Construction and Empirical Analysis of Macroeconomic Forecasting Models Based on Machine Learning
DOI:
https://doi.org/10.61173/y7a9f995Keywords:
Macroeconomic Forecasting, Machine Learning, LSTM, Time Series, Quantitative EconomicsAbstract
Macroeconomic forecasting is an important basis for policy-making and investment decision. However, traditional econometric models always show some limitations in handling non-linear and high-dimensional data environment, especially in the face of more complex and fluctuated global economic environment. The complexity gradually increases, and more adaptable and robust analytical tools are needed to capture the intricate patterns and relationships between economic variables. This paper constructs a forecasting model based on advanced machine learning methods combining time series features and macroeconomic index, and compares the LSTM neural network and Random Forest algorithm. Based on China’s macroeconomic data from 2010 to 2023, the empirical results show that machine learning models have obvious advantages over the traditional ARIMA model in forecasting two important indicators GDP growth rate and inflation rate, and the average forecasting error is reduced by 18.7%. The empirical results of machine learning models show that these models are better at capturing temporal dependencies and feature interactions than traditional models. This study not only provides new support for improving the accuracy of macroeconomic forecasting, but also verifies the strong application significance of machine learning in quantitative economics and proves that using these techniques can make economic policies and investment strategies more accurate and responsive in rapidly changing economies such as China.
References
and the lack of alternative data (such as private sector 1, 2008, pp. 71-92. data) makes the data source limited, which affects the di- [3] Breiman, Leo. „Random Forests.“ Machine Learning, vol.
versity of input features. 45, no. 1, 2001, pp. 5-32. (5) Model Limitations: Although this study compares [4] Liu, Yan, and Yilin Wang. „Macroeconomic Forecasting LSTM and Random Forest models, other advanced with Random Forest: Evidence from China.“ China Economic
machine learning algorithms (such as Transformer or Review, vol. 68, 2021, pp. 101602. XGBoost) are not compared. Moreover, the model used [5] Hochreiter, Sepp, and Jürgen Schmidhuber. „Long Short-
in this study is a single-step forecasting model (that is, Term Memory.“ Neural Computation, vol. 9, no. 8, 1997, pp. forecasting the next period based on historical data), and 1735-80. multi-step forecasting (that is, forecasting the next 3–6 pe- [6] Chen, Xiaohong, and Zhuo Song. „Forecasting China‘s GDP riods) has not been explored, which limits the application Growth Using a LSTM Neural Network.“ Economic Modelling, of the model in long-term policy planning.
vol. 80, 2019, pp. 192-201. (6) Interpretability Limitations: This study does not an- [7] Hamilton, James D. „Why You Should Never Use the alyze the interpretability of machine learning models. The “black box” model makes it difficult to explain the Hodrick-Prescott Filter.“ Review of Economics and Statistics,
internal mechanism of the model’s forecast results in a vol. 100, no. 5, 2018, pp. 831-43. clear way, which may affect the policy-makers’ trust in the [8] Diebold, Francis X., and Robert S. Mariano. „Comparing model. Predictive Accuracy.“ Journal of Business & Economic (7) Data Expansion: In the future, high-frequency data Statistics, vol. 13, no. 3, 1995, pp. 253-63. (such as daily electricity consumption, weekly retail sales) [9] Wang, Yiping, and Jia Li. „The Impact of COVID-19 on and alternative data (such as social media sentiment data, Macroeconomic Forecasting: A Machine Learning Perspective.“
satellite remote sensing data) will be used in the model to Journal of Economic Dynamics and Control, vol. 129, 2021, pp. improve the timeliness and comprehensiveness of fore- 104168. casting. For example, using satellite images to monitor the [10] Li, Wei, and Mark W Watson. „Robustness Checks in operation of industrial parks and using the above data as Macroeconomic Forecasting: A Machine Learning Approach.“
auxiliary information to predict the industrial production Journal of Applied Econometrics, vol. 36, no. 4, 2021, pp. 445- index. 67. (8) Model Innovation: Develop multi-step forecasting [11] Molnar, Christoph. Interpretable Machine Learning. Lulu. models based on deep learning algorithms (such as Trans- com, 2020. former) to realize medium- and long-term forecasting of [12] Bontempi, Gianluca, et al. „Machine Learning Strategies macroeconomic indicators (such as next 2 years of GDP for Time Series Forecasting.“ Business Intelligence: Second
growth rate). At the same time, construct hybrid models European Summer School, eBISS 2012, Brussels, Belgium, July
that combine multiple machine learning algorithms (such 15-21, 2012, Tutorial Lectures, edited by Marie-Aude Aufaure as LSTM-Transformer hybrid model) to integrate the ad-
and Esteban Zimányi, Springer, 2013, pp. 62-77. vantages of different models. [13] Vaswani, Ashish, et al. „Attention Is All You Need.“ (9) Interpretability Improvement: Apply interpretability Advances in Neural Information Processing Systems, vol. 30, tools like SHAP and LIME to explain the contribution of 2017. each input variable to the forecasting results, for example, [14] Zhang, Guoqiang Peter. „Time Series Forecasting Using a the impact of M2 growth rate and interest rate changes on CPI forecasting, so as to make the decision-making pro- Hybrid ARIMA and Neural Network Model.“ Neurocomputing,
vol. 50, 2003, pp. 159-75.
Downloads
Published
Issue
Section
License
Copyright (c) 2025 by the authors.
This work is licensed under a Creative Commons Attribution 4.0 International License.
