ACE-Trader: Adversarial Co-Evolution for Generator-Critic Trading Systems

Chang Zhou; Jiarun Zhou; Xingtong Yu; Hong Cheng; Hui Zhang; Xinming Zhang

cs.LGOct 19, 2026

ACE-Trader: Adversarial Co-Evolution for Generator-Critic Trading Systems

Authors: Chang Zhou, Jiarun Zhou, Xingtong Yu, Hong Cheng, Hui Zhang, Xinming Zhang

Organizations: University of Science and Technology of China · The Chinese University of Hong Kong

Abstract

Generator--critic trading systems pair an LLM-based trader that makes trading decisions with a red team agent that evaluates these decisions and provides feedback for revision. However, improving individual trading decisions does not ensure that both agents learn from experience and improve together over time. This raises an important question: how can the trader and red-Team select useful experience and learn from each other's evolving behavior? On one hand, experience that benefits decisions under one market condition may become ineffective or harmful under another. On the other hand, trading outcomes reflect both agents' actions, making it difficult to derive appropriate experience for each role. In this work, we propose ACE-Trader, an adversarial co-evolution trading framework. Specifically, we first develop separate context-dependent memory selectors that learn to identify useful experience for the current trading decision. Both selectors receive shared feedback from net-return differences between complete memory-assisted and memory-free trading pipelines, with online updates emphasizing recent outcomes. We further design a co-evolution mechanism that links role-specific memories through shared trading records. We use these records to evaluate the Red-Team's risk assessments and the effects of the Trader's revisions against realized market outcomes, then derive experience for their respective memories. By linking this experience to the corresponding records, we enable the Trader to improve its proposals based on prior critiques and the Red-Team to adjust its review focus as the Trader's behavior changes. Experiments on U.S. stock trading show that ACE-Trader outperforms all 22 active baselines in cumulative return and achieves the highest Sharpe ratio and lowest maximum drawdown across both active methods and passive benchmarks.

Figures & tables

Figure 1: Comparison of memory mechanisms in trading agents.

Category	Model	Cum. Ret. (%) $\uparrow$	Ann. Ret. (%) $\uparrow$	Sharpe $\uparrow$	MDD. (%) $\downarrow$	Vol. (%) $\downarrow$	HR. (%) $\uparrow$	Turnover
Passive	NASDAQ Buy & Hold	+12.82	+28.29	1.340	12.84	20.10	57.38	0.000
Passive	Equal-Weight Buy & Hold	+8.68	+18.77	1.375	5.33	13.14	52.48	0.000
ML	LightGBM	+4.86	+10.30	0.502	13.13	26.48	50.74	1.341
	XGBoost	-21.18	-38.84	-1.610	29.53	28.07	50.16	1.329
	LSTM	+2.81	+5.90	0.399	9.25	18.60	47.52	1.044
	GRU	+0.53	+1.09	0.148	11.16	17.42	49.59	1.028

Table 1: Portfolio performance on Stock.

Model	TS	RS	CE	Cum. Ret. (%) $\uparrow$	Sharpe $\uparrow$	MDD. (%) $\downarrow$	Vol. (%) $\downarrow$	HR. (%) $\uparrow$
Variant 1	$\times$	$\times$	$\times$	+5.53	0.823	7.81	14.84	53.38
Variant 2	$\checkmark$	$\times$	$\times$	+6.90	1.028	7.52	14.41	58.20
Variant 3	$\times$	$\checkmark$	$\times$	+8.56	1.261	6.33	14.25	62.30
Variant 4	$\checkmark$	$\checkmark$	$\times$	+10.07	1.529	6.36	13.57	55.74
ACE-Trader	$\checkmark$	$\checkmark$	$\checkmark$	+12.44	2.023	5.21	12.35	57.38

Table 2: Ablation study of memory selection and co-evolution.

Figure 3: Impact of $\alpha$ and $\gamma$ .

