N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting

Before N-BEATS, statistical methods like ARIMA, Exponential Smoothing (ETS), and the Theta method dominated forecasting competitions. These methods explicitly model trend, seasonality, and noise — components that practitioners understand and can inspect.

Deep learning models attempted to replace this with end-to-end learning, but they consistently underperformed. The reasons were threefold: (1) forecasting datasets are often small — thousands of short series, not millions of images; (2) standard architectures like RNNs and LSTMs added complexity without sufficient inductive bias for the task; (3) practitioners distrusted black-box models whose outputs could not be decomposed into interpretable components.

The M4 competition in 2018 made this gap painfully visible. Its 100,000 time series spanned yearly, quarterly, monthly, weekly, daily, and hourly frequencies from domains including finance, demographics, and industry. Pure ML methods were comprehensively outperformed.

The atomic unit of N-BEATS is the basic block. Think of it as a worker on an assembly line who receives a piece of the time series signal (the lookback window), does two things simultaneously, and passes the remainder downstream.

The block takes an input $\mathbf{x}_\ell$ (a window of past observations) and passes it through a stack of fully connected layers with ReLU activations. This produces a hidden representation $\mathbf{h}_\ell$ that captures patterns in the input. From this representation, the block produces two sets of expansion coefficients $\theta^b_\ell$ and $\theta^f_\ell$ via separate linear layers.

These coefficients are then multiplied by basis functions $g^b$ and $g^f$ to produce two outputs: a backcast $\hat{\mathbf{x}}_\ell = g^b(\theta^b_\ell)$ which is the block's best reconstruction of its input, and a forecast $\hat{\mathbf{y}}_\ell = g^f(\theta^f_\ell)$ which is the block's prediction of future values.

The mental model is simple: the backcast says "here is what I understood about the past", the forecast says "here is what I predict for the future based on what I understood."

The backcast is not just a diagnostic — it is the mechanism that makes the whole architecture work. After block $\ell$ produces its backcast $\hat{\mathbf{x}}_\ell$, the input to the next block is the residual: the original input minus what block $\ell$ explained.

This is the backward residual: each block peels away the signal component it captured, so the next block works on a cleaner, simpler remainder. Imagine stripping layers of paint from a canvas — each layer reveals what is underneath.

Simultaneously, each block's forecast $\hat{\mathbf{y}}_\ell$ is accumulated into the final prediction. This is the forward residual: the overall forecast is the sum of partial forecasts from all blocks. No single block needs to predict the entire future — each contributes its piece.

Together, these two residual paths give the architecture its name: doubly residual stacking. The backward path ensures progressive signal decomposition, while the forward path ensures all components contribute to the final prediction.

Blocks are grouped into stacks. Within each stack, blocks share the same type of basis functions $g^b$ and $g^f$, and in the interpretable configuration they even share weights across blocks in the same stack.

In the generic (N-BEATS-G) configuration, each block uses learnable linear layers as basis functions — no constraints, maximum expressiveness. The model typically uses 30 stacks with 1 block each.

In the interpretable (N-BEATS-I) configuration, only two stacks are used: a trend stack followed by a seasonality stack, each containing 3 blocks. The trend stack uses polynomial basis functions, the seasonality stack uses Fourier basis functions. The doubly residual stacking ensures that the trend is removed before the seasonality stack sees the signal — exactly mirroring classical decomposition.

The interpretable configuration constrains the basis functions to known mathematical forms. The key idea is that if we force the network to express its outputs using specific shapes, those shapes become human-readable.

Trend basis: a polynomial of small degree $p$. Given a time vector $\mathbf{t} = [0, 1, \ldots, H-1]/H$, the trend output is $g^f_t(\theta) = \sum_{i=0}^{p} \theta_i \cdot \mathbf{t}^i$. A degree-2 polynomial captures constant, linear, and quadratic trends — the same shapes a human analyst would sketch. The network does not learn the polynomial structure — it only learns the coefficients $\theta$.

Seasonality basis: a truncated Fourier series. The basis functions are $\cos(2\pi k \mathbf{t})$ and $\sin(2\pi k \mathbf{t})$ for harmonics $k = 1, \ldots, K$. This constrains the output to periodic patterns — exactly the shapes needed to capture weekly, monthly, or yearly cycles.

N-BEATS uses model ensembling as its primary regularization strategy — not dropout or L2 penalties. The authors found that training multiple models with different random seeds and averaging their forecasts was far more effective than standard regularization.

The final N-BEATS ensemble averages 180 models: 3 lookback window lengths × 3 loss functions (MAPE, MASE, sMAPE) × 20 random seeds. This ensemble width is what closes the gap from strong to state-of-the-art. The diversity in lookback windows means some models focus on short-term patterns while others capture longer dependencies. The diversity in loss functions ensures the ensemble is robust to different error metrics.

On the M4 dataset (100K series), N-BEATS-G achieved an OWA of 0.821 — beating the M4 competition winner (ES-RNN with OWA 0.838) by 3% and the statistical benchmark by 11%. The interpretable variant N-BEATS-I achieved OWA 0.847, which still outperformed the M4 winner.

On M3 (3,003 series) and TOURISM (1,311 series), N-BEATS similarly achieved state-of-the-art results. These datasets span vastly different domains and frequencies, demonstrating that the architecture generalizes without any domain-specific modifications.

The ablation study confirmed that doubly residual stacking (vs parallel or last-forward alternatives) was essential. Removing the backward residual degraded performance significantly, validating the core architectural innovation.

N-BEATS fundamentally changed how the forecasting community views deep learning. Before it, the question was whether DL could work at all for time series. After it, the question became how to build on its success. Its core ideas — progressive residual decomposition, basis expansion for interpretability, and heavy ensembling — have become foundational patterns in neural forecasting.