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

Forecasting from numerical history alone misses textual signals, while LLMs lack precise numerical pattern mechanisms. Nexus separates coarse and granular outlooks before synthesis, and matches or outperforms TSFMs across Zillow and stock forecasting.

  • Problem

    TSFMs rely on numerical sequences and miss unstructured real-world signals, while LLMs can reason over such context but lack mechanisms for precise numerical pattern recognition.

    Source: §1 · p. 1

  • Method

    Nexus generates separate coarse-level trend and granular-level feature-and-catalyst outlooks, then synthesizes them into a forecast.

    Source: §1 · p. 2

  • Results

    On seasonal Zillow and volatile Stocks datasets, Nexus achieves state-of-the-art results, matching or outperforming TSFMs and outperforming zero-shot CoT baselines.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The findings frame real-world forecasting as an agentic reasoning problem extending beyond sequence modeling.

    Source: §6 · p. 11

    Pre-cutoff historical evaluations can inflate scores because models may recall numerical values or related events from pretraining.

    Source: §4.1 · p. 5

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

Nexus frames forecasting as joint temporal and contextual reasoning, separating coarse trends from granular features and catalysts before synthesizing and calibrating forecasts. The authors report state-of-the-art results across volatile stock and seasonal Zillow data.

Introduction

  • TSFMs capture numerical patterns but can miss contextual shocks, while LLMs parse unstructured context yet struggle with precise temporal dependencies and numerical forecasting.This leaves a compromise between discarding qualitative context and relying on zero-shot LLM forecasts that may fail to capture time-series properties.Source: §1 · pp. 1–2
  • Nexus treats agentic forecasting as integrating textual evidence with temporal reasoning, rather than merely automating numerical workflows.The paper argues volatile-domain forecasts require statistical properties and fundamental drivers because numerical models can miss shock events while LLMs struggle with multi-seasonal periodicity.Source: §1 · p. 2
  • Nexus separates coarse-level trends from granular time-series features and catalysts, then synthesizes the outlooks into a mathematically grounded forecast.A domain-level calibration loop evaluates past errors across historical splits and generates review guidelines to help the synthesizer weigh conflicting signals.Source: §1 · pp. 2–3
  • Nexus consistently outperforms TimesFM-2.5 [Das et al., 2024] and Zero-Shot CoT baselines [Kojima et al., 2022] across stock and Zillow forecasting, with superior numerical accuracy and interpretable reasoning.Evaluation uses data strictly after the underlying LLMs’ knowledge cutoffs, covering seven stock tickers and Zillow Home Counts for 15 major US metropolitan areas.Source: §1 · pp. 2–3

2. Problem Formulation

The task is to forecast future numerical values from historical numerical and textual context while generating natural-language reasoning about the catalysts and events behind those predictions. It is framed as a mapping from multimodal context to forecasts and their justifications.

  • Multimodal time series forecasting with explicit reasoning jointly predicts future values and generates their underlying causal rationale.Source: §2 · p. 3
  • The historical context pairs a univariate numerical sequence with timestep-associated unstructured text, such as news, financial reports, or macroeconomic summaries.Source: §2 · p. 3
  • Given that context, the task predicts the next T numerical values and generates natural-language reasoning about the fundamental catalysts and events driving them.Source: §2 · p. 3
  • Formally, a mapping F synthesizes quantitative data and qualitative context to output predicted values together with their justifications.Source: §2 · p. 3

3. The Nexus Framework

Nexus replaces monolithic forecasting with three stages: contextualizing historical multimodal data, generating macro- and micro-resolution outlooks, and synthesizing a forecast with explicit reasoning. Calibration can refine this process using backtest-derived guidelines that pass a validation threshold.

