| 263 | 0 | 450 |
| 下载次数 | 被引频次 | 阅读次数 |
消费者价格指数（CPI）是衡量通胀水平的核心指标。传统统计模型在处理通胀数据时，往往难以刻画其所包含的非线性结构与动态依赖关系，且现实经济数据规模有限、噪声水平较高，导致利用神经网络建模预测CPI的效果受限，并且缺乏经济结构解释，呈现“黑箱化”特征。为此，本文提出一种融合通胀机制约束的时序卷积网络模型，通过软约束正则化方式，将货币数量论、成本推动通胀、汇率传导效应和通胀惯性4类通胀生成机制嵌入模型损失函数，并在约束项中引入可学习的结构参数和偏置项，使模型在数据驱动的同时具备对经济规律的自适应表达能力。此外，在不改变主模型的前提下，进一步构建时间与状态门控的补充模型，用于对主模型计算的机制残差进行时变加权，从而识别不同经济阶段4类机制的相对贡献及其阶段性切换，实现对预测误差的结构性归因以及探索不同经济周期下各类机制对通胀的主导作用是否存在动态变化。基于2002—2024年我国月度宏观经济数据分析，本文所提出的模型在趋势识别、转折点检测和整体预测一致性等方面均优于不包含约束的基线模型，兼具较强的预测性能与经济解释力，门控模型则进一步提供了对不同时期主导机制的阶段性刻画与可视化证据。
Abstract：The Consumer Price Index(CPI) is a core gauge of inflation. Traditional statistical models often struggle to capture the nonlinear structure and dynamic dependencies inherent in inflation data, which limits their predictive performance. In practice, small sample sizes and high noise levels further constrain neural network-based CPI forecasting and reduce structural interpretability, reinforcing a “black-box” characteristic. To address these issues, this study proposes a time convolutional network(TCN) with embedded inflation-mechanism constraints. Through soft constraint regularization in the loss function, we embed four canonical mechanisms—quantity theory of money, cost-push inflation, exchange-rate pass-through, and inflation inertia—and introduce learnable structural parameters and intercept terms within the constraints, enabling data-driven learning while adaptively representing economic regularities. In addition, without altering the main model, we develop a supplementary time-and-state gated module that applies time-varying weights to the mechanism residuals computed by the static model. This will identify the relative contributions of the four mechanisms across economic regimes and their phase shifts, supporting structural attribution of forecast errors and assessing whether the dominant drivers of inflation vary over different economic cycles constraint. Using Chinese monthly macroeconomic data from 2002 to 2024, the static mechanism-embedded TCN outperforms unconstrained baseline models in trend tracking, turning-point detection, and overall forecast consistency, combining strong predictive accuracy with economic interpretability. The gated module further provides stage-specific characterization and visual evidence on the time-varying dominance of mechanisms.
[1]任英华,江劲风,倪青山.基于图神经网络模型的金融危机预警研究——全行业间信息溢出视角[J].统计研究, 2022, 39(8):141–160.
[2]王小燕,冮建伟,王洁丹,等.多源异质数据下深度神经网络的整合分析及其应用[J].统计研究, 2025, 42(2):122–134.
[3]Buckmann M, Joseph A, Robertson H. Opening the Black Box:Machine Learning Interpretability and Inference Tools with an Application to Economic Forecasting[M]. Data Science for Economics and Finance:Methodologies and Applications. Cham:Springer International Publishing,2021:43–63.
[4]Cao Y, Liu X, Zhai J. Option Valuation under No-Arbitrage Constraints with Neural Networks[J]. European Journal of Operational Research,2021, 293(1):361–374.
[5]Chen Z, Liu Y, Sun H. Physics-Informed Learning of Governing Equations from Scarce Data[J]. Nature Communications, 2021, 12(1):6136:1–13.
[6]Chronopoulos I, Raftapostolos A, Kapetanios G. Forecasting Value-At-Risk Using Deep Neural Network Quantile Regression[J]. Journal of Financial Econometrics, 2024, 22(3):636–669.
[7]Dash T, Chitlangia S, Ahuja A, et al. A Review of Some Techniques for Inclusion of Domain-Knowledge into Deep Neural Networks[J].Scientific Reports, 2022, 12(1):1040.
[8]Fischer M M, Hauzenberger N, Huber F, et al. General Bayesian Time-Varying Parameter Vector Autoregressions for Modeling Government Bond Yields[J]. Journal of Applied Econometrics, 2023, 38(1):69–87.
[9]Gonçalves S, Herrera A M, Kilian L, et al. State-Dependent Local Projections[J]. Journal of Econometrics, 2024, 244(2):105702.
[10]Gur Y E. Development and Application of Machine Learning Models in US Consumer Price Index Forecasting:Analysis of a Hybrid Approach[J]. Data Science in Finance and Economics, 2024, 4(4):469–513.
[11]Hamiane S, Ghanou Y, Khalifi H, et al. Comparative Analysis of LSTM, ARIMA, and Hybrid Models for Forecasting Future GDP[J].International Journal of Sustainable Development and Planning, 2024, 29(3):853–861.
