N-BEATS

N-BEATS (Oreshkin, Carpov, Chapados, Bengio, ICLR 2020) is the model that challenged the assumption that forecasting needs domain-specific machinery. It stacks fully connected blocks connected by doubly residual links: each block backcasts the history and forecasts the horizon, and the sum of block outputs is the final prediction. Two configurations ship: a generic one with no time-series-specific components, and an interpretable one whose blocks are constrained to trend and seasonality basis functions so the forecast decomposes into readable components.

On the M4 competition's 100,000 heterogeneous series, the authors report an 11% accuracy improvement over a statistical benchmark and 3% over the previous M4 winner, the hand-crafted ES-RNN hybrid. The official implementation lives in ServiceNow/N-BEATS (the project moved from Element AI to ServiceNow after the 2021 acquisition), and the model is also maintained in Nixtla's NeuralForecast. Note the official repo carries no SPDX-detected license, so commercial users should check terms; the NeuralForecast implementation is MIT.

Author-reported (Oreshkin et al., ICLR 2020): 11% better than the M4 statistical benchmark, 3% better than the previous M4 winner across M3, M4, and TOURISM datasets. Independent evaluations still rank it strongly for short-horizon univariate forecasting, though long-horizon work favors its successor N-HiTS.