PaperlayerView exact revision on arXiv

Source-linked AI summary

Nexus : An Agentic Framework for Time Series Forecasting

Sarkar Snigdha Sarathi Das, Palash Goyal, Mihir Parmar, Nanyun Peng, Vishy Tirumalashetty, Chun-Liang Li, Rui Zhang, Jinsung Yoon, Tomas Pfister

arXiv:2605.14389v1cs.AIcs.CLcs.LG

TL;DR

Real-world time-series forecasting requires combining numerical patterns with unstructured contextual signals, which TSFMs and LLMs handle incompletely in isolation. Nexus decomposes forecasting into macro- and micro-level outlooks, contextualization, and synthesis, and consistently matches or outperforms TSFMs and LLM baselines on Zillow and stock-market data.

  • Problem

    TSFMs largely miss textual signals while LLMs lack precise numerical pattern-recognition mechanisms for context-dependent forecasting.

  • Method

    Nexus uses a multi-agent pipeline that contextualizes inputs, generates coarse and granular forecasts, and synthesizes them with domain-level calibration.

  • Results

    Nexus consistently matches or outperforms dedicated TSFMs and LLM baselines across seasonal Zillow data and volatile stock-market forecasting while producing interpretable reasoning.

  • Takeaways & Limitations

    The findings support treating real-world forecasting as an agentic reasoning problem that extends beyond sequence modeling.

  • Takeaways & Limitations

    The evaluation excludes pre-training-cutoff historical data to control for leakage from memorized values or events.

Abstract

from arXiv · show

Time series forecasting is not just numerical extrapolation, but often requires reasoning with unstructured contextual data such as news or events. While specialized Time Series Foundation Models (TSFMs) excel at forecasting based on numerical patterns, they remain unaware to real-world textual signals. Conversely, while LLMs are emerging as zero-shot forecasters, their performance remains uneven across domains and contextual grounding. To bridge this gap, we introduce Nexus, a multi-agent forecasting framework that decomposes prediction into specialized stages: isolating macro-level and micro-level temporal fluctuations, and integrating contextual information when available before synthesizing a final forecast. This decomposition enables Nexus to adapt from seasonal signals to volatile, event-driven information without relying on external statistical anchors or monolithic prompting. We show that current-generation LLMs possess substantially stronger intrinsic forecasting ability than previously recognized, depending critically on how numerical and contextual reasoning are organized. Evaluated on data strictly succeeding LLM knowledge cutoffs spanning Zillow real estate metrics and volatile stock market equities, Nexus consistently matches or outperforms state-of-the-art TSFMs and strong LLM baselines. Beyond numerical accuracy, Nexus produces high-quality reasoning traces that explicitly show the fundamental drivers behind each forecast. Our results establish that real-world forecasting is an agentic reasoning problem extending well beyond only sequence modeling.

1. Introduction

The introduction frames forecasting as a problem requiring both numerical temporal modeling and qualitative context, then presents Nexus as a multi-agent solution that separates macro and micro reasoning before synthesis. Evaluated beyond LLM knowledge cutoffs on volatile stocks and seasonal Zillow data, Nexus matches or outperforms TSFMs and LLM baselines while producing interpretable reasoning.

  • Motivation: Time Series Foundation Models unify forecasting but remain vulnerable to structural breaks and regime shifts because they rely primarily on structured numerical sequences.Real-world series often reflect qualitative events and unstructured textual signals that historical data alone cannot capture.
  • Motivation: LLM-only numerical forecasting can underperform TSFMs because autoregressive continuous-value prediction lacks an intrinsic mechanism for temporal dependencies.This creates a tradeoff between statistical models that discard qualitative context and zero-shot LLM reasoning that may miss time-series properties.
  • Nexus: Nexus separates coarse-level trend modeling from granular feature and catalyst analysis, then uses a synthesizer agent to produce a mathematically grounded forecast.The framework is fully LLM-driven and avoids forcing one model to handle all forecasting requirements simultaneously.
  • Evaluation: Nexus is evaluated on post-cutoff data spanning 7 volatile stock tickers and Zillow Home Counts across 15 major US metropolitan areas.The evaluation is designed to prevent knowledge leakage and covers both text-driven volatile-market forecasting and intrinsic numerical modeling of periodic real-estate data.
  • Results: Nexus consistently outperforms TimesFM-2.5 and Zero-Shot CoT baselines, achieving superior numerical accuracy and highly interpretable reasoning across both domains.The introduction characterizes these results as state-of-the-art performance on seasonal Zillow and volatile stock datasets.