[12]Hauzenberger N, Huber F, Koop G, et al. Fast and Flexible Bayesian Inference in Time-Varying Parameter Regression Models[J]. Journal of Business&Economic Statistics, 2022, 40(4):1904–1918.
[13]Hazell J, Herreno J, Nakamura E, et al. The Slope of the Phillips Curve:Evidence from US States[J]. The Quarterly Journal of Economics,2022, 137(3):1299–1344.
[14]Iorio F D, Triacca U. A Comparison Between VAR Processes Jointly Modeling GDP and Unemployment Rate in France and Germany[J].Statistical Methods&Applications, 2022, 31(3):617–635.
[15]Kumar S. Analysis of CPI Time Series across Different Economic Regimes Using ARIMA Models[J]. International Journal of High School Research, 2024, 6(2):133–138.
[16]Li Z, Zheng H, Kovachki N, et al. Physics-Informed Neural Operator for Learning Partial Differential Equations[J]. ACM/JMS Journal of Data Science, 2024, 1(3):1–27.
[17]Lo A W, Singh M. Deep-Learning Models for Forecasting Financial Risk Premia and Their Interpretations[J]. Quantitative Finance, 2023, 23(6):917–929.
[18]Lu C J, Lee T S, Chiu C C. Financial Time Series Forecasting Using Independent Component Analysis and Support Vector Regression[J].Decision Support Systems, 2009, 47(2):115–125.
[19]Müller K R, Smola A J, Rätsch G, et al. Predicting Time Series with Support Vector Machines[A]. International Conference on Artificial Neural Networks[C]. 1997:999–1004.
[20]Ning Y, Kazemi H, Tahmasebi P. A Comparative Machine Learning Study for Time Series Oil Production Forecasting:ARIMA, LSTM, and Prophet[J]. Computers&Geosciences, 2022, 164:105126.
[21]Nonejad N. An Overview of Dynamic Model Averaging Techniques in Time-Series Econometrics[J]. Journal of Economic Surveys, 2021,35(2):566–614.
[22]Penwarden M, Zhe S, Narayan A, et al. A Metalearning Approach for Physics-Informed Neural Networks(PINNs):Application to Parameterized PDEs[J]. Journal of Computational Physics, 2023, 477:111912.
[23]Shi H, Wei A, Xu X, et al. A CNN-LSTM Based Deep Learning Model with High Accuracy and Robustness for Carbon Price Forecasting:A Case of Shenzhen’s Carbon Market in China[J]. Journal of Environmental Management, 2024, 352:120131.
[24]Theoharidis A F, Guillén D A, Lopes H. Deep Learning Models for Inflation Forecasting[J]. Applied Stochastic Models in Business and Industry, 2023, 39(3):447–470.
[25]Wang T, Beard R, Hawkins J, et al. Recursive Deep Learning Framework for Forecasting the Decadal World Economic Outlook[J]. IEEE Access, 2024(3):152921–152944.
[26]Zhang R, Meng Q, Zhu R, et al. Monte Carlo Neural PDE Solver for Learning PDES via Probabilistic Representation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025(6):5059–5075.
(1)因篇幅所限，1Smooth L损失函数定义以附录1展示，见《统计研究》网站所列附件。下同。
(1)因篇幅所限，TrendAcc 、 TurningPt 、 Corr 、 SMAPE公式以附录2展示。
(2)因篇幅所限，贝叶斯优化策略与采样机制具体公式以附录3展示。
(3)因篇幅所限，静态TCN模型输入以附录4展示。
(1)因篇幅所限，测试集预测对比图（前8佳模型+基线模型）以附图1展示。
(2)因篇幅所限，最佳模型训练过程中损失曲线以附图2展示。
(1)因篇幅所限，参数筛选过程图以附图3展示。
(2)因篇幅所限，4个典型高波动经济事件官方名称来源说明以附录5展示。
(3)因篇幅所限，关键经济事件期间的局部走势图以附图4展示。
(1)因篇幅所限，机制贡献的时变加权动态TCN模型设定的4个机制残差公式以附录6展示。
(1)训练中对强度向量α施加的是向锚点α0的二次收缩项■，以稳定长期强度尺度。同时，对状态映射权重矩阵W加入正则项■抑制过度敏感。动态门控训练阶段经济结构参数保持固定。
(2)因篇幅所限，能量分摊口径公式以附录7展示。
(3)因篇幅所限，逐机制贡献曲线与主导机制色带图、误差改善份额图分别以附图5和附图6展示。
基本信息:
DOI：10.19343/j.cnki.11-1302/c.2026.02.010
中图分类号:F726;TP183
引用信息:
[1]李佼瑞,邓迪,张雨筱.多种通胀机制约束的时序卷积网络模型改进及其在CPI预测中的应用[J].统计研究,2026,43(02):131-143.DOI:10.19343/j.cnki.11-1302/c.2026.02.010.
基金信息:
国家自然科学基金面上项目“经济–环境系统的分数阶随机动力学建模与分析”（1157223）
2025-06-23
2025
2026-01-17
2026-01-27
2026
1
2026-02-25
2026-02-25