N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting

TL;DR

N-HiTS tackles long-horizon time-series forecasting by introducing multi-rate input sampling and hierarchical interpolation, enabling blocks to specialize in different frequencies and to synthesize predictions at multiple time scales. The architecture combines backcast/forecast projections with a multi-resolution interpolation scheme, backed by a neural basis approximation perspective. Empirical results across six large-scale datasets show state-of-the-art accuracy with substantial reductions in computation and memory, while ablation studies confirm the complementary value of each component and the top-down hierarchical ordering. The work also emphasizes interpretability through decomposed forecast components and demonstrates broad efficiency gains, suggesting strong practical impact for scalable time-series systems.

Abstract

Recent progress in neural forecasting accelerated improvements in the performance of large-scale forecasting systems. Yet, long-horizon forecasting remains a very difficult task. Two common challenges afflicting the task are the volatility of the predictions and their computational complexity. We introduce N-HiTS, a model which addresses both challenges by incorporating novel hierarchical interpolation and multi-rate data sampling techniques. These techniques enable the proposed method to assemble its predictions sequentially, emphasizing components with different frequencies and scales while decomposing the input signal and synthesizing the forecast. We prove that the hierarchical interpolation technique can efficiently approximate arbitrarily long horizons in the presence of smoothness. Additionally, we conduct extensive large-scale dataset experiments from the long-horizon forecasting literature, demonstrating the advantages of our method over the state-of-the-art methods, where N-HiTS provides an average accuracy improvement of almost 20% over the latest Transformer architectures while reducing the computation time by an order of magnitude (50 times). Our code is available at bit.ly/3VA5DoT

Figures (11)

• Figure 1: (a) The computational costs in time and memory (b) and mean absolute errors (MAE) of the predictions of a high capacity fully connected model exhibit evident deterioration with growing forecast horizons. (c) Specializing a flexible model's outputs in the different frequencies of the signal through hierarchical interpolation combined with multi-rate input processing offers a solution.
• Figure 1: Datasets' partition into train, validation, and test sets used in our experiments (ETTm$_2$, ECL, Exchange, ILI, TrafficL, and Weather). All use the last 20% of the total observations as test set (marked by the second dotted line), and the 10% preceding the test set as validation (between the first and second dotted lines), except for ETTm$_2$ that also use 20% as validation. Validation provides the signal for hyperparameter optimization. We construct test predictions using rolling windows.
• Figure 2: N-HiTS architecture. The model is composed of several MLPs with ReLU nonlinearities. Blocks are connected via doubly residual stacking principle with the backcast $\mathbf{\tilde{y}}_{t-L:t,\ell}$ and forecast $\mathbf{\hat{y}}_{t+1:t+H,\ell}$ outputs of the $\ell$-th block. Multi-rate input pooling, hierarchical interpolation and backcast residual connections together induce the specialization of the additive predictions in different signal bands, reducing memory footprint and compute time, improving architecture parsimony and accuracy.
• Figure 2: Proposed pooling configurations.
• Figure 3: N-HiTS composes its predictions hierarchically using blocks specialized in different frequencies based on controlled signal projections, through expressiveness ratios, and interpolation of each block. The coefficients are locally determined along the horizon, allowing N-HiTS to reconstruct non-periodic/stationary signals, beyond constant Fourier transform projections.

Theorems & Definitions (3)

• proof
• proof
• proof