Wavelet-based Multiscale Approach in High-Dimensional Forecasting Models

Wavelet-based Multiscale Approach in High-Dimensional Forecasting Models

• Autor
• Aldryn Dylan Mamani Quispe
• Resumo

• This paper proposes a multiscale forecasting framework that combines wavelet-based decompositions with high-dimensional estimation techniques. By decomposing inflation into frequency components and estimating separate models for each component using frequency-matched predictors, the approach captures heterogeneous dynamics across short- and long-run horizons. An optimization step selects, for each forecast horizon, the combination of components and models that minimizes out-of-sample forecasting error. We apply the methodology to forecast Brazilian inflation and evaluate its performance against standard benchmarks, including random walk and autoregressive models, survey-based expectations, and high-dimensional forecasting methods. The results show that the proposed method delivers substantial gains at short horizons and generally improves upon survey and high-dimensional benchmarks, although gains narrow at longer horizons. The component selection reveals that different estimation strategies specialize across frequency bands. Overall, the findings suggest that exploiting frequency-specific information through wavelet decompositions, combined with high-dimensional methods, provides a promising and interpretable tool for inflation forecasting.

• Palavras-chave
• Wavelets; Multiscale modeling; Frequency-specific forecasting; High-dimensional forecasting; Inflation forecasting.
• Modalidade
• Comunicação oral
• Área Temática
• Econometria Financeira

Voltar Download