XALPHA: A Memory-Driven AI Quant Researcher for Hypothesis-to-Code Alpha Discovery

Abstract: Financial markets are noisy, non-stationary, and high-dimensional, making it difficult to discover predictive and robust trading signals. Alpha discovery has evolved from manual factor design to machine learning, evolutionary search, and recent LLM-based frameworks, improving the efficiency of factor generation, search, and evaluation. However, existing methods still mostly automate isolated steps, rather than functioning as end-to-end quant researchers that can absorb external knowledge, close the hypothesis-to-code validation loop, and learn from accumulated discovery feedback. To fill this gap, we introduce XAlpha, a memory-driven AI Quant Researcher for continuous hypothesis-to-code alpha discovery. XAlpha maintains a multi-source research memory system that integrates report-grounded financial knowledge with discovery feedback from prior generations and research cycles. Guided by this memory system, a Macro Brain plans research themes and selects suitable Archetypes; a Micro Brain transforms the planned hypothesis pool into executable factor code and verifies ex-ante tri-alignment among the hypothesis idea, code logic, and financial plausibility; and a Cross Brain consolidates empirical outcomes into generation-level feedback, cycle-level summaries, and archetype-level research cues for future exploration. In this way, XAlpha turns alpha mining from isolated factor generation into a closed-loop research process that continuously reads, hypothesizes, implements, validates, reflects, and evolves. Experiments on CSI300 show that XAlpha achieves stronger overall alpha discovery performance than representative baselines.

• The paper introduces XAlpha, a closed-loop AI quant researcher that combines report-to-memory absorption, hypothesis planning, executable factor evolution, leakage testing, and mechanism-level feedback.
• XAlpha achieves the strongest IC (0.0413), ICIR (0.2573), RankICIR (0.3412), annualized return (17.92%), excess return (14.12%), and information ratio (1.5853) among the reported baselines on CSI300 data, although CogAlpha records the highest RankIC (0.0590).
• The system produces diverse, interpretable factors and improves search efficiency through memory-driven routing, but its generalizability remains unproven because experiments use one market, horizon, data contract, backend, and set of empirically chosen thresholds.

Motivation and problem statement

Alpha mining—the discovery of predictive trading signals mapping historical market observations to future asset returns—has progressed from manual factor design through machine learning and evolutionary search to LLM-based frameworks. The authors of XAlpha (2607.08332) argue that this progression has largely automated isolated factor-level operations (search, generation, evaluation) rather than the full hypothesis-driven workflow of a human quant researcher: absorbing market knowledge, formulating hypotheses, translating them into executable code, validating financial plausibility, evaluating empirically, and reusing both successes and failures. Existing LLM-based systems such as AlphaGPT, AlphaAgent, AlphaJungle, RD-Agent(Q), FactorMiner, and CogAlpha each address parts of this loop but do not integrate external knowledge absorption, research memory, hypothesis planning, code implementation, empirical validation, and feedback consolidation into a single closed-loop agent. XAlpha is proposed as an "AI Quant Researcher" that closes this gap by adapting the autonomous research-agent paradigm (e.g., The AI Scientist, AI Co-Scientist, EvoScientist) to quantitative finance.

Architecture

XAlpha comprises a Report-to-Memory Absorption (RMA) layer and three coordinated "brains," all operating over a multi-source research memory that combines report-grounded knowledge with discovery feedback from prior generations and cycles.

RMA layer. External research reports are converted into structured, retrievable knowledge via a three-layer A/B/C taxonomy rather than being placed raw into prompts. The A-layer screens report fragments for eligibility under the daily OHLCV data contract: fragments whose mechanisms cannot be observed or stably proxied from daily open/high/low/close/volume (e.g., order-book states, fundamentals) are dropped. The B-layer assigns retained evidence to broad mechanism families (trend/momentum, reversal, price-volume interaction, etc.). The C-layer specializes these into 48 Research Archetypes—structured research cues that are explicitly not factor formulas. This design prevents downstream planning of hypotheses that cannot be implemented under the available data contract.

Macro Brain. Operating at cycle level, it selects a routing mode (fixed-theme, coarse-guided, or memory-driven), determines the cycle theme, chooses one primary B-layer plus three or four supporting B-layers, retrieves compatible C-layer archetypes, and materializes a fixed bundle of eight active research agents with an archetype-guided hypothesis pool. By default, five coarse-guided cycles initialize feedback before switching to memory-driven routing.

Micro Brain. This inner loop transforms hypotheses into executable Python factor code over OHLCV primitives, then evolves candidates through mechanism-level mutation, crossover, and refinement operators, with periodic novelty injection at generations 3 and 7 of each 10-generation cycle. Every candidate must pass a quality pipeline consisting of an AST gate (syntax, unsupported fields, static future leakage such as negative shifts, complexity limits), an ex-ante tri-alignment judge verifying consistency among hypothesis idea, code logic, and financial rationale, numerical validation (invalid-value ratio capped at 30%), and unit tests based on truncation and future-noise perturbation to detect dynamic leakage not caught statically.

