# N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting

**Authors:** Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza, Max Mergenthaler, Artur Dubrawski  
**arXiv:** [2201.12886](https://arxiv.org/abs/2201.12886)  
**Date:** 2022-01-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.12886v1/

## Key points

- Authors: Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza, Max Mergenthaler, Artur Dubrawski
- Introduces N-HiTS model for long-horizon forecasting challenges
- Challenges in long-horizon forecasting: prediction volatility and computational complexity
- N-HiTS tackles challenges with hierarchical interpolation and multi-rate data sampling techniques
- Sequential generation of predictions emphasizing components with varying frequencies and scales
- Outperforms state-of-the-art methods for long-horizon forecasting by 25%
- Reduces computation time significantly compared to Transformer architectures
- Code available on GitHub repository

## Summary

In their paper titled "N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting," authors Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza, Max Mergenthaler, and Artur Dubrawski introduce the novel model N-HiTS to address the challenges of long-horizon forecasting. Despite recent advancements in neural forecasting and improved performance of large-scale systems, long-horizon forecasting remains complex due to prediction volatility and computational complexity. The N-HiTS model tackles these challenges by incorporating innovative hierarchical interpolation and multi-rate data sampling techniques. These allow for sequential generation of predictions while emphasizing components with varying frequencies and scales by decomposing the input signal and synthesizing the forecast. Through extensive empirical evaluation, the authors demonstrate that N-HiTS outperforms state-of-the-art methods for long-horizon forecasting. On various multivariate tasks, N-HiTS shows an average accuracy improvement of 25% compared to the latest Transformer architectures while significantly reducing computation time. The code for implementing N-HiTS is available in their GitHub repository. Overall, this paper presents a promising approach to improving accuracy and efficiency in long-horizon forecasting through advanced interpolation and data sampling techniques in neural networks.