3. The Nexus Framework

  • Nexus decomposes forecasting into Contextualization, Dual-Resolution Forecast Outlook Generation, and Forecast Synthesis and Calibration, combining structured context and distinct outlooks into a forecast with interpretable reasoning.The framework structures raw multimodal context, projects future outlooks at different resolutions, and merges them into numerical predictions and explicit reasoning.Source: §3 · p. 3

3.1. Contextualization

  • The Historical Context Agent organizes each timestep’s numerical value and textual context into a chronological timeline of key drivers, filtering noise for downstream forecasters.This structured, step-by-step representation links values with concise summaries of their driving factors rather than producing a generic monolithic summary.Source: §3.1 · pp. 3–4

3.2. Dual-Resolution Forecast Outlook Generation

  • Nexus generates complementary macro and micro outlooks: the macro agent sets a horizon-wide regime, while the micro agent evaluates timestep-specific catalysts and volatility.The two resolutions address the risk of missing short-term details when focusing only on broad trends, or losing broader shifts when considering changes step by step.Source: §3.2 · pp. 4–5

3.3. Forecast Synthesis and Calibration

  • The Forecast Synthesizer merges structured history with macro and micro outlooks to produce each numerical prediction and reasoning that explains how it weighted the two views.Calibration generates critique guidelines from errors and reasoning across training folds, intersects fold-specific rules, and applies the resulting guidelines only if they improve performance by at least k% on the hidden validation fold.Source: §3.3 · p. 5

4. Experiments

Experiments evaluate Nexus on post-cutoff multimodal and numerical-only forecasting, and find it outperforms the CoT baseline, matches or outperforms a TSFM, and produces reasoning preferred by judges.

4.1. Experimental Setup

  • The zero-shot evaluation used datasets from after the models’ January 2025 knowledge cutoff to reduce the risk of forecasts benefiting from recalled values or events.It covered Zillow weekly inventory and seven equities, with three years of numerical history for Zillow and one year for Stocks. Gemini-3.1-Pro and Claude-4.5-Sonnet were evaluated at temperature 0.1 across 4-, 8-, and 13-week Zillow horizons and 6-, 13-, and 26-week stock horizons.Source: §4.1 · pp. 6–7

4.2. Forecasting with Multimodal Context

  • Nexus consistently outperformed the LLM CoT baseline in multimodal forecasting, with the performance gap especially pronounced on Zillow.With Claude-4.5-Sonnet, CoT showed substantial degradation that the authors attribute possibly to long-context difficulty and reliance on simple trend extrapolation; Nexus tracked temporal dynamics and contextual events across both domains.Source: Table 2 · p. 7, §4.2 · pp. 7–8

4.3. Forecasting with Numerical Context Only

  • Nexus consistently matched or outperformed TimesFM-2.5 across Zillow and Stocks when given numerical context only.The comparison also included the CoT baseline, and the authors conclude that Nexus captures time-series dynamics beyond its contextual reasoning capability.Source: Table 3 · p. 8, §4.3 · p. 8

4.4. Reasoning Quality Evaluation

  • In cross-judged pairwise evaluation, Nexus was preferred by judge LLMs most of the time and showed substantially better reasoning than CoT.Judges assessed domain relevance, event relevance and plausibility, logic-to-number consistency, and analytical depth; Gemini outputs were judged by Claude and vice versa.Source: §4.4 · pp. 8–9, Table 4 · p. 9

4.5. Component Analysis

  • On Stocks, disabling Micro Reasoning raised MAPE from 0.0866 to 0.0877, while disabling Macro Reasoning raised it to 0.0882; full Nexus beat ablations and standard CoT.The authors link micro reasoning to capturing short-term volatility and macro reasoning to overarching trend guidance.Source: Table 5 · p. 9, §4.5 · pp. 9–10

5. Related Works

The related-work landscape spans language-model adaptation and temporal pretraining, then extends toward semantic, adaptive, and agentic forecasting. Nexus is positioned in this latter direction through explicit LLM reasoning, diagnostic feedback, and iterative calibration over temporal evidence.

LLMs for Time Series Forecasting

  • LLM forecasting ranges from string-based zero-shot prompting to modality-aligned or reprogrammed models, while ablations question whether linguistic pretraining drives gains.Examples include Zhang et al. [2024] and Gruver et al. [2023] on string encoding, GPT4TS/FPT Zhou et al. [2023], TEMPO Cao et al. [2024], Time-LLM Jin et al. [2024b], UniTime Liu et al. [2024b], and Tan et al. [2024]'s ablations.Source: §5 · p. 10

