expectations remain dominant drivers of long-term yields, aligning with the term premium and expectation theory of interest rates [1,4].
5.2. Interpretation and Theoretical Implications
The superior performance of nonlinear models such as random forest (RF) and multilayer perceptron (MLP) over linear regression (LR) highlights the limitations of linear modeling in capturing the complex dynamics between macroeconomic variables and bond yields. However, this performance gain comes with a notable trade-off in interpretability.
While the RF and MLP models can uncover intricate nonlinear interactions and offer high predictive accuracy, they often function as “black boxes,” providing limited transparency into how specific inputs influence outputs. In contrast, LR offers full interpretability via model coefficients, which is particularly valuable in policymaking and regulatory environments, where understanding causality, accountability, and model rationale is essential.
For financial analysts, traders, and policymakers, this trade-off means that model selection should be context dependent. In high-stakes environments, such as central banking decisions or fiscal policy evaluation, slightly lower predictive performance might be acceptable if the model provides interpretable insights. Conversely, in trading or algorithmic forecasting scenarios, where real-time accuracy is paramount, models like RF may be preferable despite reduced transparency.
Future work should explore explainability techniques, such as SHAP (SHapley Additive Explanations) or LIME (Local Interpretable Model-agnostic Explanations), to bridge this gap by enhancing the interpretability of complex models. These tools can help decompose model predictions into intuitive variable contributions, making even black-box models more suitable for policy-related applications.
5.3. Practical Implications
This study has several practical implications for investors, policymakers, and system developers:
For bond traders and investors: The RF model offers a reliable tool for tactical asset allocation and risk management. Accurate short-term yield forecasts are crucial for pricing fixed-income securities and adjusting portfolio duration.
For policymakers: Understanding the leading indicators of bond yields helps assess market expectations and the transmission of monetary and fiscal policy. For example, persistent yield curve inversion could signal recession risks and influence rate-setting decisions.
For financial system developers: The integration of RF or MLP models into trading platforms or risk assessment systems can enhance decision support by delivering timely, data-driven forecasts [11,13].
5.4. Limitations
While deep learning models such as Long Short-Term Memory (LSTM) networks have shown promise in financial time-series forecasting due to their ability to capture long-range dependencies [19], they were not utilized in this study for several reasons. First, LSTM models are data-hungry and typically require large volumes of high-frequency or granular data to generalize effectively [20]. Given our dataset’s monthly frequency and modest size, implementing LSTM would likely have resulted in overfitting and reduced robustness. Second, this study prioritized model transparency and interpretability, particularly for stakeholders such as policymakers and analysts. Deep neural architectures like LSTM are often criticized for their “black box” nature, which can be a barrier in policy-driven or regulated environments [21]. Finally, LSTM implementation involves significant computational overhead and hyperparameter tuning, which was beyond the scope of this study’s focus on accessible and widely used forecasting methods. Future work will explore the application of deep learning methods including LSTM and Transformer-based architectures to address these challenges and potentially enhance forecasting performance in longer time horizons or higher-frequency datasets.
5.5. Directions for Future Research
Future studies could explore the following avenues:
Incorporating advanced deep learning techniques, such as Long Short-Term Memory (LSTM) networks or Transformer-based architectures, to capture sequential dependencies and long-range temporal patterns in macro-financial data [14].
Ensembling diverse machine learning models beyond random forest, such as Gradient Boosting Machines (e.g., XGBoost, LightGBM) or Stacked Generalization (Stacking) frameworks, which could enhance performance through model diversity and meta-learning.
Expanding the feature set to include international capital flows, term premium components, yield curve slope, market volatility indices (e.g., VIX), and global macroeconomic indicators, which may improve model responsiveness to external shocks and reflect global interdependencies in bond markets.
Scenario analysis and stress testing, using simulation-based forecasts under hypothetical macroeconomic conditions (e.g., policy rate shocks or geopolitical crises), to assess model stability and provide decision-support tools for risk managers and central banks.
Enhancing the interpretability of complex models using tools such as SHAP or LIME to make high-performing ensemble or neural models more transparent and suitable for policy analysis and regulatory oversight [22].
These directions can help build more robust, explainable, and globally informed forecasting frameworks, aligning predictive modeling efforts with the evolving complexity of financial systems.