2. Problem Formulation

Nexus formulates multimodal time series forecasting as jointly predicting future numerical values and generating explicit causal rationales from numerical histories and associated unstructured text. The task learns a mapping that synthesizes quantitative data and qualitative context to produce forecasts with justifications.

  • Problem Formulation: The task jointly predicts future sequence values and generates their underlying causal rationale from multimodal historical context.The context combines numerical observations with associated unstructured textual data, including news, financial reports, and macroeconomic summaries.
  • Problem Formulation: The numerical input X1:𝜏 is a univariate sequence observed over a context window of length 𝜏, paired with textual data E1:𝜏 at each timestep.The textual context may include news, financial reports, or macroeconomic summaries.
  • Problem Formulation: The output comprises forecasts X𝜏+1:𝜏+𝑇 for the subsequent 𝑇 timesteps and natural-language reasoning R𝜏+1:𝜏+𝑇 explaining their fundamental catalysts and events.The framework learns a mapping F that synthesizes quantitative data and qualitative context to output predicted values and justifications.

3. The Nexus Framework

Nexus decomposes forecasting into contextualization, dual-resolution outlook generation, and forecast synthesis and calibration. Its agents structure multimodal history, model macro and micro dynamics, then combine forecasts with calibrated guidelines to produce numerical predictions and interpretable reasoning.

  • Framework Overview: Nexus organizes forecasting into three stages: Contextualization, Dual-Resolution Forecast Outlook Generation, and Forecast Synthesis and Calibration.The framework replaces direct monolithic approximation with staged processing of multimodal forecasting inputs.
  • Forecast Synthesis and Calibration: The Forecast Synthesizer Agent dynamically merges macro trajectories, micro catalysts, and structured history to generate the final forecast and reasoning.Synthesis is conditioned on learned guidelines G from calibration.
  • Contextualization: The Historical Context Agent transforms numerical values and external textual information into a structured chronological timeline emphasizing factors driving value changes.This preprocessing reduces cognitive overload from long numerical sequences mixed with dense unstructured text.
  • Dual-Resolution Forecast Outlook Generation: The Macro-Reasoning Agent establishes a broad forecast-horizon regime, while the Micro-Reasoning Agent evaluates each future timestep for immediate catalysts and localized volatility.The complementary outlooks preserve both broader fundamental shifts and short-term events.
  • Forecast Synthesis and Calibration: The Calibration Agent uses sequential forward-simulation backtesting to derive generalized critique guidelines and applies them only after hidden-validation improvement of at least k%.Rules from n−1 training folds are intersected to reduce overfitting to temporary anomalies.

4. Experiments

Nexus is evaluated in leakage-controlled, zero-shot experiments on post–January 2025 Zillow and stock datasets, using multimodal and numerical-only settings across multiple horizons. It consistently outperforms CoT baselines and matches or exceeds TimesFM-2.5, while ablations show that macro, micro, and calibration components are important for accuracy.

  • Experimental Setup: The evaluation uses data immediately following the models’ January 2025 knowledge cutoff to prevent leakage and test genuine zero-shot forecasting.Historical pre-cutoff data could allow models to recall numerical values or real-world events from pretraining.
  • Datasets: Experiments cover Zillow weekly sale inventory across 15 US MSAs and weekly closing prices for seven publicly traded companies, with three years of numerical history provided for each task.The Zillow evaluation spans February–October 2025; the stocks include AAPL, GOOGL, RKLB, JNJ, MSFT, NFLX, and NVDA.
  • Evaluation Settings & Horizons: LLMs are evaluated with numerical context only or with numerical sequences plus chronological unstructured text, across short-, medium-, and long-horizon forecasts.Horizons are 4, 8, and 13 weeks for Zillow and 6, 13, and 26 weeks for Stocks; Nexus uses n=6 backtest splits and a 5% calibration threshold.
  • Multimodal Contextual Forecasting: Nexus consistently outperforms the LLM-based CoT baseline in multimodal forecasting, with the largest performance gap on Zillow.Claude-4.5-Sonnet’s CoT baseline over-relies on simple trend extrapolation, whereas Nexus tracks nuanced temporal dynamics and contextual events across both domains.
  • Numerical-only Forecasting: In numerical-only forecasting, Nexus consistently matches or outperforms TimesFM-2.5 across Zillow and Stocks, indicating strong intrinsic time-series pattern recognition.The comparison includes both the CoT baseline and TimesFM-2.5, a flagship time-series foundation model.
  • Ablation Study: On Stocks, removing Micro Reasoning increases MAPE from 0.0866 to 0.0877, while removing Macro Reasoning increases it to 0.0882; the full pipeline outperforms ablations and standard CoT.These results identify macro, micro, and calibration as critical components of Nexus’s forecasting accuracy.

