The era of relying solely on historical numerical data for critical business predictions is ending. Salesforce AI Research has unveiled MoiraiAgent, an agentic framework that fundamentally redefines time-series analysis by integrating qualitative, real-world context alongside traditional metrics. This shift toward Agentic time-series forecasting acknowledges that modern outcomes are driven as much by policy shifts and external shocks as by trends and seasonality.
The core innovation lies in moving from a static model pipeline to a dynamic, LLM-coordinated agent. Traditional models often fail when external factors—like regulatory changes or supply chain disruptions—create regime shifts that historical data cannot predict. MoiraiAgent uses a powerful large language model to reason over heterogeneous data, allowing it to interpret natural language context and apply that insight to numerical prediction tasks. This capability is essential for enterprise use cases, where early warning systems and capacity planning require adaptive, real-time responsiveness to external variables like weather or policy announcements.
A critical practical feature is the intelligent expert-selection mechanism. Recognizing that no single time-series foundation model dominates all domains—Chronos-2 excels in WebOps, while others struggle with Nature data—MoiraiAgent employs a lightweight 3B-parameter LLM to choose the best fit for the specific task. This selector analyzes historical values, temporal characteristics, and cross-validation errors from candidate models, which include state-of-the-art systems like TimesFM-2.5 and Tirex. The result is a system that consistently outperforms all individual state-of-the-art models on the GIFT-Eval benchmark, demonstrating the power of orchestration over singular model strength.
Contextual Reasoning: Beyond Pre/Post-Processing
The agentic framework applies context in three sophisticated ways: trimming the lookback window during phase transitions, refining anomaly detection by identifying non-persistent patterns, and modifying raw predictions based on anticipated future events. This is far more complex than simple data cleaning; it involves dynamic tool orchestration, where the LLM might invoke a Python sandbox to remove abnormal impacts or adjust the historical window to focus on a new operational regime. To validate this complex reasoning, the team introduced GIFT-CTX, a new benchmark specifically designed to test joint reasoning over numerical series and natural language context, where neither input alone suffices for accuracy.
The performance metrics on GIFT-CTX are highly instructive, revealing the limitations of existing approaches. MoiraiAgent significantly outperformed specialized time-series foundation models and frontier generalist LLMs like GPT-5.2 and Gemini 3.0 Pro in contextual scenarios. According to the announcement, this validates the necessity of a dedicated, tool-orchestrated framework for contextual forecasting, rather than relying on generalist LLMs that struggle with long numerical sequences. This sets a new standard for accuracy in dynamic environments, proving that context-aware prediction is now a measurable, achievable goal for complex enterprise systems.
MoiraiAgent represents a necessary evolution in predictive analytics, moving forecasting from a statistical exercise to a form of applied reasoning. The integration of LLMs as dynamic coordinators, capable of selecting experts and interpreting qualitative data, signals the future direction for all high-stakes forecasting systems. While the current framework uses pre/post-processing for context, the next frontier will involve end-to-end models that embed contextual signals directly within the numerical modeling process, ensuring that predictions are inherently responsive to the complex, unpredictable forces of the real world.
