Source-linked AI summary
AutoScientist-Quant: Self-Evolving Coding Agents for Automatic Research in Quantitative Investment
Zongqian Li, Yaoyiran Li, Yaohui Guo, Ming Zhang, Nigel Collier, Eugene Ie
TL;DR
Existing alpha agents use fixed searches, stop before library and model selection, and can contaminate evaluation through feedback overlap. AutoScientist-Quant unifies adaptive alpha discovery, library filtering, and model search under a budget-aware controller with separated feedback and test windows. On CSI universes, it achieves the best value of nearly every metric across settings, backbones, and markets.
Problem
Existing agentic methods cannot adapt search decisions during runs, usually stop at alpha generation, and may overlap feedback with final evaluation.
Method
A single budget-aware controller adaptively searches alphas, filters the library, and tunes model configurations while using shared memory and a held-out window only for final evaluation.
Results
AutoScientist-Quant achieves the best value of all eight metrics in all four alpha settings among agentic methods on CSI 300, with the same broad lead across tested backbones and markets.
Takeaways & Limitations
The results support treating quantitative research as one adaptive search problem spanning discovery, library selection, and model choice.
Takeaways & Limitations
The held-out window is used exactly once for final evaluation and is never observed by any search stage.
Abstract
from arXiv · showhide
Large language model agents can discover alphas, yet current methods have three weaknesses. The search cannot adapt during the run, automation usually ends at alpha generation while library selection and model choice stay manual, and alpha discovery can read the test window through loop feedback or code problems. We present AutoScientist-Quant, a self evolving search process that regards quantitative research as one budgeted search problem. A single controller conditions every decision on the remaining budget, choosing at each round whether to improve, combine, pivot, or stop, which node to expand, how many alphas to generate, and how to retrieve past trajectories from the shared memory. The same core then selects from the library and tunes the model, closing the loop from hypothesis to deployable strategy. We also review the evaluation pipeline reused from prior work, fix two lookahead problems, and keep the feedback window disjoint from the held out test window, so every comparison tests true generalization. On CSI universes, the framework attains the best value of nearly every metric in every setting, and these conclusions hold across several backbones and markets.
1 INTRODUCTION
Quantitative alpha research needs scalable, continually renewed discovery, but existing agentic systems keep the search structure fixed, end automation at alpha generation, and risk test-window contamination. AutoScientist-Quant addresses these gaps with adaptive, closed-loop search and corrected evaluation separation.
- Markets are noisy and nonstationary, so individual alphas can become crowded and new alphas must be discovered and renewed at scale.
- Existing agentic methods use fixed directions, rounds, alpha counts, and operations, without adapting search depth to evidence or sharing validated experience across branches.
- Prior automation usually stops at alpha generation, leaving library selection and model search fixed and producing strong information quality but negative excess returns after costs.
- AutoScientist-Quant makes cost and structure decisions during the run, using one controller to select operations, branches, retrieval, and alpha counts under a global budget.
- The framework closes the research loop across alpha discovery, alpha filtering, and model search, while correcting lookahead problems and separating feedback from held-out evaluation.
2 METHODS
AutoScientist-Quant formulates quantitative research as one budgeted search over alpha libraries, subsets, and model configurations. A shared controller adapts each stage to remaining resources and evidence, while the held-out window is used only for final evaluation.
- 2.1 PROBLEM FORMULATION: Figure 1 organizes alpha discovery, alpha filtering, and model search as three linked budgeted tree searches sharing one decision core and memory.
- 2.1 PROBLEM FORMULATION: The framework jointly searches a discovered alpha library, an optional standard library, and a predictor to maximize feedback-window utility under global budget B.
- 2.2 ALPHA DISCOVERY: Alpha discovery builds trajectories of hypotheses, executable alpha sets, and backtest feedback, with all directions stored in shared memory visible to every branch.
- 2.2 ALPHA DISCOVERY: At each discovery round, the controller chooses IMPROVE, COMBINE, PIVOT, or STOP, selects a parent, and sets the number of generated alphas.
- 2.3 ALPHA FILTERING: Alpha filtering searches remove, add, and swap edits over a redundant library, scoring subsets by eight metrics standardized against the historic evaluated-node distribution.
