Forecasting the Future with Collaborative AI

by

Author: Denis Avetisyan

A new framework combines the power of time series models and large language models to deliver more accurate and explainable predictions.

The Nexus framework establishes a multi-agent system predicated on a three-stage process: contextualization of historical data into structured signals, generation of both macro and micro-scale forecasts to capture differing resolutions, and finally, synthesis of these perspectives alongside a calibration process learned from prior inaccuracies-a design intended to optimize predictive accuracy through layered analysis and iterative refinement of its forecasting model.
The Nexus framework establishes a multi-agent system predicated on a three-stage process: contextualization of historical data into structured signals, generation of both macro and micro-scale forecasts to capture differing resolutions, and finally, synthesis of these perspectives alongside a calibration process learned from prior inaccuracies-a design intended to optimize predictive accuracy through layered analysis and iterative refinement of its forecasting model.

Nexus, a multi-agent system, leverages multimodal learning and causal reasoning for state-of-the-art time series forecasting with improved calibration.

Traditional time series forecasting often struggles to integrate unstructured contextual data alongside numerical patterns, limiting its ability to capture real-world dynamics. To address this, we introduce ‘Nexus : An Agentic Framework for Time Series Forecasting’, a multi-agent system that decomposes prediction into specialized stages, effectively combining the strengths of time series foundation models and large language models. Our results demonstrate that current LLMs possess surprisingly strong intrinsic forecasting abilities when organized within an agentic framework, achieving state-of-the-art accuracy and providing interpretable reasoning traces. Does this agentic approach represent a fundamental shift towards viewing real-world forecasting as a problem of reasoning, rather than solely sequence modeling?

The Inherent Limitations of Conventional Forecasting

Conventional time series forecasting methods, while effective for relatively stable and linear trends, frequently encounter difficulties when confronted with the intricacies of real-world data. These methods often assume consistent relationships between past and future values, a simplification that breaks down in the face of non-linear dynamics – such as accelerating growth, sudden shifts, or cyclical patterns. Furthermore, external factors, like economic shocks, policy changes, or even unforeseen events, introduce influences that are not inherently captured within the historical time series itself. Consequently, forecasts generated by traditional approaches can exhibit significant inaccuracies, particularly when dealing with complex systems where multiple interacting forces are at play. The inability to adequately model these nuances highlights a critical need for more sophisticated forecasting techniques capable of accommodating both internal complexities and external disruptions.

Large Language Models, despite their impressive ability to process and understand information from diverse sources – a capability known as multimodal context – fundamentally struggle with the task of time series forecasting. These models excel at identifying relationships between concepts, but lack an intrinsic understanding of sequential data and the crucial dependencies that unfold over time. Unlike systems designed to analyze temporal patterns, LLMs treat each data point largely independently, failing to effectively capture the inertia, trends, and seasonality inherent in time-dependent phenomena. This limitation means that while an LLM might recognize what is happening in a series, it often cannot accurately predict when something will happen, or how past events influence future outcomes, hindering their performance in forecasting applications.

The challenge of forecasting in dynamic systems demands a departure from conventional methods, as reliance on either purely temporal models or those focused solely on contextual understanding proves insufficient. A novel strategy is required that actively synthesizes these two crucial elements – the ability to discern patterns extending across long time horizons and the capacity to incorporate relevant external information. This integration isn’t simply a matter of appending context to a traditional time series analysis; instead, it necessitates an architectural shift, one that allows the model to simultaneously process historical dependencies and contextual signals to generate more robust and accurate predictions. Such an approach acknowledges that future states aren’t solely determined by past values, but are also shaped by a complex interplay of external factors, demanding a forecasting system capable of understanding both the ‘when’ and the ‘why’ behind evolving patterns.