6. Conclusions
The U.S. 10-year Treasury bond yield remains a central financial indicator, influencing global investment decisions, interest rate benchmarks, and macroeconomic policy. This study examined its short-term forecasting using a comparative approach across four models: linear regression (LR), decision tree (DT), random forest (RF), and multilayer perceptron (MLP). By integrating key macroeconomic indicators, including the federal funds rate, core CPI, GDP growth, and federal debt growth, we evaluated each model’s ability to predict bond yield changes over 1- to 3-month horizons.
The results highlight the superior performance of the random forest model, which achieved the lowest forecasting error and the highest R2 value among the models tested. This finding is consistent with previous research demonstrating the effectiveness of ensemble learning in financial forecasting tasks [6,9]. In contrast, the LR model, while interpretable and simple to implement, was less effective in capturing the nonlinear dynamics of bond yields, an observation also supported by Stock and Watson [8].
This study’s results carry several practical and theoretical implications. First, they suggest that machine learning models, particularly ensemble and neural network approaches, offer meaningful improvements over traditional econometric methods for yield forecasting. These models can better accommodate the complexity of modern financial systems, especially in the presence of structural breaks, regime shifts, and nonlinear interactions among economic variables [7].
Second, the findings support the ongoing integration of artificial intelligence (AI) into financial decision-making systems. Bond traders, risk managers, and institutional investors can incorporate these models into real-time analytics tools to enhance market responsiveness and improve risk-adjusted returns. Moreover, central banks and policymakers may benefit from model outputs as supplementary indicators of market sentiment and yield expectations.
However, several limitations must be acknowledged. The scope of the model is restricted to a specific set of macroeconomic indicators and a particular bond maturity. Broader inclusion of global economic and sentiment indicators could enrich predictive power. Furthermore, deep learning methods such as LSTM or hybrid models combining econometrics and AI might yield superior performance in long-term forecasting [14].
In conclusion, this research affirms the viability of machine learning methods for short-term bond yield forecasting, particularly in environments characterized by economic uncertainty and nonlinear dynamics. Future studies should continue exploring the fusion of interpretable AI models and economic theory to build more robust and transparent forecasting systems. As the financial landscape becomes increasingly data-driven, the strategic application of ML offers a compelling path forward for yield modeling and macro-financial analysis.
Y.-F.W. conceptualized the study and played a key role in drafting the manuscript. M.Y.-F.W. designed the questionnaires used in the pilot implementation and approved the final version of the manuscript. L.-Y.T. provided coding and technical support. All authors have read and agreed to the published version of the manuscript.
Not applicable.
Not applicable.
Dataset available on request from the authors.
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Figure 1 Linear regression model structure.
Figure 2 Decision tree model structure.
Figure 3 Random forest model structure.
Figure 4 MLP model structure.
Figure 5 Training loss trend chart of four models in 50 training cycles.
Model performance on training set.
| Model | Data | MSE | MAE | R2 |
|---|---|---|---|---|
| LR | Testing Set | 6.356 | 1.906 | 0.208 |
| Training Set | 6.363 | 1.901 | 0.211 | |
| DT | Testing Set | 3.496 | 1.078 | 0.564 |
| Training Set | 3.222 | 0.962 | 0.601 | |
| RF | Testing Set | 3.400 | 1.072 | 0.576 |
| Training Set | 3.247 | 0.994 | 0.597 | |
| MLP | Testing Set | 5.545 | 1.723 | 0.309 |
| Training Set | 5.431 | 1.695 | 0.327 |
1. Diebold, F.X.; Li, C. Forecasting the term structure of government bond yields. J. Econom.; 2006; 130, pp. 337-364. [DOI: https://dx.doi.org/10.1016/j.jeconom.2005.03.005]
2. Ang, A.; Piazzesi, M.; Wei, M. What does the yield curve tell us about GDP growth?. J. Econom.; 2006; 131, pp. 359-403. [DOI: https://dx.doi.org/10.1016/j.jeconom.2005.01.032]
3. Bernanke, B.S.; Blinder, A.S. The federal funds rate and the channels of monetary transmission. Am. Econ. Rev.; 1992; 82, pp. 901-921.