Time-Series Foundation Models

  • Temporal foundation models use large-scale temporal pretraining for zero-shot forecasting but generally remain static, single-pass predictors without explicit reasoning or external-evidence interaction.Examples include Chronos Ansari et al. [2024], TimesFM Das et al. [2024], Lag-Llama Rasul et al. [2023], MOMENT Goswami et al. [2024], and MOIRAI Woo et al. [2024].Source: §5 · p. 10

Semantic, Adaptive, and Agentic Forecasting

  • Recent semantic forecasting work studies calibration, temporal distillation, and asynchronous text-time-series fusion.These directions are explored by LoFT-LLM You et al. [2025], T-LLM Guo et al. [2026], and TimeSAF Zhang et al. [2026].Source: §5 · p. 10
  • Adaptive and agentic systems add model arbitration, feature analysis, selection, planning, tool use, reflection, or memory, but mostly operate over numerical histories or derived outputs.Examples include Synapse Das et al. [2025], TimeCopilot Garza and Rosillo [2025], TimeSeriesScientist Zhao et al. [2025], AlphaCast Zhang et al. [2025], and Cast-R1 Tao et al. [2026].Source: §5 · pp. 10–11
  • Nexus is positioned within agentic forecasting through explicit LLM reasoning, diagnostic feedback, and iterative calibration over prior temporal evidence rather than one-shot numerical extrapolation.Agentic time series forecasting Cheng et al. [2026] argues for iterative workflows involving perception, planning, reflection, and memory.Source: §5 · p. 11

6. Conclusion

Nexus decomposes multimodal contextual time-series forecasting into specialized stages that combine macro-level trajectories with event-driven micro-level catalysts. In zero-shot evaluations on Zillow and stock-market datasets, it consistently outperformed TimesFM-2.5 and strong Chain-of-Thought LLM baselines while producing explicit reasoning.

6. Conclusion

  • Nexus consistently outperformed TimesFM-2.5 and strong Chain-of-Thought LLM baselines in zero-shot evaluations on Zillow real-estate and stock-market datasets.The evaluations covered real-world datasets.Source: §6 · p. 11
  • Nexus organizes forecasting into Contextualization, Dual-Resolution Forecast Outlook Generation, and Forecast Synthesis and Calibration.The stages manage long numerical sequences alongside unstructured text.Source: §6 · p. 11
  • Nexus synthesizes broad macro-level trajectories with granular, event-driven micro-level catalysts to produce numerical predictions alongside explicit reasoning.The framework is presented as an approach for interpretable, robust, and adaptable forecasting in complex real-world domains.Source: §6 · p. 11

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

The appendix describes the study’s evaluation limits and documents the prompts used for Nexus agents, the CoT baseline, and the LLM-as-a-Judge comparator.

A. Limitations

  • The evaluation covers only Zillow and Stock data, uses post-January 2025 observations to mitigate leakage, and reports single-run results because repeated runs are computationally infeasible.The authors cite the scarcity of paired timestamped numerical and textual datasets and the high cost of querying very large models as constraints.Source: A. Limitations · p. 15

B.1. Historical Context Agent; B.2. Macro-Reasoning Forecaster Agent

  • The appendix supplies agent prompts and templates; the historical context agent builds a timestamped event-impact timeline, while the macro forecaster reasons over context to predict exactly the requested horizon.The historical agent is instructed to capture all listed facts for each timestamp, and the macro agent returns reasoning and forecast values in specified tags.Source: B. Nexus : Agent Prompts · p. 15, §B.1 · p. 15, §B.2 · p. 16

B.3. Micro-Reasoning Forecaster Agent

  • The micro-reasoning agent predicts future events and values, returning a timestamped forecast with movement labels, key drivers, reasoning, and adjusted values for each step.Its required output is a single JSON object covering every forecast timestamp at the specified frequency.Source: §B.3 · pp. 16–17

B.4. Calibration Agent

  • The calibration agent critiques a value predictor’s reasoning and numerical estimates against ground truth, then produces generalized review guidelines for future predictions.Its prompt emphasizes whether the reasoning translated into numerical values appropriately, including over- or underestimating impacts.Source: §B.4 · p. 17

B.5. Value Predictor Agent

  • The value predictor combines historical context with macro and micro forecasts, adjusting those perspectives to produce final numerical values for each future step.The prompt also supplies calibration guidelines when available and requires reasoning and exactly the requested number of forecasts.Source: §B.5 · pp. 17–18

