N-BEATS

The core architectural innovation of N-BEATS is its doubly residual stacking principle. The architecture is composed of stacks, each containing multiple blocks. Each block receives the residual signal from the previous block's forecast, producing a partial forecast that is subtracted from the input. Simultaneously, the block's output is summed to produce the final global forecast. This dual residual flow—one backward through the input signal, one forward through the forecast aggregation—allows the network to iteratively refine its prediction without vanishing gradient issues.

N-BEATS enforces interpretability by constraining the output layer of each block to learn specific basis functions. In the generic architecture, blocks learn a polynomial approximation of the signal. In the interpretable configuration, stacks are specialized:

• Trend Stack: Constrained to learn a monotonic polynomial function, capturing long-term directional movement.
• Seasonality Stack: Constrained to learn a periodic Fourier series, capturing cyclical patterns. This explicit decomposition into trend and seasonality mimics classical time series decomposition but is learned end-to-end via gradient descent.

Each N-BEATS block is a simple yet effective multi-layer perceptron (MLP) with ReLU activations. The block takes the lookback window as input and passes it through four fully connected layers before branching into two output heads:

• Backcast (θb): The block's reconstruction of the input signal, used to compute the residual passed to the next block.
• Forecast (θf): The block's contribution to the final prediction at the specified horizon. This dual-output design is what enables the doubly residual topology, forcing each block to explain a portion of the signal before passing the unexplained remainder forward.

Unlike classical methods like ARIMA or even hybrid deep learning approaches, N-BEATS does not require manual feature engineering or domain-specific preprocessing. The model ingests raw, univariate historical values directly. It learns to implicitly extract relevant patterns—trend, seasonality, and complex non-linear dynamics—purely from the data. This makes it highly portable across different forecasting domains, from retail demand to financial time series, without requiring the practitioner to handcraft lag features, moving averages, or holiday indicators.

N-BEATS achieved state-of-the-art performance on the M4 competition dataset, a widely recognized benchmark for univariate time series forecasting. The model outperformed a large ensemble of classical statistical methods and other neural network architectures by a statistically significant margin. Key performance characteristics:

• Pure DL approach: No ensembling with statistical models required.
• Horizon flexibility: A single architecture handles multiple forecast horizons.
• Interpretability trade-off: The generic variant achieves highest accuracy; the interpretable variant sacrifices minimal accuracy for full decomposability.

N-BEATS does not predict time steps directly. Instead, it learns to predict expansion coefficients that are projected onto a set of predefined basis functions.

• Trend basis: Models monotonic behavior using polynomial functions (e.g., a linear ramp).
• Seasonality basis: Models cyclical patterns using Fourier series (sine/cosine waves).
• This forces the network to decompose the signal structurally, acting as a learned, deep version of classical time series decomposition.

The architecture's core innovation is its doubly residual topology, which creates a clean, interpretable decomposition path.

• Backward residual: Each block receives the residual signal from the previous block, iteratively removing explainable structure.
• Forward residual: Each block outputs a partial forecast, and the sum of all partial forecasts across blocks forms the final prediction.
• This design prevents gradient interference and ensures each block specializes in a distinct signal component.

N-BEATS offers two distinct configurations, trading off flexibility for human-readability.

• Generic architecture: The basis functions are learned end-to-end by the network, maximizing predictive accuracy without human priors.
• Interpretable architecture: The basis functions are hard-coded as trend and seasonality waveforms, allowing practitioners to extract and visualize the exact decomposition the model learned.
• This dual-mode design makes it uniquely suited for both pure performance and auditable forecasting.

N-BEATS represents a paradigm shift from statistical to pure deep learning for univariate forecasting.

• vs. ARIMA/ETS: N-BEATS makes no assumptions about stationarity or data-generating processes; it learns the structural priors directly from data.
• vs. DeepAR: Unlike autoregressive RNNs, N-BEATS does not feed predictions back as inputs, avoiding error accumulation in multi-step forecasts.
• vs. TFT: N-BEATS is univariate by design and does not natively ingest static or known future covariates, making it ideal for large-scale, isolated SKU forecasting.

N-BEATS can be trained as a meta-learner across many heterogeneous time series, enabling few-shot forecasting.

• Cross-series learning: A single N-BEATS model can learn shared structural patterns (e.g., weekly seasonality) from thousands of different SKUs.
• Fast adaptation: When a new time series with limited history is introduced, the model can generate a reasonable forecast by leveraging patterns learned from the broader dataset.
• This property directly addresses the cold-start problem in retail and inventory management.

The additive nature of the forward residual connections allows for direct visual inspection of the forecast composition.

• A final forecast is the sum of outputs from multiple stacks (e.g., Trend Stack + Seasonality Stack).
• Practitioners can plot each stack's output to see exactly how much of the prediction is driven by a long-term trend versus a weekly cycle.
• This provides a level of algorithmic explainability that is rare in deep learning, satisfying governance requirements for high-stakes supply chain decisions.