Forecast First, Ask Later: How DCATS Makes Time Series Smarter with LLMs

When it comes to forecasting traffic patterns, weather, or financial activity, the prevailing wisdom in machine learning has long been: better models mean better predictions. But a new approach flips this assumption on its head.

Instead of chasing ever-more complex architectures, the DCATS framework (Data-Centric Agent for Time Series), developed by researchers at Visa, suggests we should first get our data in order—and let a language model do it.

The Agentic Turn in AutoML

DCATS builds on the trend of integrating Large Language Model (LLM) agents into AutoML pipelines, but with a twist. While prior systems like AIDE focus on automating model design and hyperparameter tuning, DCATS delegates a more fundamental task to its LLM agent: curating the right data.

This is a subtle but radical shift. DCATS treats time series metadata—city population, road type, volume patterns—not just as documentation, but as actionable knowledge. The LLM-agent reasons over this metadata to suggest which auxiliary time series to include in training. It’s like asking a data-savvy assistant, “To forecast traffic in Campbell, CA, which nearby sensors or similar towns should I learn from?”

Here’s how it works:

What emerges is an iterative, explainable data selection loop—a sort of AutoRAG (retrieval-augmented generation) for time series forecasting.

Why Focus on the Data?

The logic behind DCATS is grounded in recent observations: simple models often outperform sophisticated ones when trained on better data. Lightweight forecasters like SparseTSF and UltraSTF can match or exceed Transformer-based models—if the input data is clean, relevant, and context-aware.

This puts DCATS in line with the broader data-centric AI movement, which argues that automation should start with understanding and improving datasets, not just chasing architectural novelty. It also opens a new frontier: AutoML for dataset construction, not just pipeline configuration.

Results That Matter

Tested on the LargeST traffic dataset, covering 8,600 California sensors, DCATS delivered a consistent ~6% improvement across MAE, RMSE, and MAPE metrics for all four forecasting models it augmented:

While a few percentage points might seem modest, they are highly impactful in domains like traffic forecasting, where even a 1% boost in accuracy can optimize millions in logistics, congestion planning, or ride-share pricing.

More Than a Black Box

What’s equally compelling is how transparent the process becomes. Every decision made by the LLM-agent is documented—why a certain neighbor was chosen, what metadata it leveraged, and how performance changed across iterations.

This isn’t just about performance gains. It’s about human-compatible reasoning in machine learning pipelines. Forecasting doesn’t remain a mysterious output—it becomes a dialogue.

Where This Could Go

The DCATS framework is a powerful signal of where AutoML and time series forecasting might be headed:

• Data-aware AutoML: Future systems may not just tweak model knobs—they’ll reason about why certain data sources belong in the training set.

• Multimodal metadata fusion: Combining time series with spatial, demographic, or even textual metadata could unlock higher accuracy with less data.

• Domain-transferable forecasting agents: Trained on traffic, deployable on energy, finance, or supply chains—if the metadata is rich enough.

Final Thoughts

DCATS demonstrates that LLMs aren’t just chatbots or coding assistants—they can be principled curators of data. In the often-overlooked world of time series forecasting, it shows how reasoning over metadata might matter more than wrestling with models.

When forecasting matters, maybe the smartest question isn’t “What model should I use?” but “What data should I include?”

Cognaptus: Automate the Present, Incubate the Future.