C. CoT-Baseline Prompt

  • The CoT baseline is prompted to forecast from historical records and event intelligence, consider seasonality, and provide brief reasoning with exactly the requested number of values.It is instructed to rely solely on its own deductive reasoning from the supplied context.Source: C. CoT-Baseline Prompt · pp. 18–19

D. LLM-as-a-Judge Prompt

  • The judge compares forecast reasoning rather than numerical accuracy because it is not given ground-truth values, assessing domain and event relevance, logic-to-number consistency, and analytical depth.It selects an overall preference based on coherence, depth, and plausibility, with ties allowed.Source: D. LLM-as-a-Judge Prompt · p. 19

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

Figure 3 presents qualitative forecasts comparing Nexus with TimesFM-2.5 and CoT, while Example 1 describes a 26-week upward forecast driven by macroeconomic and company-specific factors.

E. Qualitative Forecast Examples

F.1. Example 1

  • The 26-week forecast projects a steady climb despite minor seasonal pullbacks as the market absorbs trade policy impacts and anticipates strong earnings.The passage describes a gradual recovery from early 2025 tariff shocks and ongoing enterprise software resilience and AI advancements.Source: §F.1 · p. 26

F. Qualitative Reasoning Examples

  • The reasoning combines a gradual recovery from early 2025 tariff shocks with Microsoft AI monetization and cloud infrastructure growth.These macro-level expectations and micro-level catalysts jointly support the projected trajectory.Source: §F.1 · p. 26

F.2. Example 2

Example 2 attributes the forecasted stock growth to strong product sales and hardware launches, anticipated catalysts, and holiday-season demand outweighing temporary macroeconomic uncertainty.

F.2. Example 2

  • Strong initial iPhone 17 sales and the M5 hardware launch provide a foundation for continued growth.Source: §F.2 · p. 26
  • Upcoming Q4 earnings and a potential Federal Reserve rate cut are expected to outweigh temporary uncertainty from the government shutdown.Source: §F.2 · p. 26
  • Robust consumer demand and Black Friday momentum are anticipated to drive the stock steadily higher through late November.Source: §F.2 · p. 26

F.3. Example 3 · F.4. Example 4

The examples forecast upward trajectories for Rocket Lab and Netflix, grounding each in company-specific catalysts, earnings expectations, and macro context.

F.3. Example 3

  • Rocket Lab’s inclusion in the $5.6 billion NSSL Phase 3 Lane 1 contract is presented as a positive catalyst validating Neutron and countering short-seller claims.The forecast describes this as macro-level validation.Source: §F.3 · p. 26
  • Over the next six weeks, Rocket Lab is expected to gain steadily as the market digests the contract, strong Q1 launch cadence, and anticipated Q1 2025 earnings.Source: §F.3 · p. 26
  • Rocket Lab’s forecast describes a gradual recovery from its recent price correction, aligning contract validation with micro-level earnings momentum.Source: §F.3 · p. 26

F.4. Example 4

  • Netflix’s forecast cites robust ad-tier growth and strategic live sports investments as strong fundamentals despite recent macroeconomic volatility.Source: §F.4 · p. 26
  • Over the next 26 weeks, Netflix is expected to recover from tariff-induced dips and rise steadily, driven by anticipated Q1 and Q2 earnings, summer content, and expanding ad-tech.Source: §F.4 · p. 26
  • Netflix’s projected trajectory reaches toward the 115 range by October 2025.Source: §F.4 · p. 26

F.5. Example 5 … F.9. Example 9

Examples 5–9 show forecasts combining macro-level seasonal or market patterns with micro-level event and regional drivers. Across the examples, forecasts describe seasonal recovery and normalization, while the stock example projects a bullish recovery tied to company milestones and anticipated earnings.

