N-HiTS Hierarchical Interpolation

Forecasting a series that moves on several timescales at once — a slow trend, a weekly cycle, and jittery daily noise — is hard for a single neural network block to do well, because a filter tuned to catch the slow trend blurs out the fast wiggles, and a filter tuned to the fast wiggles gets lost trying to also track the trend. N-HiTS solves this by giving up on doing it all in one block and instead stacking several specialized ones. Each block first downsamples its input, so it only ever sees a coarsened, low-resolution version of the series — one block might see a heavily smoothed version (catching only the trend), while another sees only mild smoothing (catching finer detail). Each block produces a small forecast at its own coarse resolution, and then interpolates that forecast back up to the full, fine-grained resolution the reader actually wants — spreading a few coarse points into a full detailed curve, similar to stretching a low-resolution image up to full size. The interpolated outputs of every block are summed to produce the final forecast, and each block also passes its unexplained "residual" to the next block, exactly like N-BEATS Interpretable Forecasting's residual stacking.

The gain from this multi-rate design is mostly computational: coarser blocks operate on far fewer points, so N-HiTS is both faster to train and better at long-horizon forecasts than models that try to predict every timescale at full resolution simultaneously, since the model is never forced to represent a slow, smooth trend using thousands of jittery high-frequency parameters.