PaperlayerView exact revision on arXiv

Source-linked AI summary

MACTS-EM: Multi-Agent Collaborative Time Series Forecasting with Emergent Memory

Ahmad Shahi, Mamehgol Yousefi

arXiv:2610.02255v1cs.LGcs.MA

TL;DR

Time-series forecasting methods struggle with regime shifts, cross-domain transfer, and multimodal integration. MACTS-EM coordinates specialised forecasting agents with emergent memory and related components; it improves forecasting accuracy, transfer, and distribution-shift outcomes in most reported scenarios.

  • Problem

    Existing approaches struggle with regime shifts, cross-domain knowledge transfer, and multimodal data integration in dynamic, interconnected systems.

    Source: §1 · p. 1

  • Method

    MACTS-EM coordinates specialised forecasting agents through a meta-cognitive layer and emergent memory, alongside multimodal integration and adversarial robustness components.

    Source: §1 · pp. 2–3, §2 · p. 3

  • Results

    Across most domains and horizons, MACTS-EM improves forecasting accuracy by 8.2–12.0%, improves zero-shot transfer by 22.0–26.8% in five of six scenarios, and enhances distribution-shift resilience by 16.2–20.6% in three of four domains.

    Source: §4.4 · pp. 12–13

  • Takeaways & Limitations

    The framework offers potential uses in early warning, resource allocation during rapid shifts, cross-domain forecasting in data-scarce contexts, and interpretable predictions.

    Source: §5 · p. 13

    MACTS-EM incurs 36% more training and approximately 42% more inference overhead than TimeMixer, and performance diminishes for forecasts beyond 90 days.

    Source: §5 · p. 13

Abstract

from arXiv · show

Time series forecasting remains a critical challenge across numerous domains. Despite significant advancements, existing approaches struggle with complex phenomena such as regime shifts, cross-domain knowledge transfer, and multimodal data integration. This paper introduces Multi-Agent Collaborative Time Series Forecasting with Emergent Memory (MACTS-EM), a novel framework where specialised agents collaborate to achieve superior forecasting performance. The MACTS-EM architecture integrates: (1) domain-specialised forecasting agents for pattern recognition, anomaly detection, causal inference, and uncertainty quantification; (2) a meta-cognitive layer for dynamic agent allocation; (3) an emergent memory mechanism enabling cross-domain pattern transfer; (4) multimodal contextual integration; and (5) adversarial robustness components. Evaluation across financial markets, climate patterns, energy consumption, and pandemic propagation demonstrates that MACTS-EM outperforms existing approaches in most scenarios, with 8-12% improvement in forecasting accuracy, 22-27% better zero-shot transfer capability, 16-21% enhanced resilience during regime shifts, and 15-18% faster recovery after distribution shifts. Our findings suggest that collaborative, agentic approaches to time series forecasting represent a promising direction beyond traditional architectures, particularly for complex real-world scenarios requiring multi-resolution temporal understanding and contextual adaptation.

1 Introduction

Forecasting methods struggle with regime shifts, cross-domain transfer, and multimodal data integration in dynamic, interconnected systems. MACTS-EM proposes collaborative forecasting by specialised agents coordinated through a meta-cognitive layer and emergent memory.

1 Introduction

  • Existing forecasting approaches struggle with regime shifts, cross-domain knowledge transfer, and multimodal integration in dynamic, interconnected systems.The paper situates this challenge across real-world systems including finance, healthcare, climate science, and energy management.Source: §1 · p. 1
  • MACTS-EM reframes forecasting as collaboration among specialised agents coordinated by a meta-cognitive layer and emergent memory.Its agents communicate intermediate representations, while memory supports cross-domain transfer and allocation adapts to context-specific performance.Source: §1 · p. 2
  • The proposed architecture combines specialised forecasting components and explicit communication protocols, with memory-based transfer, dynamic allocation, multimodal integration, and robustness components.The paper identifies quantitative transfer experiments and evaluation against state-of-the-art baselines among its contributions.Source: §1 · pp. 2–3

2 MACTS-EM Framework

MACTS-EM organizes forecasting around specialised agents that communicate and are coordinated by a meta-cognitive layer. Emergent memory supports cross-domain pattern transfer, while multimodal inputs, robustness measures, and staged training complete the framework.