F.5. Example 5

  • The San Diego forecast applies historical spring growth to a lower 2025 baseline, projecting a steady weekly climb driven by spring break, cultural festivals, and baseball season.The final values align with the micro-reasoning projections and combine the historical late-February-to-mid-April seasonal rise with local event drivers.Source: §F.5 · pp. 26–27

F.6. Example 6

  • The forecast stabilizes in early February after a winter and wildfire-induced trough, then rises toward a mid-April peak before slight post-holiday normalization.It aligns the historical January–February low and spring climb with macro-reasoning and micro-level events including the Super Bowl, Academy Awards, LA Marathon, and Coachella.Source: §F.6 · p. 27

F.7. Example 7

  • The eight-week forecast anticipates stabilization after January 2025 wildfire declines, followed by a pronounced March rise as seasonal events, milder weather, and spring-break travel return.It also accounts for a minor historical plateau in the second week of February and compares the recovery with prior years’ spring patterns.Source: §F.7 · p. 27

F.8. Example 8

  • The forecast combines the historical post-January recovery pattern with wildfire impacts, anticipating early-February stabilization and later normalization as cultural events and resumed activities support an upward trajectory.Historical trends from 2023 and 2024 inform the outlook, which names the Grammy Awards, Super Bowl festivities, and holiday weekends as drivers.Source: §F.8 · p. 27

F.9. Example 9

  • The stock forecast projects a climb from the low $20s toward the mid-to-upper $20s as confidence returns and the short-seller narrative is priced out.It attributes the V-shaped recovery to Neutron Stage 2 qualification, inclusion in NSSL and EWAAC, anticipated Q1 earnings, and the approaching H2 2025 Neutron debut window.Source: §F.9 · p. 27
  • The stock forecast links its bullish momentum to Neutron Stage 2 qualification and inclusion in the NSSL and EWAAC defense frameworks.It also cites anticipated strong Q1 earnings and the approaching H2 2025 window for the Neutron rocket’s debut as reasons momentum may continue.Source: §F.9 · p. 27

F.10. Example 10 … F.16. Example 16

These examples combine macro-level trends with micro-level events or seasonal cues to forecast stock prices and regional activity. Forecasts range from event-driven equity gains to seasonal spring increases in local activity.

F.10. Example 10

  • Alphabet’s forecast rises from 210.45 to 238.10 over 13 weeks, combining a bullish trend with Fed policy, earnings, antitrust developments, and holiday advertising demand.The path includes pullbacks around regulatory and economic uncertainty, followed by gains after dovish Fed signals, strong Q3 earnings, and holiday-ad demand expectations.Source: §F.10 · p. 28

F.11. Example 11

  • The eight-week forecast projects a gradual rise in activity through April and May, as seasonal spring demand and holidays support growth despite a lower 2025 baseline.It anticipates a peak toward the end of May.Source: §F.11 · p. 28

F.12. Example 12

  • The four-week forecast projects steady weekly increases from late February through March, following the recurring seasonal rise despite a lower year-over-year baseline since 2022.The projection combines historical seasonal growth with spring events and improving weather.Source: §F.12 · p. 29

F.13. Example 13

  • Rocket Lab rose from the $11 range to over $64 over the past year; with preliminary Q3 earnings released October 3, its November 10 report will rely more on forward guidance than revenue surprises amid six-week catalysts.Source: §F.13 · p. 29
  • Rocket Lab’s forecast climbs from 66.50 on October 31 to 78.40 on December 5, with launches, earnings guidance, analyst reactions, and year-end milestones as catalysts.The projected sequence reaches 72.80 after the November 10 earnings report and 75.20 the following week, assuming positive forward-looking statements and analyst support.Source: §F.13 · p. 29

F.14. Example 14

  • Rocket Lab’s forecast anticipates near-term gains from Q4 2024 earnings and analyst reactions, followed by brief consolidation before renewed growth around Spring 2025 catalysts and possible defense contracts.The forecast combines the company’s backlog and mission execution with upcoming event-driven catalysts.Source: §F.14 · p. 29

F.15. Example 15

  • The 13-week Washington, DC activity forecast projects a gradual spring increase by aligning the historical January dip and spring climb with the Cherry Blossom Festival and Easter.The cited historical cycle spans 2022 to 2025.Source: §F.15 · p. 30

F.16. Example 16

  • The eight-week Washington, DC activity forecast projects week-over-week rises from the late-January trough, with Presidents’ Day and the Cherry Blossom Festival expected to drive surges.The projection applies the early-2025 year-over-year baseline increase to the historical seasonal growth curve.Source: §F.16 · p. 30

Related papers

Loading 2605.14389v1…