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Phase	Decision, evidence, and outcome
Trader: Commit	Proposal: Buy additional AMZN worth 2% of NAV; sell JPM worth 1% and XOM worth 1.5% of NAV. Rationale: AMZN’s RSI of 23.35 suggests an oversold rebound.
RT: Challenge	Evidence: AMZN’s 10- and 20-day returns are -5.01% and -8.42%, and its price is 6.61% below the 20-day moving average. Challenge: Low RSI alone does not justify adding exposure while the downtrend persists. Risk condition: AMZN’s return at the end of the five-trading-day horizon is below -3%.
Trader: Revise	Accept the challenge and cancel the additional AMZN purchase. Retain the existing AMZN shares without selling them. Keep the JPM and XOM sell orders unchanged.
Risk evaluation	AMZN returns -4.42% over the evaluation window ending on June 25, satisfying the pre-specified risk condition. The warning is a hit .
Revision evaluation	The original proposal yields -0.0148%, whereas the revised plan yields +0.0745%. Canceling the purchase improves net return by 0.0893% , turning a small evaluated loss into a gain.

Appendix

Table 3: Case from the initial proposal to the revised decision.

Absolute predicted return	Position-size level $q_{t,j}$
$0.2\%\leq\|\hat{r}_{t,j}\|\leq 0.6\%$	1
$0.6\%<\|\hat{r}_{t,j}\|\leq 1.2\%$	2
$1.2\%<\|\hat{r}_{t,j}\|\leq 1.8\%$	3
$1.8\%<\|\hat{r}_{t,j}\|\leq 2.4\%$	4
$\|\hat{r}_{t,j}\|>2.4\%$	5

Appendix

Table 4: Mapping absolute predicted returns to position-size levels.

Algorithmic Trading

Algorithmic trading uses computational models to generate signals, allocate portfolios, and execute orders across financial markets.

See all 149 papers

EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents
Sep 15, 2026Sehee Kim, Yumin Choi, Minki Kang +1LLM Agent Self-Improvement Prompt Optimization
KAIST
Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models
Apr 18, 2026Yuchen Pan, Soung Chang LiewLLM Evaluation Financial Reasoning in Language Models
Department of Information Engineering, The Chinese University of Hong Kong
ContestTrade: A Multi-Agent Trading System Based on Internal Contest Mechanism
Aug 1, 2025Rui Sun, Li Zhao, Zuoyou Jiang +5Multi-Agent LLM Systems LLM Agent Orchestration
1Stepfun · 2FinStep

Sep 15, 2026J/K move · Enter open · S save

EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents

Sehee Kim, Yumin Choi, Minki Kang, Sung Ju Hwang

KAIST

LLM Agent Self-Improvement Prompt Optimization Algorithmic Trading Self-Evolving Agents RL for Tool Use

Open paper AskPDF arXiv

Large language model (LLM) trading agents can combine market data, news, and executable analysis, but their behavior is often controlled by static hand-written tool-use policies that are fixed before deployment. This limits their ability to adapt how they gather evidence, invoke tools, verify signals, and manage risk under changing market regimes. We introduce EvolveTrade, a self-evolving framework that treats the system prompt of a tool-using trading agent as a text-parameterized policy. After each update interval, a Policy Agent revises this policy using accumulated decision traces and realized portfolio feedback, while keeping the backbone LLM fixed. The updated policy is then used for the next batch of trading decisions, enabling the agent to refine its information-acquisition and portfolio-construction procedure over time. Experiments across multiple market regimes and two LLM backbones show that EvolveTrade often improves Sharpe Ratio and Cumulative Return over fixed-policy LLM baselines, achieving the improved SR and CR in most evaluated settings. Behavioral analyses further show that self-evolved policies increase code-mediated analysis and activate regime-relevant computations; case-level policy-to-return attributions trace how policy-induced allocation changes contribute to realized return differences. These results suggest that adapting the reusable procedure governing tool use is a key direction for building more robust LLM trading agents.

Absolute predicted return	Position-size level $q_{t,j}$
$0.2\%\leq\|\hat{r}_{t,j}\|\leq 0.6\%$	1
$0.6\%<\|\hat{r}_{t,j}\|\leq 1.2\%$	2
$1.2\%<\|\hat{r}_{t,j}\|\leq 1.8\%$	3
$1.8\%<\|\hat{r}_{t,j}\|\leq 2.4\%$	4
$\|\hat{r}_{t,j}\|>2.4\%$	5

Abstract

Algorithmic Trading

Explore similar work

EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents

Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models

ContestTrade: A Multi-Agent Trading System Based on Internal Contest Mechanism