Nexus improves forecasting accuracy and reasoning by uniquely integrating macro- and micro-level time series analysis with external multimodal context, addressing the limitations of both traditional time series forecasting methods and large language model approaches that often struggle with temporal dynamics.
Nexus improves forecasting accuracy and reasoning by uniquely integrating macro- and micro-level time series analysis with external multimodal context, addressing the limitations of both traditional time series forecasting methods and large language model approaches that often struggle with temporal dynamics.

Deconstructing Temporality: A Multi-Agent Framework

Nexus employs a multi-agent system to improve forecasting accuracy by separating overall trends from detailed time series characteristics. This builds on agentic forecasting principles, where individual agents focus on specific aspects of the data; in Nexus, these agents work in parallel to deconstruct the time series into coarse-level, overarching patterns and granular, high-frequency features. This decomposition allows for independent analysis and modeling of each component, ultimately enhancing the framework’s ability to capture both broad tendencies and subtle nuances within the data, leading to more robust and precise forecasts.

The DualResolutionForecast stage within Nexus implements a dual-resolution methodology to address time series forecasting challenges by producing both macro-level and micro-level outlooks. This involves analyzing the time series at differing scales; the macro outlook captures broad, overarching trends, while the micro outlook focuses on granular, short-term fluctuations and specific feature patterns. By generating forecasts at these two resolutions in parallel, the framework aims to improve predictive accuracy and robustness, especially in complex time series where both large-scale trends and fine-grained details contribute to future values. This approach allows for the disentanglement of underlying patterns that might be obscured when analyzing the data at a single resolution.

The Nexus framework employs two distinct agent types, the MacroReasoningAgent and the MicroReasoningAgent, to address time series forecasting at varying levels of granularity. The MacroReasoningAgent focuses on high-level, coarse-grained trends within the time series data, providing a broad contextual outlook. Simultaneously, the MicroReasoningAgent analyzes the time series at a finer, more detailed resolution, capturing localized patterns and anomalies. These agents operate independently and in parallel, with their respective forecasts then combined to generate a comprehensive and nuanced prediction that benefits from both macro and micro perspectives. This parallel processing enhances the framework’s ability to disentangle complex temporal dynamics.

Nexus outperforms both TimesFM-2.5 and Chain-of-Thought baselines in generating qualitative forecasts, as demonstrated by visual comparison of their predictions.
Nexus outperforms both TimesFM-2.5 and Chain-of-Thought baselines in generating qualitative forecasts, as demonstrated by visual comparison of their predictions.

Dynamic Calibration: Learning from Predictive Discrepancies

The CalibrationAgent within Nexus functions as a dynamic error-correction mechanism, continuously analyzing discrepancies between predicted and actual values. This analysis informs iterative refinements to forecasting strategies, adjusting model parameters and algorithmic weights to minimize future errors. Beyond performance improvement, the CalibrationAgent also generates review guidelines, flagging specific instances of significant error or bias to facilitate human oversight and identify areas for further model development. The agent’s learning process is continuous, adapting to evolving data patterns and ensuring the system’s predictive capabilities remain optimized over time.

The CalibrationAgent within Nexus functions as a dynamic performance evaluator, continuously assessing the outputs of individual forecasting agents. This monitoring process includes the identification of systematic biases – consistent over or under-predictions – across various time series. Based on these evaluations, the agent adjusts the weighting assigned to each forecasting agent; agents demonstrating higher accuracy and lower bias receive increased influence in the overall forecasting process, while those with poorer performance are downweighted. This adaptive weighting mechanism allows Nexus to dynamically optimize its forecasting strategy and improve overall prediction accuracy without requiring manual intervention or retraining of individual agents.

Nexus incorporates Time Series Foundation Models (TSFM) to specifically address inherent volatility within time series data. These models identify and quantify fluctuations, enabling more accurate forecasting compared to baseline approaches. Performance evaluations using the Zillow and Stock datasets demonstrate consistent outperformance in both Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE). This indicates that the TSFM effectively models the dynamic characteristics of the data, leading to improved predictive accuracy and reduced error rates across diverse time series applications.