5. Related Works

Related work adapts LLMs to time-series forecasting through prompting, numerical tokenization, modality alignment, and cross-modal bridging, while time-series foundation models discretize observations for probabilistic zero-shot prediction. Recent semantic, adaptive, and agentic approaches add calibration, fusion, arbitration, planning, tool use, reflection, and memory, increasingly framing forecasting as an iterative workflow.

  • LLMs for Time Series Forecasting: LLM-based forecasting research explores direct prompting, numerical tokenization, modality alignment, and cross-modal bridging for adapting language models to continuous time series.These directions investigate how models originally trained on discrete text can process numerical observations for forecasting.
  • Time-Series Foundation Models: Chronos discretizes and quantizes continuous time-series values into a fixed vocabulary, then trains transformer language-model architectures with cross-entropy loss for probabilistic zero-shot forecasting.The approach is evaluated across diverse datasets and treats forecasting explicitly as language modeling.
  • Semantic, Adaptive, and Agentic Forecasting: Semantic, adaptive, and agentic methods extend forecasting beyond sequence modeling through semantic calibration, temporal distillation, asynchronous text-time-series fusion, and arbitration among models.The cited systems include LoFT-LLM, T-LLM, TimeSAF, and Synapse.
  • Semantic, Adaptive, and Agentic Forecasting: Recent agentic systems introduce multi-step planning, tool use, reflection, or memory, although most primarily operate over numerical histories, diagnostics, model outputs, or tool-generated features.AlphaCast and Cast-R1 are cited as examples of these developments.
  • Semantic, Adaptive, and Agentic Forecasting: Agentic time-series forecasting reframes prediction as an iterative workflow involving perception, planning, and reflection rather than static model-centric prediction.This direction emphasizes workflow-level forecasting beyond a single static model output.

6. Conclusion

Nexus is a multi-agent framework for multimodal contextual time-series forecasting that manages numerical sequences and unstructured text through structured forecasting stages.

  • Framework overview: Nexus addresses multimodal contextual time-series forecasting with a novel multi-agent framework.The framework is designed for forecasting that combines numerical data with unstructured text.
  • Framework overview: The framework decomposes forecasting into Contextualization, Dual-Resolution Forecast Outlook Generation, and Forecast Synthesis and Calibration.These stages structure the processing of long numerical sequences alongside unstructured text.
  • Framework overview: Nexus manages the complexity of combining long numerical sequences with unstructured text through structured processing stages.

A. Limitations … D. LLM-as-a-Judge Prompt

The appendix notes that evaluation is limited by dataset availability and potential training-data contamination, then specifies prompts for Nexus agents, a CoT baseline, and an LLM-based reasoning judge. These prompts operationalize historical event analysis, macro/micro forecasting, calibration, final synthesis, and qualitative comparison using structured outputs.

