N-BEATS

As explained in Time Series Forecasting Using Foundation Models

N-BEATS is a neural forecasting model organized as stacks of blocks for modeling time series without requiring explicit series decomposition or input transformations.

N-BEATS is a neural forecasting model that learns time-series relationships through stacks of fully connected blocks and basis expansion, without explicit decomposition or input transformations. Its generic architecture makes it useful for exploring pretrained forecasting models and transfer learning. section 2.0 section 2.1 section 2.2

Each block produces a forecast and a backcast. The forecast contributes predicted future values; the backcast reconstructs information learned from the input. Within a stack, later blocks receive residuals after earlier backcasts are subtracted, allowing successive blocks to capture what remains. Stack predictions are then summed into the final forecast. section 2.2.1 section 2.2.2 section 2.2.3

Figure 2.2 Architecture of N-BEATS. N-BEATS consists of stacks of blocks. Each block outputs a forecast and a backcast.

Use the architecture diagram to follow forecast aggregation and residual flow from blocks through stacks. section 2.2

A practical N-BEATS workflow includes:

• Set the forecast horizon and input size, then initialize NBEATS through neuralforecast. section 2.4
• Pretrain with a chosen number of gradient-update steps, and save the resulting model for reuse. section 2.4
• Load the model and perform zero-shot forecasting by passing an unseen dataset without additional training. section 2.5
• Compare pretrained, fine-tuned, and data-specific variants with plots, MAE, and sMAPE. section 2.7

Monthly versus daily forecasting

A model pretrained on monthly M3 data achieved MAE 1.59 and sMAPE 3.56% with zero-shot forecasting on Australian monthly drug prescriptions. Applied unchanged to daily data, it reached MAE 2.59 and sMAPE 8.69%, while a data-specific model reached 1.34 and 4.87%. section 2.7 section 2.8

How Time Series Forecasting Using Foundation Models covers it

How deep it goes

Explains blocks, basis expansion, backcasts, residuals, stacks, training steps, and evaluation, but not detailed layer mathematics.

Examples

Demonstrates monthly M3 pretraining, Australian drug-prescription forecasting, daily temperature forecasting, prediction plots, MAE, and sMAPE.

What you can do after

Apply N-BEATS via neuralforecast, reuse a pretrained model, generate forecasts, and evaluate alternatives with MAE and sMAPE.

How it connects

Connects basis expansion and fully connected layers with residual connections, transfer learning, zero-shot forecasting, and foundation models.

liveBookmanning.com

Read Time Series Forecasting Using Foundation Models on liveBook Get the book at manning.com

Related concepts

• fully connected neural networkuses · A fully connected neural network is a feedforward neural architecture in which each unit in a layer is connected to every unit in the adjacent layer.
• language modeluses · A language model is a computational model that represents patterns in linguistic tokens to estimate sequence probabilities and predict or generate language from context.
• M3 datasetuses · The M3 dataset is a collection of 3,003 time series from the M3 forecasting competition sampled at yearly, quarterly, and monthly frequencies.

This explainer was generated by Manning's knowledge graph from the text of Time Series Forecasting Using Foundation Models. The section links point to the parts of the book it draws on.