2 MACTS-EM Framework

  • MACTS-EM defines forecasting as collaboration among specialised agents coordinated by a meta-cognitive layer and emergent memory, with optional covariates and multimodal data.The framework is represented by components for agents, coordination, memory, multimodal integration, and adversarial robustness.Source: §2 · p. 3

2.1 Specialised Agent Modules

  • Four specialised agents address pattern recognition, anomaly detection, causal inference, and uncertainty, exchanging features, forecast confidence, and abstract knowledge.Their models include modified N-BEATS, a TCN, BlockRNN with LSTM cells, and LightGBM quantile regression.Source: §2.1 · p. 4

2.2 Meta-Cognitive Layer

  • The meta-cognitive layer extracts time-series characteristics and computes agent weights using parameters learned during training.The learned parameters are optimized for diverse forecasting scenarios.Source: §2.2 · p. 4

2.3 Emergent Memory Mechanism

  • Emergent memory stores domain identifiers and pattern vectors, identifies related domains, and adapts transferred patterns using learned coefficients.Meta-learning and cross-validation tune adaptation coefficients; contrastive learning with triplet loss clusters similar patterns.Source: §2.3 · pp. 4–5

2.4 Multimodal Integration and Robustness

  • Modality-specific encoders combine text, visual, and categorical data through cross-modal attention, while adversarial scenarios assess robustness to shifts.A two-level weighting scheme adjusts agent weights under distribution shifts and defaults to base weights during stable periods.Source: §2.4 · p. 5

2.5 Training and Inference

  • Training first pretrains agents independently, then trains them collaboratively; inference selects agents, forecasts, and aggregates predictions with uncertainty estimates and explanations.The ensemble loss balances forecasting accuracy with diversity, robustness, and transfer.Source: §2.5 · p. 5

3 Experimental Setup

The experiments evaluate forecasting across four domains using accuracy, robustness, and transfer metrics, with comparisons against statistical, deep-learning, foundation-model, and ensemble baselines. Protocols test forecast horizons, zero-shot transfer, distribution shifts, multimodal inputs, and component ablations.

3.1 Datasets and Metrics

  • The evaluation covers financial markets, climate patterns, pandemic propagation, and energy consumption, measuring MAPE, domain-normalised RMSE, shift resilience, recovery time, and zero-shot target-domain MAPE.DSR is defined so higher values are better, with DSR = 1 indicating no degradation after a shift.Source: §3.1 · pp. 5–6

3.2 Baselines and Implementation

  • MACTS-EM is compared with statistical, deep-learning, LLM-based and foundation models, and a stacking ensemble; implementation uses walk-forward cross-validation and selective agent activation.Selective activation, incremental memory updates, and parallel execution reduce inference time by 42% versus a naive implementation.Source: §3.2 · p. 6

3.3 Experimental Protocols

  • Protocols test short-, medium-, and long-horizon forecasting, zero-shot domain transfer, natural and synthetic distribution shifts, multimodal inputs, and component ablations.Forecast horizons span 1–7, 8–30, and 31–90 days; inputs include domain-specific news sentiment, satellite imagery, mobility data, and weather forecasts.Source: §3.3 · pp. 6–7

4 Results and Analysis

MACTS-EM generally outperformed baselines across domains and horizons, with gains in forecasting accuracy, zero-shot transfer, shift resilience, and recovery. Domain-specific results and component analyses show where these gains were strongest and how performance varied with context.

4.1 Forecasting Accuracy

  • 8.2–12.0% improvement over the best baselines occurred in most forecasting scenarios, with statistical significance (p < 0.01).MACTS-EM slightly underperformed in short-term climate and medium-term energy forecasting.Source: §4.1 · p. 7

Domain-Specific Performance

  • 38.8% improvement vs. TimeMixer during the COVID-19 crash accompanied financial gains of 7.3% during stable periods; climate long-term projections had 12.0% lower error rates, including 27.3% improvement for Arctic sea ice.News sentiment integration contributed 25% of the financial improvement; satellite imagery reduced climate error by 18.2% during summer melting periods, and anomaly detection weight rose from 15% to 37% during extreme events.Source: Domain-Specific Performance · p. 8
  • Pandemic forecasts detected variants 2.3 days earlier and estimated peak timing at ±1.8 days versus ±4.6 days for baselines; energy forecasting improved 8.4% routinely, 23.7% during heatwaves, and 28.4% during cold snaps.Mobility integration reduced pandemic MAPE by 24.7% during interventions, while weather data reduced energy error by 32.3% in rapidly changing conditions; energy prediction intervals had 92% coverage and 35% narrower width than baselines.Source: Domain-Specific Performance · pp. 8–9