  • A. Limitations: Evaluation is limited to Zillow and Stock datasets because paired, timestamped numerical values and related textual context are scarce, while publicly available data may overlap with LLM training.The paper identifies dataset scarcity and contamination risk as current limitations.
  • B.1. Historical Context Agent: The Historical Context Agent extracts explicit and implicit factors, links events to historical target changes, and outputs a chronological timestamped timeline.Its prompt requires values, textual content, and coverage of every listed fact or event.
  • B.2. Macro-Reasoning Forecaster Agent: The Macro-Reasoning Forecaster predicts exactly the requested horizon from historical context, placing step-by-step reasoning and numerical forecasts in separate tags.The prompt states a January 2025 knowledge cutoff and requires exactly {horizon} values.
  • B.3. Micro-Reasoning Forecaster Agent: The Micro-Reasoning Forecaster predicts future events and target values at the specified frequency using a JSON timestamp_forecasts list with movement labels, drivers, reasoning, and adjusted values.The required schema generates forecasts for every horizon timestamp.
  • B.4. Calibration Agent: The Calibration Agent compares predictions with ground truth, diagnoses numerical and logical errors, and produces generalized review guidelines for future sanity checks.It considers agent error, MAPE, upstream macro and micro MAPE, actual events, and actual values.
  • B.5. Value Predictor Agent: The Value Predictor combines historical data with macro and micro reasoning, adjusts both perspectives for each future step, and outputs exactly {horizon} final values.Its prompt uses upstream forecasts and calibration guidelines as inputs.
  • C. CoT-Baseline Prompt: The CoT baseline relies on deductive reasoning from historical records and event intelligence, considers seasonality, and returns exactly {horizon} values inside <prediction> tags.The prompt specifies comma-separated numerical output and a January 2025 knowledge cutoff.
  • D. LLM-as-a-Judge Prompt: The LLM-as-a-Judge compares two models without ground-truth numerical values, judging domain relevance, event plausibility, logic-to-number consistency, analytical depth, and overall preference in strict JSON.It evaluates reasoning quality, coherence, and plausibility rather than requiring perfect numerical accuracy.

E. Qualitative Forecast Examples · F. Qualitative Reasoning Examples · F.1. Example 1

The qualitative examples compare Nexus forecasts with TimesFM-2.5 and CoT baselines across Zillow metropolitan-area series. A reasoning example shows Nexus combining macro tariff-shock recovery expectations with Microsoft-specific AI and cloud-growth catalysts to project a steady 26-week rise despite minor seasonal pullbacks.

  • E. Qualitative Forecast Examples: Figure 3 compares Nexus predictions with TimesFM-2.5 and CoT baselines.The figure presents qualitative forecast comparisons.
  • E. Qualitative Forecast Examples: The examples include San_Diego_CA_msa forecasts at horizons h8.The labeled example is identified as panel (e).
  • E. Qualitative Forecast Examples: The examples include Los_Angeles_CA_msa forecasts at horizons h13, h8, and h4.These are labeled as panels (f), (g), and (h).
  • E. Qualitative Forecast Examples: Additional examples cover Houston_TX_msa_TX at h8, Riverside_CA_msa_CA at h4, and Washington_DC_msa_VA at h13 and h8.These correspond to panels (k), (l), (o), and (p).
  • F. Qualitative Reasoning Examples: The reasoning examples illustrate Nexus integrating macro-level and micro-level drivers before producing a forecast.The example combines tariff-related recovery expectations with company-specific catalysts.
  • F.1. Example 1: Over the 26-week horizon, Microsoft’s projected price steadily climbs as tariff impacts are absorbed and strong earnings are anticipated.The trajectory is attributed to enterprise software resilience and ongoing AI advancement.
  • F.1. Example 1: Minor seasonal pullbacks do not disrupt the upward trajectory driven by AI monetization, cloud infrastructure growth, and enterprise software resilience.The forecast attributes the final values to continued AI advancement alongside the broader recovery expectation.

F.2. Example 2

The forecast projects continued stock growth, driven by strong iPhone 17 sales, the M5 hardware launch, upcoming Q4 earnings, and a potential Federal Reserve rate cut. Robust holiday consumer demand and Black Friday momentum are expected to outweigh temporary government-shutdown uncertainty.

  • Growth drivers: Strong initial iPhone 17 sales and the M5 hardware launch provide a foundation for continued stock growth.These product signals are identified as the forecast’s primary basis for continued growth.
  • Catalysts: Upcoming Q4 earnings and a potential Federal Reserve rate cut are expected to act as positive catalysts.The passage presents both events as supportive catalysts for the stock.
  • Seasonal outlook: Holiday consumer demand and Black Friday momentum are anticipated to drive the stock higher through late November despite temporary government-shutdown uncertainty.The forecast expects seasonal demand to outweigh the temporary macroeconomic disruption.

F.3. Example 3 · F.4. Example 4

The examples illustrate Nexus’s contextual forecasting by linking company-specific catalysts, fundamentals, and macroeconomic conditions to gradual upward price-recovery trajectories over distinct horizons.

