NHITS¶
NHITS (Neural Hierarchical Interpolation for Time Series) Forecasting.
N-HITS is particularly suitable for hierarchical time series forecasting tasks,
where the data can be naturally decomposed into multiple levels of aggregation
such as sales data at different geographical levels or product hierarchies
how it works¶
It utilizes a special technique called hierarchical interpolation to efficiently forecast future values. This technique breaks down the forecasting process into smaller components focusing on different frequencies within the data.
Additionally, NHITS employs multi-rate input processing, meaning it can analyze data at various granularities to capture both short-term and long-term patterns.
Strengths¶
Accuracy: NHITS boasts high accuracy, especially for short-term forecasting, often outperforming established models like LSTMs.
Complex seasonality: It excels at capturing intricate seasonal patterns and trends in time series data.
Efficiency: Compared to other high-accuracy models, NHITS can be computationally efficient, requiring less training time in some cases.
Flexibility: It can handle multivariate forecasting (multiple time series at once) and incorporate past covariates (additional data points that might influence the forecast).
Weaknesses¶
Data requirements: While generally efficient, NHITS might still require more data for training compared to simpler models.
Interpretability: Similar to other neural networks, understanding how NHITS arrives at predictions can be challenging.
help¶
A research paper on NHITS: https://arxiv.org/abs/2201.12886
Documentation on NHITS implementation: https://www.nixtla.io/