CompEvo: Competition-Induced Evolution for Multi-Agent in News-Driven Time Series Forecasting

Yuxuan Zhang; Yangyang Feng; Yong Guan; Daifeng Li; Kexin Zhang; Junlan Chen; Bowen Deng; Jun Liu

cs.NESep 3, 2026

CompEvo: Competition-Induced Evolution for Multi-Agent in News-Driven Time Series Forecasting

Authors: Yuxuan Zhang, Yangyang Feng, Yong Guan, Daifeng Li, Kexin Zhang, Junlan Chen, Bowen Deng, Jun Liu, +1 more

Organizations: Sun Yat-sen University, Guangzhou, China · North China Electric Power University, Beijing, China · Unilumin Group Co., Ltd., Shenzhen, China

Abstract

News-driven time series forecasting uses evolving textual events together with historical observations to predict future values, supporting applications such as market risk monitoring and resource scheduling. In multi-agent settings, two challenges still remain. The first is degeneration of thought, where agents converge to similar evidence-seeking behaviors. The second is insufficient theoretical grounding, where strategy updates are often heuristic and lack a principled formulation. To address the above challenges, we propose CompEvo, a competition-induced evolution framework for multi-agent news-driven time series forecasting. For theoretical grounding, we introduce an evolutionary game formulation to guarantee equilibrium existence and optimization convergence. Building on this formulation, we construct a trainable multi-agent evolution framework that integrates strategy execution, fitness-based differentiable selection, and competition-induced strategy evolution. CompEvo enables heterogeneous agents to explore diverse news evidence, converts forecasting feedback into differentiable influence weights, and evolves agent strategies under competitive pressure to preserve effective logic while maintaining diversity. Experiments on four real-world datasets show that CompEvo reduces RMSE by 27.3% and MAPE by 26.2% on average over strong baselines. Further analysis indicates that CompEvo successfully maintains diverse and specialized agent behaviors.

Figures & tables

Figure 2: Overview of CompEvo for multi-agent NTSF.

Dataset	Metrics	CompEvo	DLin.	iTrans.	FiLM	Pyra.	FED.	GPT4TS	N2F	GPT4MTS	LTP-LLM
Electricity	MAE	220.15 $\pm$ 2.4	255.70	233.58	254.05	544.64	238.77	236.91	250.71	229.40	232.10
	MSE ${}_{\times 10^{3}}$	125.44 $\pm$ 1.8	135.05	97.61	153.90	625.00	133.96	142.60	192.24	133.50	136.80
	RMSE	354.18 $\pm$ 2.5	367.49	312.42	392.38	790.54	365.44	377.62	438.45	365.38	369.86
	MAPE (%)	6.45 $\pm$ 0.05	7.29	6.86	6.81	8.26	7.27	6.72	7.60	6.72	6.85
Exchange	MAE ${}_{\times 10^{-3}}$	4.35 $\pm$ 0.08	6.96	27.04	5.24	40.18	35.19	15.05	13.43	4.82	4.75
	MSE ${}_{\times 10^{-4}}$	0.35 $\pm$ 0.01	9.06	11.59	0.77	24.50	18.45	4.01	17.72	0.46	0.49

Table 1: Performance comparison on four datasets. Results for CompEvo are reported as mean

\pm

standard deviation over 5 runs. A one-sided Wilcoxon signed-rank test against LTP-LLM and GPT4MTS gives

p<0.05

on each dataset. Bold and underlined values mark the lowest and second-lowest errors.

Method	Agents	Model(s)	Trainable params	Train steps
SA	1	Llama-3.1-8B	41.9M	2,193
MAEL	10	Llama-3.1-8B	419.4M	2,193
MAE-GPT	10	GPT-4o + Llama-3.1-8B	419.4M	2,193
MAD-RL	10	Llama-3.1-8B	1,258.4M	2,193
CompEvo	10	Llama-3.1-8B	922.8M	2,193

Table 2: Agent-based baseline configurations. All methods use a 4,096-token context. Counts exclude frozen backbone parameters and the frozen GPT-4o selector in MAE-GPT.

Dataset	Metrics	CompEvo	SA	MAEL	MAE-GPT	MAD-RL
Electricity	MAE	220.15	245.50	246.13	254.61	248.76
	MSE ${}_{\times 10^{3}}$	125.44	185.10	177.68	195.15	166.67
	RMSE	354.18	430.20	421.52	441.76	408.25
	MAPE (%)	6.45	7.45	7.42	7.49	6.98
Exchange	MAE ${}_{\times 10^{-3}}$	4.35	13.10	11.26	11.69	9.03
	MSE ${}_{\times 10^{-4}}$	0.35	17.20	12.82	18.32	8.24

Table 3: Agent baseline comparison.

Model	RMSE	MSE	MAE	MAPE (%)
Mistral-v0.1-7B	366.84	134.57	245.13	7.18
Gemma-2-9B	364.12	132.58	228.14	6.69
Llama-2-7B	368.53	135.81	248.79	7.25
Llama-3.1-8B	354.18	125.44	220.15	6.45
Llama-3.1-70B	342.57	117.35	212.08	6.22
Qwen-2.5-7B	363.27	131.97	227.43	6.68

Table 4: Backbone robustness on Electricity. MSE is reported in

10^{3}

, and MAPE is reported in percentage.

