N-BEATS

N-BEATS operates directly on raw historical values without requiring manual extraction of trend, seasonality, or holiday effects. The architecture learns a hierarchical decomposition internally through stacked blocks of fully connected layers. Each block produces backcast and forecast signals; the backcast is subtracted from the input before passing to the next block, forcing the network to sequentially model residual patterns. This eliminates the need for domain-specific preprocessing and allows the model to discover latent structures automatically.

The architecture can be configured with interpretable basis functions that enforce specific structural decompositions. In the interpretable configuration, each block's output is projected onto learned trend and seasonality bases:

• Trend basis: A polynomial function of time (e.g., degree p) that captures monotonic or slowly varying patterns
• Seasonality basis: A Fourier series that captures periodic fluctuations This constraint produces human-readable decompositions where each block's contribution can be visually inspected, addressing the black-box criticism of deep learning while maintaining predictive accuracy.

N-BEATS employs a doubly residual stacking architecture with two distinct residual connections per block:

• Backcast residual: The block predicts the portion of the input it can explain and subtracts it, passing the unexplained remainder to the next block
• Forecast residual: Each block's partial forecast is summed to produce the final prediction This topology allows the network to decompose the signal hierarchically, with earlier blocks capturing dominant patterns and deeper blocks modeling finer residuals. The architecture naturally prevents gradient vanishing through its residual design.

N-BEATS offers two operational modes that trade off flexibility against explainability:

• Generic configuration: Blocks use unconstrained basis functions learned end-to-end, maximizing predictive performance on arbitrary time series
• Interpretable configuration: Blocks are constrained to trend and seasonality bases, producing decompositions that domain experts can validate Both configurations share the same doubly residual topology. The generic variant won the M4 forecasting competition, while the interpretable variant provides audit-ready outputs for regulated industries like pharmaceutical supply chains.

N-BEATS achieved state-of-the-art performance on the M4 forecasting competition, which includes 100,000 diverse time series across business, financial, and economic domains. The model outperformed both traditional statistical methods and hybrid ML approaches by a significant margin:

• sMAPE improvement: 3% reduction in symmetric mean absolute percentage error compared to the competition benchmark
• OWA improvement: Achieved the lowest overall weighted average across all forecasting horizons This victory validated pure deep learning as a viable alternative to statistical forecasting methods for univariate time series.

N-BEATS integrates naturally into ensemble forecasting frameworks. Multiple N-BEATS instances can be trained with different:

• Horizon lengths: Specializing in short, medium, or long-range predictions
• Block counts: Varying model depth to capture different levels of signal complexity
• Basis configurations: Mixing generic and interpretable blocks
• Random initializations: Leveraging the stochastic nature of SGD training Averaging predictions across this diverse ensemble typically yields a Continuous Ranked Probability Score (CRPS) improvement of 5-10% over a single model, making it ideal for probabilistic demand forecasting where uncertainty quantification is critical.

The core mechanism of N-BEATS. Instead of predicting values directly, the network predicts expansion coefficients for a set of predefined basis functions. These functions represent generic patterns like trends and seasonality. The final forecast is the sum of these weighted basis functions, ensuring the output is a structured, interpretable decomposition rather than an arbitrary vector. This forces the network to learn a hierarchical decomposition of the signal.

The architectural backbone of N-BEATS. The network is organized into stacks, and each stack consists of multiple blocks. Each block produces two outputs: a partial forecast and a backcast. The backcast is subtracted from the input before passing it to the next block, and the partial forecasts are aggregated. This double residual principle allows deeper stacks to refine the signal by removing components already explained by earlier blocks, leading to state-of-the-art performance on univariate benchmarks.

A specific instantiation of N-BEATS where the basis functions are constrained to human-understandable forms. The architecture uses two stacks:

• Trend Stack: Uses polynomial basis functions to model monotonic or slowly varying behavior.
• Seasonality Stack: Uses Fourier series basis functions to model periodic, cyclical patterns. This configuration decomposes a time series into its structural components without sacrificing predictive accuracy, making it highly valuable for supply chain planners who need to explain forecasts.

An alternative N-BEATS instantiation where the basis functions are not pre-defined but are learned end-to-end from the data. The network learns its own latent basis for the signal. While this configuration can achieve slightly higher accuracy on some datasets by discovering non-standard patterns, it sacrifices the direct interpretability of the trend and seasonality decomposition. It is the preferred choice when raw predictive performance is the sole objective.

N-BEATS is a pure univariate model, meaning it forecasts a single time series based solely on its own history. It does not natively ingest exogenous covariates like price, promotions, or weather. This design choice is a deliberate departure from models like the Temporal Fusion Transformer. The architecture's strength lies in its ability to extract maximum signal from the target's own past, making it an ideal component in a larger ensemble where other models handle external drivers.

N-BEATS achieved a landmark victory in the M4 forecasting competition, a major benchmark for time series methods. It outperformed a large field of statistical and machine learning models by a statistically significant margin. This result validated the hypothesis that a pure deep learning architecture, without feature engineering or domain-specific scaling, could surpass heavily tuned traditional methods on a diverse set of univariate series, cementing its status as a foundational modern forecaster.