N-BEATS

N-BEATS does not use any time-series-specific feature engineering or input scaling. The model learns a hierarchical decomposition of the signal directly from the target history. Each block outputs a backcast (to remove explained signal) and a forecast (partial prediction), which are aggregated across blocks. This design eliminates the need for domain-specific pre-processing like Fourier transforms or moving averages.

The architecture enforces interpretable basis functions within its blocks:

• Trend block: Uses a polynomial basis to model monotonic or slowly-varying patterns, constrained to a small degree p (e.g., p=2 for linear, p=3 for quadratic).
• Seasonality block: Uses a Fourier basis with sine and cosine waveforms to capture periodic fluctuations. This structural inductive bias makes the model's internal decomposition human-readable, unlike black-box RNNs or Transformers.

N-BEATS organizes blocks into stacks, each specializing in a signal component. The architecture uses two residual connections per block:

• The backcast residual subtracts the block's approximation of the input, passing the unexplained remainder to the next block.
• The forecast residual aggregates partial forecasts from all blocks to form the final prediction. This topology allows deep stacking (e.g., 30 blocks) without vanishing gradients, enabling the model to iteratively refine its decomposition.

N-BEATS offers two operational modes:

• Generic mode: All blocks use a fully-learned basis, maximizing predictive accuracy without enforcing interpretability. This variant won the M4 forecasting competition.
• Interpretable mode: Blocks are explicitly assigned to trend or seasonality stacks with constrained basis functions, enabling analysts to inspect and validate the decomposition logic. The choice depends on whether the use case prioritizes raw accuracy or model transparency for regulatory or risk management purposes.

Each N-BEATS block contains a simple multi-layer perceptron (MLP) with fully connected layers and ReLU activations. The MLP takes the input window and generates expansion coefficients (theta), which are then projected onto the block's basis functions to produce backcast and forecast signals. This design is computationally efficient, leveraging matrix multiplications that are highly optimized on GPUs, and avoids the sequential bottlenecks of recurrent architectures.

N-BEATS achieved first place in the M4 forecasting competition (2018), outperforming both statistical benchmarks and hybrid ML approaches across 100,000 diverse time series. The winning configuration used a generic model with 30 blocks and no time-series-specific pre-processing. This result demonstrated that a pure deep learning approach, with the right architectural priors, can surpass heavily engineered ensembles of exponential smoothing and ARIMA models.

N-BEATS uses a pure deep learning architecture with no feature engineering, time-series-specific components, or handcrafted inputs. The network learns a hierarchical basis expansion where each block produces a partial forecast by projecting the input onto learned basis functions. The architecture consists of:

• Trend blocks: Model low-frequency, monotonic components using polynomial basis functions
• Seasonality blocks: Capture cyclical patterns using Fourier-like harmonic basis functions
• Generic blocks: Learn arbitrary basis functions when decomposition is not enforced The final forecast is the sum of all block outputs, creating an interpretable additive decomposition without any statistical assumptions.

N-BEATS employs a doubly residual topology that enables hierarchical decomposition of the forecasting signal. Each block receives the backcast (residual of the input signal) from the previous block and outputs both a forecast and a backcast. The architecture operates as follows:

• Block input: Receives the residual signal not explained by previous blocks
• Block output 1 (backcast): The portion of the input signal explained by this block, subtracted from the input for the next block
• Block output 2 (forecast): The block's contribution to the final prediction, summed across all blocks This topology ensures each subsequent block focuses on the unexplained residual, preventing redundancy and enabling clean signal decomposition.

Each N-BEATS block contains a simple multi-layer perceptron (MLP) with fully connected layers and ReLU activations. The block's final layer projects the hidden representation into expansion coefficients for the basis functions. Key design choices:

• No recurrent connections: Enables parallel training across the entire sequence
• No attention mechanisms: Avoids quadratic complexity while maintaining performance
• Shared block structure: All blocks have identical architecture but learn different basis functions
• Theta parameters: Each block learns both expansion coefficients (theta_f for forecast, theta_b for backcast) from the same hidden representation This simplicity allows N-BEATS to train faster than recurrent or transformer-based alternatives while achieving state-of-the-art accuracy on M3, M4, and TOURISM competitions.

N-BEATS offers a configurable inductive bias through its interpretable architecture variant. When configured for decomposition, the network enforces a clean separation of signal components:

• Trend basis: Uses polynomial functions (e.g., linear, quadratic) to model slow-moving directional components. The trend is constrained to be monotonic or slowly varying.
• Seasonality basis: Uses periodic sine and cosine functions with learned frequencies and phases to capture repeating patterns.
• Stack organization: Multiple trend blocks followed by seasonality blocks create a hierarchical decomposition from low to high frequency. This interpretable mode allows practitioners to inspect and validate each component separately, making the model suitable for regulated financial applications where explainability is mandatory.

N-BEATS achieved state-of-the-art results in the M4 forecasting competition, a benchmark spanning 100,000 time series across diverse domains. Key performance metrics:

• Overall winner in the pure machine learning category
• 3% improvement in symmetric mean absolute percentage error (sMAPE) over the statistical benchmark
• First pure DL model to outperform statistical ensembles on this benchmark
• Demonstrated cross-domain generalization without any domain-specific tuning The model's success validated that a generic, purely data-driven architecture could outperform carefully engineered statistical methods across business, finance, demographics, and economics domains without any feature engineering.

While transformers dominate NLP, N-BEATS offers distinct advantages for univariate time-series forecasting:

• Computational efficiency: O(n) complexity vs O(n²) for self-attention, enabling faster training on long sequences
• No positional encoding needed: The sequential nature is implicit in the residual stacking
• Interpretable outputs: Direct signal decomposition vs attention weight inspection
• Lower data requirements: MLP-based architecture trains effectively on smaller datasets However, transformers excel at multivariate forecasting and capturing cross-series dependencies. For pure univariate tasks, N-BEATS often matches or exceeds transformer performance with significantly less compute. The choice depends on whether cross-series interactions are critical to the forecasting problem.