Abstract
This paper evaluates the interpretability and analytical utility of spectral emission heads in the Nonstationary Spectral Decomposition Network (NS-SDN) using the 10-Year Treasury yield (DGS10) as a benchmark macroeconomic time series. Rather than focusing on recursive forecasting, we examine the model’s ability to decompose nonstationary signals into adaptive spectral components under one-step conditional forecasting. Results show that emission heads provide meaningful decomposition into trend, amplitude, instantaneous frequency, and a learned gating signal, and reveal regime-dependent spectral behavior, including frequency convergence during the 2020 macroeconomic shock. These findings position NS-SDN as a neural spectral analysis framework rather than solely a forecasting model.