Model Variant	Electricity				Exchange
Model Variant	RMSE	MSE $\times 10^{3}$	MAE	MAPE $(\%)$	RMSE $\times 10^{-2}$	MSE $\times 10^{-4}$	MAE $\times 10^{-3}$	MAPE $(\%)$
Simplified Logic Gen	445.53	198.50	255.24	7.65	4.10	16.81	14.84	4.85
Uniform Weights	452.12	204.41	265.53	7.93	4.82	23.23	20.53	5.65
w/o $L_{\text{diversity}}$	410.56	168.56	240.20	7.05	3.88	15.05	9.50	3.84
w/o $L_{\text{pruning}}$	405.22	164.20	238.51	6.95	3.76	14.13	9.81	3.92
w/o GRPO	388.51	150.94	242.08	7.12	1.82	3.31	7.97	2.82

Table 5: Architectural ablations across four datasets. Each variant changes one component while retaining the remaining model and training setup.

Metric	Original test	2026 inference
MSE	125440.00	111141.83
RMSE	354.18	333.38
MAE	220.15	225.02
MAPE (%)	6.45	6.99

Table 6: Post-training inference on temporally later Electricity samples.

Figure 3: Training dynamics on the Electricity dataset.

Figure 4: Agent reward during training, where the plateau aligns with the Nash-equilibrium interpretation.

Metric	MAEL	MAD-RL	CompEvo
Train time (h/ep)	11.4	40.0	14.5
Latency (s/step)	5.61	19.70	8.57
Peak memory (GB)	14.8	24.6	16.4

Table 7: Measured efficiency on Electricity under the three-agent setting.

Figure 5: Strategy-update diversity and agent specialization. (a) Average LUD per agent across epochs. (b) Rank plot of unique news counts over 100 steps.

Figure 6: Diversity and pruning regularization.

Figure 7: RMSE under news perturbations.

Figure 8: Agent strategy embeddings, red (MAD-RL) and blue (CompEvo).

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Domain	Time-series source	Variables	Time points	News source
Electricity	Energy	AEMO operational demand Australian Energy Market Operator (n.d.)	19	52,560	news.com.au/GDELT Leetaru and Schrodt (2013)
Exchange	Finance	Standard exchange-rate series Lai et al. (2018)	7	1,460	news.com.au/GDELT Leetaru and Schrodt (2013)
Bitcoin	Cryptocurrency	BitInfoCharts-derived series BitInfoCharts (n.d.)	18	858	GDELT/domain crawls Leetaru and Schrodt (2013)
Traffic	Transportation	Caltrans PeMS California Department of Transportation (n.d.)	862	17,544	GDELT/domain crawls Leetaru and Schrodt (2013)

Appendix

Table 8: Dataset provenance and time-series scale.

Dataset	Representative news example
Electricity	South Australia is only days away from a heatwave which will last for almost a week and has left Tour Down Under organisers anxiously watching the weather forecast.
Exchange	The RBA has dramatically revised down its economic forecasts amid the ongoing property market correction, prompting the Australian dollar to plunge again.
Traffic	A funnel cloud was spotted over Waterford in northern California on April 27 as a line of storms brought heavy rain and hail to the area.
Bitcoin	Personal finance expert Peter Adeney, known as ‘Mr. Money Mustache,’ has warned against investing in bitcoin, calling it a speculative asset rather than a true investment.

Appendix

Table 9: Representative news examples for each dataset.

Hyperparameter	Value
Number of agents	10
Validation folds	5
Adapter rank	16
Adapter alpha	32
Adapter dropout	0.05
GRPO group size	8

Appendix

Table 10: Main implementation hyperparameters.

Method	Agents	Model(s)	Trainable params	Train steps
SA	1	Llama-3.1-8B	41.9M	2,193
MAEL	10	Llama-3.1-8B	419.4M	2,193
MAE-GPT	10	GPT-4o + Llama-3.1-8B	419.4M	2,193
MAD-RL	10	Llama-3.1-8B	1,258.4M	2,193
CompEvo	10	Llama-3.1-8B	922.8M	2,193

Appendix

Table 11: Multi-agent baseline configurations. All methods use a context length of 4,096 tokens. Trainable counts exclude frozen backbone parameters and the frozen GPT-4o selector in MAE-GPT.

Metric	Original test	2026 inference
MSE	125440.00	111141.83
RMSE	354.18	333.38
MAE	220.15	225.02
MAPE (%)	6.45	6.99

Appendix

Table 12: Leakage-controlled inference results on Electricity. The 2026 samples are used only for post-training inference and are not used for training, validation, model selection, or hyperparameter tuning.

Figure 9: RMSE under news perturbations.

Panel of Figure 10 — Figure 10: Sensitivity analysis on diversity and pruning regularization. Both regularizers work best at moderate strength.

Figure 11: Impact of the initial number of agents on prediction accuracy.

