Core Concepts
Wavelet features improve forecasting accuracy across various machine learning models.
Abstract
This article explores the combination of wavelet analysis techniques with machine learning methods for time series forecasting. It investigates the use of Daubechies wavelets with varying numbers of vanishing moments as input features, comparing non-decimated wavelet transform and non-decimated wavelet packet transform. The experiments suggest significant benefits in replacing higher-order lagged features with wavelet features for one-step-forward forecasting. For long-horizon forecasting, modest benefits are observed when using wavelet features with deep learning-based models. The study highlights the importance of considering wavelet features for improved forecasting accuracy.
Stats
Wavelet numbers selected during cross-validation: 1, 7, 5, 1, 1, 1, 1, 9, 3
Mean SMAPE for one-step-ahead forecasts: Ridge (33.23%), SVR (42.51%), Forest (36.94%), XGBoost (36.15%), MLP (36.49%)
Quotes
"Our experiments suggest significant benefit in replacing higher-order lagged features with wavelet features across all examined non-temporal methods."
"The latter include state-of-the-art transformer-based neural network architectures."
"Using NDWT and NWPT multivariate inputs result in superior forecasts for seven out of nine deep learning models."