4.2 Cross-Domain Knowledge Transfer and Robustness

  • 22.0–26.8% better zero-shot transfer appeared in most scenarios, while distribution-shift resilience improved 16.2–20.6% in three of four domains and recovery was 15.3–18.2% faster.Pandemic→Energy transfer slightly underperformed by 1.1%; dynamic weighting favored anomaly and uncertainty agents during shifts, and adversarial training exposed the model to perturbations.Source: §4.2 · pp. 9–10

4.3 Multimodal Integration and Ablation Studies

  • Multimodal integration yielded 1.6–1.7× greater MAPE reduction than baselines, while removing the meta-cognitive layer, emergent memory, or multimodal integration increased MAPE by 13.4%, 11.1%, or 10.1%, respectively.With 50% of multimodal data, the framework retained 83% of its performance gain, and with 25% it retained 62%.Source: §4.3 · pp. 10–11
  • During stable markets, pattern recognition received 42% weight, whereas volatility shifted the largest weights to anomaly detection (39%) and uncertainty quantification (32%).In behavioral profiles, Sentinel messaging rose sharply in crises while Analyst activity was highest outside crises; Analyst confidence ranked highest and Realist confidence lowest.Source: §4.3 · pp. 11–12
  • Energy–Climate (0.78), Financial–Pandemic (0.63), and Climate–Financial (0.57) were the strongest reported pattern similarities; optimized MACTS-EM-RT inference was 3.7× faster with 2.3% accuracy reduction.MACTS-EM required 36% more training time than TimeMixer; optimized medium-term inference averaged 285 ms versus 201 ms for TimeMixer.Source: §4.3 · p. 12

4.4 Summary of Findings

  • Across most domains and horizons, forecasting accuracy improved 8.2–12.0%; zero-shot transfer improved 22.0–26.8% in five of six scenarios, and shift resilience improved 16.2–20.6% in three of four domains.Recovery after perturbations was 15.3–18.2% faster, and multimodal integration benefits exceeded those of baselines.Source: §4.4 · pp. 12–13

5 Implications and Limitations

MACTS-EM is presented as a shift to collaborative forecasting, with emergent memory suggesting shared patterns across domains. The paper describes practical uses but reports computational, domain-specific, tuning, and long-horizon limitations.

5 Implications and Limitations

  • MACTS-EM’s emergent memory suggests shared pattern structures across domains, while specialised expertise with coordination outperformed general-purpose models.The paper frames this as a theoretical and practical shift from monolithic models to collaborative agent-based systems.Source: §5 · p. 13
  • The framework is proposed for early warnings, resource allocation, data-scarce cross-domain forecasting, and interpretable predictions; early warnings may gain 2–4 days.The stated examples include market crashes, disease outbreaks, healthcare, and energy management.Source: §5 · p. 13
  • Limitations include 36% training and approximately 42% inference overhead versus TimeMixer, weaker results in several domains, careful-tuning needs, and diminished performance beyond 90 days.The listed underperformances are short-term climate forecasting (−1.8%), medium-term energy forecasting (−0.9%), Pandemic→Energy transfer (−1.1%), and energy distribution-shift resilience (−2.9%).Source: §5 · p. 13

6 Conclusion and Future Directions

MACTS-EM improves on state-of-the-art models in most scenarios, with its strongest reported advantages in challenging events. The conclusion also points to architectural, methodological, and application extensions for future work.

6 Conclusion and Future Directions

  • 38.8% improvement during market crashes and 23.7–28.4% improvement during extreme weather accompany broader gains in most scenarios; pandemic waves saw earlier variant detection.The conclusion identifies these challenging situations as where the framework’s strongest advantages emerge.Source: §6 · p. 13
  • Ablation studies indicate that MACTS-EM’s coordinated information flow achieves capabilities exceeding individual component contributions.The paper presents this as a promising shift toward collaborative AI systems inspired by human cognitive organisation.Source: §6 · p. 14
  • Future directions include hierarchical architectures, dynamic agent creation, continual learning, enhanced causal inference, spatiotemporal forecasting, and optimisation integration.The authors also plan LLM-based agents for textual context and natural-language explanations, plus reinforcement learning for communication and meta-learning for adaptation.Source: §6 · p. 14
Loading 2610.02255v1…