High-frequency inflation forecasting: A two-step machine learning methodology
Прогнозирование инфляции с высокой частотой: двухэтапная методология машинного обучения
2025-05-17
SCID: 54.1/kdafss48
Discuss with AI
Ridge regressionhigh-frequency inflation forecastingmachine learningmixed-frequency indicatorsnowcasting
Figures from the paper
Abstract (AI)
This study introduces a novel two-step machine learning methodology to generate high-frequency (daily and weekly) inflation forecasts in developing economies, where official statistics are typically available only at a monthly frequency and with delays. High-frequency forecasting here is interpreted as nowcasting or interpolation — real-time prediction ahead of official releases or within-period estimation using mixed-frequency indicators — while also serving as a data-augmentation strategy. In the first step, high-frequency predictors are aggregated to construct monthly-aligned features that serve as inputs for training machine learning models. In the second step, various feature selection techniques are evaluated and multiple machine learning algorithms are rigorously fine-tuned via hyperparameter optimization. Through systematic evaluation, a final model was selected — Ridge regression trained on an L1-regularized feature subset — that achieves superior out-of-sample accuracy. This model is then deployed to produce high-frequency year-on-year CPI inflation nowcasts. Forecasts exhibit strong temporal alignment with observed monthly values, while distributional equivalence — monthly vs. high-frequency projections — is confirmed via Kolmogorov–Smirnov tests. Compared to benchmark econometric models, the proposed approach delivers improved predictive performance, offering timely and granular insights for forward-looking monetary policy.
Key Findings
1
High-frequency predictors are aggregated into monthly-aligned features, followed by feature selection and hyperparameter optimization across multiple machine learning algorithms.
2
Kolmogorov–Smirnov tests confirm distributional equivalence between monthly observations and high-frequency inflation projections.
3
Ridge regression trained on an L1-regularized feature subset achieves the best out-of-sample forecasting accuracy among evaluated models.
4
The proposed approach outperforms benchmark econometric models and provides more timely, granular information for forward-looking monetary policy.
5
The resulting year-on-year CPI inflation nowcasts closely align temporally with observed monthly inflation values.
6
The study introduces a two-step machine learning methodology for daily and weekly inflation nowcasting in developing economies with delayed monthly statistics.
Research Object
High-frequency year-on-year CPI inflation in developing economies
Research Subject
Nowcasting and interpolation accuracy, temporal alignment, and distributional equivalence of daily and weekly inflation projections with official monthly CPI inflation
Publication Details
Publication Date
2025-05-17
Journal
Publisher
ISSN
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest