> ML_LITERATURE // ORESHKIN-2020-N-BEATS-NEURAL-BASIS-EXPANSION-FORECASTING_v1.0
N-BEATS: Neural basis expansion analysis for interpretable time series forecasting
Boris N. Oreshkin, Dmitry Carpov, Nicolas Chapados, Yoshua Bengio · International Conference on Learning Representations (ICLR) (2020)
algorithm2020industry-standardthirdPartyReproduced
Principal Contribution
Constructed a pure deep neural architecture using backward and forward residual link stacks projecting onto interpretable basis functions (polynomial trend, Fourier seasonality) without RNNs or attention.
Operational Relevance
Serves as qualified reference for implementing task-time-series-forecasting in production systems.
Assumptions
- Underlying spatio-temporal continuity and domain distributional stability hold
Limitations
- Performance scaling and computational footprint depend on receptive field depth, sequence length, and resolution