Technical Summary: Nexus – An Agentic Framework for Time Series Forecasting

Problem Formulation

Time series forecasting has traditionally relied on specialized Time Series Foundation Models (TSFMs) that excel at identifying numerical patterns, seasonalities, and long-range dependencies within structured data. However, these models operate in a "multimodal vacuum," often failing to incorporate unstructured contextual signals (e.g., news, events, macroeconomic summaries) that drive structural breaks and regime shifts in real-world data. Conversely, while Large Language Models (LLMs) possess the reasoning capabilities to parse such textual context, they frequently lack the intrinsic autoregressive mechanisms required for precise numerical pattern recognition, leading to suboptimal performance when used as standalone zero-shot forecasters.

The core challenge addressed by this work is the "compromise" researchers face: either discard critical qualitative context to utilize statistical models or rely on LLMs that struggle to capture time series properties. The paper formulates the task as a multimodal problem requiring the joint prediction of future numerical values ($X_{\tau+1:\tau+T}$) and their underlying causal rationale ($R$), based on a historical context of both numerical sequences ($X_{1:\tau}$) and associated unstructured text ($E_{1:\tau}$).

Methodology: The Nexus Framework

To bridge the gap between statistical rigor and contextual reasoning, the authors introduce Nexus, a fully LLM-driven multi-agent framework. Rather than forcing a single monolithic model to handle all aspects of forecasting, Nexus decomposes the task into three distinct logical stages:

1. Contextualization

To prevent cognitive overload in long-context scenarios, a Historical Context Agent ($A_{ctx}$) processes the raw multimodal data. It transforms raw numerical values and dense text into a highly structured, chronological timeline ($H_{1:\tau}$). For each timestep, this agent filters noise and explicitly links numerical values to concise summaries of the key driving factors, ensuring downstream agents receive a high-fidelity signal of cause and effect.

2. Dual-Resolution Forecast Outlook Generation

Nexus generates two complementary perspectives from the structured history:

• Macro-Reasoning Agent ($A_{macro}$): Adopts a top-down approach to map out a broad trajectory for the entire forecast horizon. It establishes the expected regime and ensures the final forecast aligns with fundamental shifts.
• Micro-Reasoning Agent ($A_{micro}$): Adopts a granular, step-by-step approach. It evaluates immediate catalysts, short-term shifts, and localized volatility for every future timestep, ensuring responsiveness to specific events.

3. Forecast Synthesis and Calibration

• Forecast Synthesizer Agent ($A_{syn}$): This agent dynamically merges the macro and micro perspectives. It synthesizes the structured history with the dual outlooks, conditioned on learned guidelines ($G$), to produce the final numerical forecast and explicit reasoning.
• Calibration Agent ($A_{calib}$): To adapt to specific domains without manual instruction design, Nexus employs a forward-simulation backtesting mechanism. It analyzes prediction errors across historical splits to generate critique rules ($G_i$). These rules are intersected to form robust master guidelines ($G$), which are only applied if they yield a performance improvement on a hidden validation set, preventing overfitting.

Key Contributions

The paper makes three primary contributions:

1. Disentanglement of Reasoning: It demonstrates that effective LLM forecasting requires separating coarse-level trends from granular time series features to overcome the intrinsic numerical limitations of LLMs.
2. The Nexus Framework: It introduces a multi-agent architecture that models macro (coarse) and micro (granular) outlooks before dynamically synthesizing them into a single, robust forecast.
3. State-of-the-Art Performance: It shows that Nexus achieves superior results on both highly seasonal (Zillow) and volatile (Stocks) datasets, matching or outperforming dedicated TSFMs like TimesFM-2.5 even in purely numerical settings, while generating highly interpretable reasoning traces.

Experimental Results

The authors evaluated Nexus on data strictly succeeding the knowledge cutoffs (January 2025) of the underlying LLMs to ensure zero-shot evaluation without data leakage. The datasets included:

• Zillow Real Estate Metrics: Weekly sale inventory counts across 15 major US metropolitan areas.
• Stock Market Equities: Weekly closing prices for 7 diverse public companies (e.g., AAPL, GOOGL, RKLB).

Performance Metrics:

• Multimodal Context: Nexus consistently outperformed strong Chain-of-Thought (CoT) baselines. On the Zillow dataset using Claude-4.5-Sonnet, Nexus reduced MAPE by 86.6% and RMSE by 88.6% compared to the CoT baseline.
• Numerical Context Only: Even without textual context, Nexus matched or outperformed TimesFM-2.5. For instance, on the Zillow dataset with Gemini-3.1-Pro, Nexus achieved a 10.4% improvement in MAPE over the CoT baseline and matched the performance of the specialized TSFM.
• Reasoning Quality: In pairwise evaluations using cross-judge methodologies (Gemini judging Claude and vice versa), Nexus was preferred over the CoT baseline in 63.5% to 97.1% of cases across various criteria, including domain relevance, event plausibility, and logic-to-number consistency.

Component Analysis:
Ablation studies confirmed that all components are critical. Removing the Micro-Reasoning agent increased MAPE on stock data, indicating the necessity of granular analysis for volatility. Similarly, removing the Macro-Reasoning agent degraded performance, highlighting the need for overarching trend guidance.

Significance and Claims

The paper claims that real-world forecasting is fundamentally an agentic reasoning problem that extends beyond simple sequence modeling. By decomposing the prediction process, Nexus demonstrates that current-generation LLMs possess stronger intrinsic forecasting abilities than previously recognized, provided that numerical and contextual reasoning are organized effectively.

The authors emphasize that Nexus offers a promising approach for developing interpretable, robust, and highly adaptable forecasting systems. Unlike static models that produce forecasts without revision, Nexus produces high-quality reasoning traces that explicitly show the fundamental drivers behind each forecast, bridging the gap between numerical trends and qualitative context. The work suggests that the future of forecasting lies not in monolithic models, but in multi-agent systems that can synthesize statistical properties with fundamental drivers.