AutoBNN: AI That Predicts Time Series With Confidence

Time series forecasting has long been a computational challenge, balancing the need for accuracy with the practical limitations of interpretability and scale. From predicting weather patterns to forecasting economic trends, traditional methods often force researchers to choose between statistical rigor and computational feasibility.

Google Research has released AutoBNN, an open-source package that aims to bridge this divide. The tool combines Bayesian neural networks with compositional kernel structures, offering what developers describe as a more scalable alternative to traditional Gaussian processes. This approach maintains the uncertainty quantification that makes Bayesian methods valuable while leveraging neural network efficiency.

The system builds on a decade of research into Gaussian processes with learned kernel structures. These kernels encode assumptions about data patterns—whether trends, seasonal variations, or noise—through compositional building blocks. AutoBNN replaces the Gaussian process backbone with Bayesian neural networks while preserving this modular architecture.

Bayesian neural networks differ from standard neural networks by treating weights as probability distributions rather than fixed values. This creates inherent uncertainty estimates in predictions, addressing a key limitation of conventional deep learning approaches for time series analysis. The shift to BNNs brings computational advantages, with training times scaling linearly rather than cubically with data size.

AutoBNN introduces several innovations, including WeightedSum operators that allow parallel exploration of model structures. Instead of testing combinations sequentially, the system can evaluate multiple configurations simultaneously through learnable mixing weights. This "soft" structure discovery enables more efficient model selection without expensive discrete optimization.

The package includes pre-defined model templates like sumofstumps and sumofshallow that incorporate these WeightedSum operations. Researchers demonstrated the approach on financial data from the M3 dataset, where the system automatically identified periodic patterns as the most significant component while downplaying less relevant factors.

Implementation occurs through TensorFlow Probability using JAX and the flax.linen neural network library. The system offers both MAP and MCMC estimation methods, supporting various likelihood functions for different data types. Google provides example applications, including analysis of the Mauna Loa CO2 dataset, where the model captured seasonal trends while providing confidence intervals for future projections.

As organizations increasingly rely on predictive analytics for decision-making, tools that offer both accuracy and interpretability could see broad adoption. AutoBNN's open-source availability positions it for potential use across scientific research, financial modeling, and operational forecasting applications where understanding prediction uncertainty matters as much as the predictions themselves.