$\lambda_{PG}$	Forecasting performance
$\lambda_{PG}$	RMSE	MSE	MAE	MAPE
0.0 (Ensemble)	398.50	158.00	248.00	7.25%
0.1	375.20	142.00	232.00	6.85%
0.5 (CompEvo)	354.18	125.44	220.15	6.45%
1.0	382.40	148.00	238.00	7.05%
5.0	435.00	189.00	268.00	7.95%

Appendix

Table 13: Sensitivity analysis of

\lambda_{PG}

on the Electricity dataset.

Figure 12: Impact of generation temperature on forecasting performance.

Temperature ( $\tau$ )	Forecasting performance
Temperature ( $\tau$ )	RMSE	MSE	MAE	MAPE
0.1	378.50	144.00	235.00	6.92%
0.5 (CompEvo)	354.18	125.44	220.15	6.45%
1.0	372.00	139.00	230.00	6.80%
5.0	415.00	170.00	255.00	7.60%

Appendix

Table 14: Sensitivity analysis of aggregation temperature (

\tau

) on the Electricity dataset.

Figure 13: Aggregation temperature sensitivity.

Figure 14: Robustness to paraphrasing the agent-side logic-update instruction. “Before Paraphrasing” uses the original agent-side update instruction, while “After Paraphrasing” uses a semantically equivalent rephrased version. The LLM judge prompt is kept fixed. Results are reported as MAPE across training epochs.

Judge Pair	Cohen’s $\kappa$	Pearson $r$	Agreement
Primary judge vs. Claude 4.6	0.75	0.81	Substantial
Primary judge vs. GPT-5.2	0.72	0.76	Substantial
GPT-5.2 vs. Claude 4.6	0.68	0.74	Moderate–Substantial

Appendix

Table 15: Robustness of LUD across LLM judges.

LUD Regime	Avg. LUD	Avg. $\Delta$ RMSE Gain	Interpretation
Low	0.09	1.12	Statistically weak local adjustment
Medium	0.44	7.85	Steady incremental optimization
High	0.76	14.20	High-impact strategic revision

Appendix

Table 16: Relationship between LUD magnitude and forecasting improvement on the Electricity dataset.

Panel of Figure 15 — Figure 15: Prompt templates used in CompEvo.

Explore similar work

When Tomorrow Becomes Today: Self-Evolving Policies for Agentic Time-Series Forecasting
Sep 21, 2026Yifan Hu, Xilin Dai, Zhiyuan Qu +4Time Series Forecasting Self-Evolving Agents
Ant International · Tsinghua University · The Chinese University of Hong Kong
TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent
Sep 23, 2026Jie Yang, Yan Zheng, Jiarui Sun +8Time Series QA LLM Agent Self-Improvement
University of Illinois at Chicago · Visa Research · University of Florida +2
EvoCast: Reliable Autonomous Research Agents for Iterative Forecasting Architecture Evolution
Oct 3, 2026Kaipeng Xu, Xianli Yan, Yan Wang +2Autonomous Research Agents Automated Machine Learning
School of Automation and Intelligent Science, Jiangnan University, Wuxi 214122, China · School of Automation, Southeast University, Nanjing 210096, China

Sep 21, 2026, cs.LGJ/K move · Enter open · S save

When Tomorrow Becomes Today: Self-Evolving Policies for Agentic Time-Series Forecasting

Yifan Hu, Xilin Dai, Zhiyuan Qu, Yiding Liu+3

Ant International · Tsinghua University · The Chinese University of Hong Kong

Time Series Forecasting Self-Evolving Agents

Open paper AskPDF arXiv

Agentic time series forecasting concerns systems whose underlying mechanisms evolve, making the relative effectiveness of numerical models, reasoning strategies, and intervention rules inherently time-varying. Consequently, a time series agent must adapt the forecasts it produces and the orchestration policy that determines which components to trust and how to coordinate them. The deployment process naturally provides supervision for this adaptation as forecast horizons elapse and realized targets reveal the effectiveness of earlier decisions. Committing all numerical expert forecasts and candidate agent paths before target observation allows each realized outcome to evaluate the entire alternative set, providing delayed feedback without additional annotation. However, existing time series agents primarily incorporate prior experience through forecast refinement, reflection, or retrieval, without systematically converting realized outcomes into persistent updates to the joint orchestration policy governing later origins. To exploit this delayed feedback systematically, we introduce TimEvolve, a frozen-backbone time series agent that converts each realized outcome into persistent joint updates of expert trust, agent path selection, and intervention strength. A temporally ordered predict, reveal, and update protocol applies this feedback to subsequent forecasts. Experiments across eight Time-MMD domains show that TimEvolve achieves the best average MSE and MAE ranks among fifteen methods and the lowest errors on both metrics in seven domains. These results demonstrate the value of learning forecasting policies from the futures encountered during deployment.

Abstract

Explore similar work

When Tomorrow Becomes Today: Self-Evolving Policies for Agentic Time-Series Forecasting

TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent

EvoCast: Reliable Autonomous Research Agents for Iterative Forecasting Architecture Evolution