[Paper Review] N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting

Cristian Challú, Kin G. Olivares+4 more|arXiv (Cornell University)|Jan 30, 2022

Stock Market Forecasting Methods42 citations

TL;DR

N-HiTS introduces multi-rate data sampling and hierarchical interpolation to long-horizon forecasting, achieving ~20% accuracy gains over Transformer baselines while reducing compute time by about 50x.

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

Motivation & Objective

• Motivate the need for accurate and scalable long-horizon forecasting in large-scale applications.
• Develop a neural forecasting model that reduces computational complexity while improving accuracy for long horizons.
• Decompose input signals and synthesize forecasts through hierarchical interpolation at multiple time scales.
• Enable interpretable decomposition of forecasts into frequency-band components via specialized blocks.

Proposed method

• Extend N-BEATS with two novel components: multi-rate input sampling and hierarchical interpolation.
• Use MaxPool with block-specific kernel sizes to create scale-focused inputs for each block.
• Each block learns forward and backward interpolation coefficients via an MLP to generate backcasts and forecasts.
• Use a controlled expressiveness ratio r_l to define the number of interpolation coefficients per block, enabling multi-scale synthesis.
• Aggregate block outputs by summing their interpolations to form the final horizon forecast, with backcast residuals guiding successive blocks.

Experimental results

Research questions

• RQ1 How can multi-rate data sampling and hierarchical interpolation improve long-horizon forecasting accuracy and efficiency?
• RQ2 Can forecasts be decomposed into interpretable frequency-band components that correspond to different time scales?
• RQ3 What is the impact of architectural components (multi-rate sampling vs hierarchical interpolation) on performance and compute, across diverse datasets?
• RQ4 Is a univariate model with shared parameters competitive with multivariate Transformer-based baselines on long horizons?

Key findings

• N-HiTS achieves average MAE/MSE improvements around 11–16% over strong Transformer baselines on six long-horizon datasets.
• The method delivers approximately 20% average accuracy improvement over recent Transformer architectures and reduces computation time by about 50x.
• Ablation studies show that combining both multi-rate sampling and hierarchical interpolation yields the best performance; removing either component degrades accuracy.
• N-HiTS provides interpretable non-linear decomposition of forecasts into frequency bands, enabling insight into trends and seasonality across blocks.