N-HiTS

N-HiTS (Challu, Olivares, Oreshkin, Garza, Mergenthaler-Canseco, Dubrawski, AAAI 2023) answers N-BEATS's weak spot: long-horizon forecasting. It keeps the doubly residual MLP stacking but adds two ideas. First, multi-rate data sampling: each block sees a max-pooled version of the input, so blocks specialize in different frequency bands. Second, hierarchical interpolation: each block forecasts a coarse low-resolution view of the horizon, and finer blocks interpolate the details, which provably approximates long horizons efficiently. The result is sharper long-range predictions at a fraction of the compute.

The authors report an average accuracy improvement of almost 20% over the latest Transformer architectures on long-horizon benchmarks, while reducing computation time by an order of magnitude. The original cchallu/n-hits repo (193 stars) carries no detected license, so the practical path is Nixtla's NeuralForecast, where N-HiTS is maintained under MIT with hyperparameter optimization and multivariate support.

Author-reported (Challu et al., AAAI 2023): ~20% average accuracy improvement over latest Transformer architectures (Autoformer, Informer, LogTrans, Reformer) on long-horizon datasets (ETTm2, ECL, Exchange, Traffic, Weather, ILI), with ~50x compute reduction. No large independent replication published beyond community adoption.