- 2.4 MODEL SEARCH: Model search explores complete model–hyperparameter configurations, then trains the winning library and configuration on training and validation data before one held-out evaluation.
3 EXPERIMENTAL DESIGN
The evaluation uses historical data from six equity universes, multiple information and portfolio metrics, broad baseline categories, and repeated backbone and market robustness checks. Main and ablation tables focus on CSI 300 under GPT.
- 3.1 DATA: The dataset covers daily OHLCV data for CSI 300, CSI 500, CSI 800, CSI 1000, S&P 500, and NASDAQ 100 from January 2015 to May 2026.
- 3.2 EVALUATION METRICS: Results use four information metrics and four cost-adjusted portfolio metrics, averaged over three independent runs.
- 3.5 MAIN COMPARISON: Table 1 compares machine-learning, deep-learning, and agentic methods on CSI 300 under GPT across four alpha settings, including standalone and merged libraries.
- 3.3 BASELINES: Baselines include machine-learning, deep-learning, and agentic methods evaluated across ALPHA158, ALPHA360, and ALPHA20 libraries.
- 3.4 MODELS: GPT-OSS-120B and GLM-5 provide the two comparison backbones, with robustness analyses changing the backbone and extending evaluation across additional universes.
- 3.6 ABLATIONS: Table 2 ablates AutoScientist-Quant stages on CSI 300 under GPT, reporting absolute metrics plus gains over the best agentic value and gaps to the full method.
4 RESULTS
AutoScientist-Quant outperforms agentic and conventional baselines across information and portfolio metrics, while ablations show that both adaptive discovery and downstream selection contribute. The advantage persists across backbones and markets, with robustness analyses also revealing stronger gains in smaller-cap universes.
- Main results: 0.010 IC is the best static baseline, versus 0.023–0.034 for all three agentic methods.The agentic results are roughly three times the best machine-learning or deep-learning result, indicating additional information from hypothesis-guided discovery.
- Main results: 0.034 IC and 0.219 ICIR make AutoScientist-Quant the strongest agentic method in ALPHA158.QuantaAlpha reaches 0.031 IC and 0.190 ICIR, while AlphaAgent reaches 0.030 and 0.182; the ordering also holds for rank-based metrics.
- Main results: 3.5% excess return, 0.500 information ratio, and 0.368 Calmar ratio are achieved with ALPHA158.Excess return remains positive in all four settings, and AutoScientist-Quant has the smallest maximum drawdown among agentic methods in every setting.
- Main results: 0.034 IC and 3.5% ARR improve from 0.028 IC and 1.8% ARR when CUSTOM is augmented with ALPHA158.Every agentic method benefits from merged libraries, with ALPHA158 the strongest partner for all three methods.
- Ablation studies: Removing model search costs up to 2.2 percentage points of excess return while leaving IC nearly unchanged.Removing alpha filtering as well changes CUSTOM ARR from 0.9% to −0.3%, showing that downstream stages primarily affect portfolio outcomes.
- Ablation studies: 3.8 percentage points is the largest excess-return loss from removing dynamics in the discovery ablation.Removing the searching strategy lowers IC by up to 0.002 and CUSTOM excess return from −0.3% to −2.4%; discovery components affect both information and portfolio metrics.
- Ablation studies: 2.6 percentage points higher excess return is reached by partial methods versus the best agentic baseline in some cells.The results attribute the lead to combined stage-level and component-level contributions rather than one component.
- Model robustness: 0.036 IC with ALPHA158 and positive ARR in all four settings preserve the ranking under GLM.Seven of eight baseline settings lose money after costs, while the full method remains consistently profitable.
5 CONCLUSIONS
AutoScientist-Quant treats quantitative research as a unified search process whose exploration, downstream decisions, and evidence-based stopping jointly support strong results. On CSI universes, it leads on nearly every metric across settings, while the framework is proposed for broader domains.
- AutoScientist-Quant leads on nearly every metric in every setting across two backbones and four CSI universes.
- Exploration creates information, downstream stages turn it into returns, and the search recovers from unseen dips, controls redundancy, and stops on evidence.
- The same search framework can also be used for algorithm development, reasoning, and evaluation.