Beyond Extrapolation: Integrating Context and Long-Range Dependencies

Traditional time series forecasting often relies on historical data alone, overlooking the significant influence of external occurrences. Nexus distinguishes itself by integrating event-driven forecasting, a method that actively incorporates information about real-world events and their potential impact on future values. This approach moves beyond simply extrapolating past patterns; it recognizes that sales spikes can coincide with marketing campaigns, stock fluctuations can stem from geopolitical events, and energy demands can be shaped by weather patterns. By identifying and modeling these causal relationships, Nexus enhances forecast accuracy and provides a more contextualized understanding of the underlying dynamics driving the time series, ultimately leading to more informed decision-making.

Nexus employs Time Series Feature Mapping (TSFM) to discern connections within data that extend far beyond immediate temporal proximity. This innovative approach doesn’t simply analyze sequential patterns; instead, it actively seeks out and models long-range dependencies, revealing how events or conditions distant in time can subtly influence current outcomes. By identifying these often-overlooked relationships, TSFM allows Nexus to capture complex dynamics and anticipate shifts that traditional forecasting methods would miss. The system effectively builds a more holistic understanding of the time series, recognizing that present conditions aren’t solely determined by recent history, but also by echoes of the more distant past, ultimately leading to more robust and insightful predictions.

ForecastSynthesis represents a significant advancement in time series forecasting by intelligently integrating both macro-level trends and short-term catalytic events. This approach moves beyond simple extrapolation, constructing forecasts that are demonstrably more nuanced and comprehensive. Rigorous evaluation, utilizing a large language model as an independent judge, consistently reveals that ForecastSynthesis outperforms baseline models across multiple critical dimensions – including the relevance of the forecast to the broader domain, the accurate incorporation of impactful events, the logical consistency between analytical reasoning and numerical predictions, and an overall greater analytical depth. The result is a forecasting capability that doesn’t just predict what will happen, but provides a reasoned and contextualized understanding of why, offering a level of insight previously unattainable.

The Nexus framework, as detailed in the article, operates on a principle of modularity and rigorous interaction between agents-a design philosophy echoing the sentiments of Barbara Liskov, who once stated: “It’s one of the things I’ve always believed: that you can write programs that are understandable and correct.” This focus on correctness isn’t merely about achieving accurate forecasts; it extends to the framework’s ability to provide interpretable reasoning. The agents within Nexus aren’t simply black boxes generating predictions; their individual contributions and causal reasoning pathways are exposed, allowing for a verifiable understanding of the forecasting process. This commitment to provable logic, rather than empirical success alone, is central to the framework’s design and aligns perfectly with Liskov’s emphasis on program correctness and clarity.

What Remains to be Proven?

The presented Nexus framework, while demonstrating empirical success, merely shifts the burden of proof. The confluence of time series foundation models and large language models generates forecasts, yes, but does not inherently explain them with mathematical certainty. The current reliance on LLM-generated reasoning, however fluent, remains fundamentally a post-hoc justification, not a derivation. A truly elegant solution would embed causal axioms directly within the agentic structure, allowing for forecasts provably linked to underlying principles, rather than probabilistic associations.

Calibration, though addressed, remains a perpetually receding horizon. Improving confidence interval coverage is a statistical exercise, not a solution to the problem of epistemic uncertainty. The framework’s multimodal approach introduces further complexities; the true value lies not simply in incorporating diverse data streams, but in formally defining the information-theoretic relationships between them. Absent such definitions, the system remains a sophisticated correlator, not a true understanding engine.

Future work must focus on formalizing the very notion of ‘reasoning’ within this agentic context. The pursuit of interpretable AI should not settle for explanations humans find plausible, but for derivations that are logically unassailable. Only then will such frameworks transcend the limitations of purely empirical methods and approach a genuinely elegant solution to the problem of time series forecasting.

Original article: https://arxiv.org/pdf/2605.14389.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-05-17 15:24