AutoBNN: Probabilistic time series forecasting with compositional bayesian neural networks
Google just open-sourced AutoBNN. Here's why time series forecasting just got way faster.
Why it matters
Google Research released AutoBNN, an open-source JAX library that automates interpretable time series forecasting by combining Bayesian neural networks with compositional kernel structures. This addresses a critical gap: traditional Bayesian methods require domain expertise and don't scale, while neural networks lack interpretability. The tool scales linearly with data size (vs. cubic scaling for Gaussian processes) and runs efficiently on GPUs/TPUs, making probabilistic forecasting accessible for enterprises handling weather, traffic, economic, and sensor data.
The key facts
9 to knowOpen-source library written in JAX, implemented in TensorFlow Probability
Training complexity: linear in data points (vs. cubic scaling for traditional Gaussian processes)
Supports GPU/TPU acceleration for large-scale time series
Provides compositional kernel structures: Linear, Quadratic, Matérn, ExponentiatedQuadratic, Periodic, OneLayer
Includes operators: Addition, Multiplication, ChangePoint, LearnableChangePoint, WeightedSum
Enables 216 possible model structures through sum_of_products architecture
Produces interpretable models with reliable confidence intervals
Tested on M3 and Mauna Loa CO2 datasets
scikit-learn inspired API for accessibility
Go to the source
Google Research Blogblog.research.google
Publisher excerpt: Posted by Urs Köster, Software Engineer, Google Research Time series problems are ubiquitous, from forecasting weather and traffic patterns to understanding economic trends. Bayesian approaches start with an assumption about the data's patterns (prior probability), collecting evidence (e.g., new…