4. Rudebusch, G.D.; Swanson, E.T. The bond premium in a DSGE model with long-run real and nominal risks. Am. Econ. J. Macroecon.; 2012; 4, pp. 105-143. [DOI: https://dx.doi.org/10.1257/mac.4.1.105]
5. Nelson, C.R.; Siegel, A.F. Parsimonious modeling of yield curves. J. Bus.; 1987; 60, pp. 473-489. [DOI: https://dx.doi.org/10.1086/296409]
6. Gu, S.; Kelly, B.; Xiu, D. Empirical asset pricing via machine learning. Rev. Financ. Stud.; 2020; 33, pp. 2223-2273. [DOI: https://dx.doi.org/10.1093/rfs/hhaa009]
7. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016.
8. Stock, J.H.; Watson, M.W. Forecasting output and inflation: The role of asset prices. J. Econ. Lit.; 2003; 41, pp. 788-829. [DOI: https://dx.doi.org/10.1257/jel.41.3.788]
9. Breiman, L. Random forests. Mach. Learn.; 2001; 45, pp. 5-32. [DOI: https://dx.doi.org/10.1023/A:1010933404324]
10. Dase, R.K.; Pawar, D.D. Application of artificial neural network for stock market predictions: A review of literature. Int. J. Mach. Intell.; 2010; 2, pp. 14-17.
11. Patel, J.; Shah, S.; Thakkar, P.; Kotecha, K. Predicting stock and stock price index movement using trend deterministic data preparation and machine learning techniques. Expert Syst. Appl.; 2015; 42, pp. 259-268. [DOI: https://dx.doi.org/10.1016/j.eswa.2014.07.040]
12. Ghosh, D.; Chaudhuri, K. Interest rate forecasting using machine learning techniques. Int. J. Forecast.; 2021; 37, pp. 640-654. [DOI: https://dx.doi.org/10.1016/j.ijforecast.2020.09.005]
13. Zhang, Y.; Li, S.; Xu, L.; Wang, H. Predicting government bond yield curve using machine learning algorithms. Ann. Oper. Res.; 2019; 282, pp. 681-703. [DOI: https://dx.doi.org/10.1007/s10479-017-2645-6]
14. Bao, W.; Yue, J.; Rao, Y. A deep learning framework for financial time series using stacked autoencoders and long-short term memory. PLoS ONE; 2017; 12, e0180944. [DOI: https://dx.doi.org/10.1371/journal.pone.0180944] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28708865]
15. Fisher, I. The Theory of Interest; Macmillan: New York, NY, USA, 1930.
16. Friedman, M. The role of monetary policy. Am. Econ. Rev.; 1968; 58, pp. 1-17.
17. Montgomery, D.C.; Peck, E.A.; Vining, G.G. Introduction to Linear Regression Analysis; 6th ed. Wiley: Hoboken, NJ, USA, 2021.
18. Breiman, L.; Friedman, J.; Olshen, R.A.; Stone, C.J. Classification and Regression Trees; Routledge: New York, NY, USA, 2017.
19. Hewamalage, H.; Bergmeir, C.; Bandara, K. Recurrent neural networks for time series forecasting: Current status and future directions. Int. J. Forecast.; 2021; 37, pp. 388-427. [DOI: https://dx.doi.org/10.1016/j.ijforecast.2020.06.008]
20. Zhang, G.; Aggarwal, C.C.; Qi, G.J. Stock price prediction via discovering multi-frequency trading patterns. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; Halifax, NS, Canada, 13–17 August 2017; pp. 2141-2149. [DOI: https://dx.doi.org/10.1145/3097983.3098165]
21. Samek, W.; Wiegand, T.; Müller, K.R. Explainable artificial intelligence: Understanding, visualizing and interpreting deep learning models. arXiv; 2017; [DOI: https://dx.doi.org/10.48550/arXiv.1708.08296] arXiv: 1708.08296
22. Lundberg, S.M.; Lee, S.-I. A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst.; 2017; 30, pp. 4765-4774.
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; Wang, Max Yue-Feng 2 ; Li-Ying, Tu 1 1 Institute of Information and Decision Sciences, National Taipei University of Business, Taipei 100, Taiwan; [email protected]
2 Department of Marketing, Pennsylvania State University, University Park, PA 16802, USA; [email protected]