Multichannel SSA decomposition in a hybrid SSA-N-BEATS framework for complex time series forecasting

Forecasting high-resolution time series remains challenging because of nonlinear behavior, multiple seasonalities, and high-frequency noise. Deep learning architectures such as N-BEATS offer strong forecasting capability, yet their performance often deteriorates when trained directly on raw signals that mix multiple temporal characteristics. Hybrid frameworks integrating decomposition with neural networks offer a practical solution, though effectiveness depends on the stability of the underlying data. This paper benchmarks the N-BEATS (baseline) model against two Hybrid SSA-N-BEATS scenarios: Hybrid SSA-N-BEATS (denoising) and Hybrid SSA-N-BEATS (multichannel). In the denoising scenario, singular spectrum analysis (SSA) is applied solely for denoising, isolating deterministic signals before network training. In the multichannel scenario, SSA decomposes the series into trend and seasonal subsignals, each processed in parallel through an independent N-BEATS channel. The proposed evaluation protocol assesses model behavior across stable periods and periods marked by structural level shifts. This provides a systematic framework for identifying when decomposition-based hybridization is practically beneficial. This work makes the following methodological contributions:•Provides a decomposition-based benchmarking framework for forecasting complex time series.•Introduces both denoising-based and multichannel SSA-N-BEATS pipelines.•Demonstrates that hybrid model effectiveness depends on temporal stability and structural level shifts.