نوع مقاله : علمی - پژوهشی
نویسندگان
1 گروه مالی و بانکداری، دانشکده مدیریت و حسابداری، دانشگاه علامه طباطبائی(ره)، تهران، ایران،
2 گروه مالی و بانکداری، دانشکده مدیریت و حسابداری، دانشگاه علامه طباطبائی(ره)، تهران، ایران.
3 گروه رایانه، دانشکده آمار، ریاضی و رایانه، دانشگاه علامه طباطبائی(ره)، تهران، ایران.
چکیده
کلیدواژه‌ها
عنوان مقاله [English]
نویسندگان [English]
Purpose: The yield curve is a key analytical tool in economics, offering vital insights into market expectations regarding monetary policy, economic conditions, and inflation across various time horizons. It also plays a critical role in fiscal policymaking, financial institution modeling, and investment decisions such as asset valuation and risk management. Despite its importance, the analysis and forecasting of the yield curve have received limited attention in Iran. This becomes especially significant in the context of chronic inflation, currency volatility, international sanctions, and dependence on oil revenues. The present study aims to forecast the risk-free government bond yield curve in Iran. To this end, a two-dimensional forecasting approach across both time and maturity dimensions is employed, allowing for simultaneous analysis of the term structure and its dynamic behavior over time.
Methodology: Among the various approaches to yield curve forecasting, the Dynamic Nelson-Siegel (DNS) factor model is adopted as the foundational framework due to its interpretability, dimensionality reduction capabilities, and its ability to summarize the curve through three latent factors: level, slope, and curvature. These factors have well-established economic and financial interpretations and provide a meaningful basis for strategic and policy-level decision-making. Using data from Iranian Islamic Treasury Bills (ITBs), this study forecasts the aforementioned factors using a range of models, including the Vector Autoregressive-GARCH (VAR-GARCH) model as a classical baseline, gradient boosting algorithms as shallow machine learning models, and deep learning architectures such as Convolutional-Recurrent Long Short-Term Memory (Conv-LSTM) networks and Gated Recurrent Units (GRU). These models differ in terms of complexity, interpretability, data requirements, computational demands, and their capacity to capture linear or nonlinear relationships.
Findings: The empirical results reveal that the VAR-GARCH model outperforms others in forecasting the level factor, largely due to its autoregressive structure, which is better suited for modeling stable long-term trends. Conversely, deep learning models underperform in predicting the level factor due to limited data availability and difficulty in capturing persistent trends. However, for the slope and curvature factors—more influenced by short- and medium-term fluctuations—deep learning models demonstrate superior performance, owing to their ability to capture complex nonlinear temporal patterns. In contrast, traditional statistical models exhibit limitations in handling such fluctuations due to rigid assumptions. Subsequently, the predicted factors were integrated into the DNS model, and the accuracy of the reconstructed yield curve was evaluated using the Root Mean Square Error (RMSE). The results indicate that no single model dominates in predicting all three factors simultaneously. Therefore, a hybrid model strategy, in which each factor is forecasted by the most accurate model, leads to enhanced reconstruction performance. This approach is also theoretically consistent with the DNS model’s assumption of factor independence. The optimal configuration was achieved when the level factor was predicted using either VAR-GARCH or Conv-LSTM, the slope factor using GRU, and the curvature factor using either VAR-GARCH or a gradient boosting algorithm, resulting in a reconstruction error of approximately 0.5%.
Conclusion: This study introduces an accurate and data-driven framework for yield curve forecasting in the Iranian financial market by leveraging the Dynamic Nelson-Siegel model. Unlike previous studies that primarily relied on classical approaches such as VAR, this research integrates both shallow and deep machine learning models. In the first stage, these models were evaluated based on their ability to predict the DNS factors. The VAR-GARCH model was found to be most effective for forecasting the level factor, while deep learning models were more accurate in predicting slope and curvature. In the second stage, the reconstructed yield curve, based on the predicted factors, was assessed using RMSE. The findings suggest that a tailored combination of models for each factor—specifically, VAR-GARCH or Conv-LSTM for level, GRU for slope, and VAR-GARCH or gradient boosting for curvature—results in the highest forecasting accuracy, with a reconstruction error of less than 0.5%.
کلیدواژه‌ها [English]
)