  • F.3. Example 3: Rocket Lab’s inclusion in the $5.6 billion NSSL Phase 3 Lane 1 contract is identified as a major positive catalyst validating Neutron.The catalyst counters previous short-seller claims.
  • F.3. Example 3: Over the next six weeks, Rocket Lab is forecast to show steady upward momentum as the market digests the contract and launch cadence.Anticipation of the upcoming Q1 2025 earnings report also supports the forecast.
  • F.3. Example 3: Rocket Lab’s forecast describes a gradual recovery from its recent price correction.The trajectory is associated with contract effects, strong Q1 launch cadence, and earnings anticipation.
  • F.4. Example 4: Netflix’s forecast combines macro and micro perspectives with strong underlying fundamentals, including robust ad-tier growth and strategic live sports investments.The forecast accounts for recent macroeconomic volatility.
  • F.4. Example 4: Over the next 26 weeks, Netflix is expected to recover from tariff-induced dips and follow a steady upward trajectory.The forecast specifies a longer horizon than the Rocket Lab example.
  • F.4. Example 4: Netflix’s projected growth is tied to strong Q1 and Q2 earnings reports, successful summer content slates, and expanding ad-tech capabilities.These anticipated developments serve as catalysts for the recovery trajectory.

F.5. Example 5 … F.9. Example 9

Examples 5–9 show Nexus combining macro-level seasonal or recovery patterns with micro-level events and domain-specific fundamentals to produce contextually grounded forecasts. The examples span travel activity, regional events and wildfires, and a stock’s bullish recovery.

  • F.5. Example 5: Nexus combines historical spring growth with San Diego-specific events to model a steady week-over-week rise from a lower 2025 baseline.The cited drivers include spring break travel, cultural festivals, and the start of baseball season.
  • F.5. Example 5: The final San Diego projections align with the micro-level reasoning and reflect anticipated seasonal momentum.
  • F.6. Example 6: The forecast models a variable that bottoms in January and February before steadily climbing through spring, incorporating recovery from winter and wildfire-induced troughs.
  • F.6. Example 6: Upcoming events including the Super Bowl, Academy Awards, LA Marathon, and Coachella are integrated with the macro seasonal outlook.
  • F.7. Example 7: The eight-week forecast anticipates recovery from January 2025 wildfires, early-February stabilization, and a pronounced March rise driven by seasonal events, milder weather, and spring break travel.
  • F.8. Example 8: Historical trends and wildfire impacts support initial stabilization after January, followed by stronger growth from Grammy Awards, Super Bowl festivities, holiday weekends, and resumed local activities.
  • F.9. Example 9: The stock’s V-shaped recovery is supported by Neutron Stage 2 qualification and inclusion in NSSL and EWAAC, with bullish momentum expected to continue.
  • F.9. Example 9: Anticipated strong Q1 earnings and the H2 2025 window for the Neutron rocket’s debut contribute to a projected steady climb over the next six weeks.

F.10. Example 10 … F.16. Example 16

Examples 10–16 apply Nexus’s macro–micro decomposition across stock-market and seasonal activity forecasts. The forecasts combine broad trends with event-specific catalysts, producing directional projections over horizons from four to thirteen weeks.

  • F.10. Example 10: F.10 forecasts Alphabet’s next 13 weeks by combining a sustained bullish macro trend with events, earnings, economic releases, and a recent close of 212.58.Key forward drivers include Federal Reserve policy, Q3 earnings, and antitrust developments.
  • F.11. Example 11: F.11 projects a gradual eight-week rise toward a late-period peak from seasonal spring momentum despite a slightly lower 2025 baseline.The seasonal pattern is linked to warmer weather, cultural festivals, Cinco de Mayo, and Memorial Day.
  • F.12. Example 12: F.12 synthesizes late-February and March seasonal growth with spring events and improving weather to project steady weekly increases over four weeks.The upward trajectory remains intact despite a year-over-year baseline decline since 2022.
  • F.13. Example 13: F.13 characterizes Rocket Lab’s six-week forecast as a sustained macro uptrend reinforced by launch execution, major contracts, and the anticipated Neutron debut.The stock rose from the $11 range to above $64 over the past year.
  • F.14. Example 14: F.14 forecasts Rocket Lab’s near-term rise, brief consolidation, and subsequent recovery by combining backlog and mission execution with earnings and Spring 2025 catalysts.Q4 2024 earnings and analyst reactions drive initial momentum before later catalysts are priced in.
  • F.16. Example 16: F.16 projects consistent week-over-week growth after a late-January trough, using the early-2025 baseline increase and catalysts including Presidents’ Day and the National Cherry Blossom Festival.The forecast applies the historical February–March growth curve to Washington, DC activity.
Loading 2605.14389v1…