“N-BEATS: Neural basis expansion analysis for interpretable time series forecasting” by Boris Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio was posted to arXiv in 2019. It was an important result because it showed that a plain deep neural network could win at forecasting, a domain where simple statistical methods had long held the lead.
The architecture is deliberately generic. It stacks many blocks of fully-connected layers, where each block produces both a forecast and a backcast (its reconstruction of the input), and the residual left after subtracting the backcast is passed to the next block. There are no recurrent layers, no convolutions, and nothing hand-crafted for seasonality or trend. Despite this, N-BEATS improved on the winner of the M4 competition by about three percent and beat a strong statistical benchmark by eleven percent. An interpretable variant constrains the blocks so that some learn trend and others learn seasonality, recovering a human-readable decomposition.
The paper helped trigger a wave of deep learning research on forecasting, suggesting that, given enough related series to learn from, neural networks could compete with and exceed classical methods.
Why business readers should care: N-BEATS marked the point where deep learning became a serious option for demand and sales forecasting, not just for images and text.