A new paper from Google titled “Nexus: An Agentic Framework for Time Series Forecasting” presents a novel approach to forecasting that emphasizes the importance of contextual understanding over mere historical data. Nexus redefines forecasting as a reasoning problem, claiming that models should interpret the surrounding events and not just the numerical data. This aligns with recent studies indicating that large language models are more effective on quantitative reasoning tasks when guided by structured workflows, which break down forecasting into stages such as perception and planning. In practical tests, one version based on Claude achieved an 86.6% reduction in average MAPE when compared to traditional methods, suggesting that acknowledging the complexities of the world can enhance forecasting accuracy.

Nexus: Nexus is a Google-led, multi-agent framework for time series forecasting that combines language models with structured reasoning over numerical data and textual context. In this news, Nexus is the core subject of the paper “Nexus: An Agentic Framework for Time Series Forecasting,” which shows how decomposing forecasting into contextualization, dual-resolution outlooks, and synthesis can significantly improve performance over standard chain-of-thought prompting.
Claude: Claude is Anthropic’s family of large language models designed for reasoning-intensive tasks, coding, and natural language understanding. In the Nexus paper, a Claude-based agent configuration is used as one of the forecasting backbones, demonstrating large gains in accuracy when wrapped in Nexus’s structured, multi-agent workflow compared with prompting Claude directly.
Zillow: Zillow is a major U.S. online real estate marketplace that aggregates property listings, valuations, and housing market data. In the Nexus study, Zillow’s housing inventory and related real estate time series are used as one of the main evaluation domains to show how context-aware, agentic forecasting can better explain and predict housing market dynamics.

`json
{
“LLM_structure”: “Research indicates that large language models often perform more effectively on quantitative reasoning tasks when guided by structured workflows or dedicated agents rather than as single comprehensive systems.”,
“Agentic_forecasting”: “Studies on agentic time series forecasting suggest that dividing forecasting into stages like perception, planning, and reflection can produce more robust and interpretable results compared to one-pass model predictions.”,
“Context_integration”: “In fields such as finance and real estate, there is an increasing focus on integrating unstructured textual context like news, policy changes, and company reports into time series models to detect event-driven changes that pattern-based learning might overlook.”
}
`