> ML_ALGORITHM // NEURAL-HIERARCHICAL-INTERPOLATION-N-HITS_v1.0
N-HiTS (Neural Hierarchical Interpolation)
Hierarchical neural forecasting architecture incorporating multi-rate input sampling and non-linear interpolation, dramatically outperforming Transformers on long horizons.
Pure Neural Forecasting Architecturetime-series-forecastinghigh-intrinsiclarge (>100k)
Back to All AlgorithmsComputational Complexity
Training Complexity:O(epochs * batch_size * blocks * downsampled_MLP)
Inference Complexity:O(downsampled_MLP)
Hardware Profile
CPU Friendly:Yes
Requires GPU:Yes
Memory Footprint:low
Interpretability & Data
Interpretability Tier:high-intrinsic
Training Data Needs:large (>100k)
Interpretability Assessment
Multi-rate pooling enforces that slow stacks synthesize low-frequency macro trend while fast stacks capture high-frequency ripples.
Suitable Tasks & Supported Modalities
Suitable Tasks:
long horizon forecastinghigh frequency telemetry
Supported Modalities:
time-series
Foundational Literature
Common Pitfalls & Warnings
- Subsampling too aggressively in the final block causes aliasing artifacts and loss of sharp step-transitions