Evaluation module. Factors are scored on direction-aligned winsorized IC/RankIC blends weighted 0.70/0.30, combined with rolling out-of-sample diagnostics (five-year train, six-month OOS, three-month step). Normal selection (60th percentile threshold, RankIC floor 0.005, positive-ratio floor 0.50) maintains the parent pool; elite selection (80th percentile, RankIC floor 0.01, positive-ratio floor 0.55, plus return-shape gates) preserves stronger mechanisms in an archive; cycle-final library admission requires a train–validation score above 0.75. An AST-based complexity discount penalizes over-engineered implementations for normal ranking only.

Cross Brain. It performs Factor-to-Archetype Attribution (FAA), re-labeling elite factors by their final implemented mechanism rather than their source archetype—an important safeguard against feedback contamination when repair, mutation, crossover, or refinement changes a factor's dominant logic. It then constructs GOOD/BAD feedback records that are mechanism-level rather than score-only, and writes them to archetype-, generation-, and cycle-level memories that steer future routing and novelty injection.

Experimental results

Experiments use CSI300 daily data from Qlib with a 10-day adjusted open-to-open return target, chronological splits (train 2011–2020, validation 2021, test 2022–2025), gpt-oss-120b as the LLM backend, and Ridge regression for portfolio reporting under a top-50/drop-5 strategy with transaction costs.

Model	IC	RankIC	ICIR	RankICIR	AR	AER	IR
LightGBM	0.0190	0.0330	0.2039	0.2674	0.0508	0.0267	0.3865
CNN	0.0253	0.0321	0.2023	0.2485	0.0781	0.0423	0.5430
AutoAgent	0.0308	0.0444	0.1997	0.2896	0.0650	0.0314	0.3628
RD-Agent(Q)	0.0256	0.0291	0.1855	0.2061	0.0731	0.0416	0.4689
CogAlpha	0.0382	0.0590	0.2040	0.2937	0.0639	0.0381	0.3743
XAlpha	0.0413	0.0489	0.2573	0.3412	0.1792	0.1412	1.5853

XAlpha achieves the best IC, ICIR, RankICIR, AR, AER, and IR; CogAlpha retains the highest RankIC. The portfolio gap is large: XAlpha's IR of 1.5853 is roughly 2.7× the next-best value (MLP's 0.5965), and its annualized excess return of 14.12% more than triples the strongest baseline. The authors interpret this as evidence that the closed-loop process improves not only factor quality but also the practical usefulness of the discovered library under held-out evaluation. A caveat worth noting: the comparison table does not include statistical significance tests or multiple-seed variance estimates, so the robustness of the margin itself is not directly quantified.

Component analysis and diversity

Qualitative traces illustrate two mechanisms. In RMA analysis, an accounting-ratio fragment is correctly dropped as OHLCV-infeasible while a return-decomposition fragment is retained, classified to B1 trend-and-momentum, and routed to distinct archetypes separating initiation momentum, mature continuation, and momentum failure. In memory-routing analysis, GOOD/BAD cycle feedback converts prior outcomes into a concrete information gap ("continuous pressure-to-response mapping under regimes and liquidity stress"), producing the next theme regime_lagged_directional_pressure_response with B14 lagged response as primary layer—a controlled change of search space rather than simple factor reuse.

Elite-factor correlation analysis supports genuine diversification: within a single theme (50 factors), mean absolute pairwise correlation is 0.234 with 68.0% of pairs below 0.3; across six themes (60 factors), these fall to 0.142 and 87.1%, respectively. A representative elite factor—regime-dependent reversal after downside overshoot with volatility-conditioned decay speed—is presented with full code and achieves test-split IC 0.0440 and RankIC 0.0634, demonstrating that code-based discovery can yield mechanically readable signals.

Computational cost

Factor generation averages about 15 seconds per factor, one generation about 16 minutes, and one full mining cycle about 3 hours, on two H100 GPUs. Because evolution uses a locally deployed gpt-oss-120b, no external API cost is incurred. These figures make multi-cycle experimentation feasible but indicate that full runs remain nontrivial in compute.

Limitations and open questions

The paper concedes several boundaries. Evaluation is confined to CSI300 with a single 10-day horizon and daily OHLCV inputs; extension to broader universes, other horizons, richer data sources, and live trading is deferred to future work. Portfolio backtesting is framed explicitly as a held-out diagnostic rather than a full portfolio-construction claim. The A-layer screening discards any mechanism requiring fundamentals, order-book, sentiment, or macro data, which bounds the hypothesis space by construction. Several design choices—the fixed eight-agent bundle, the 0.70/0.30 metric blend, gate thresholds, and the complexity-discount parameters—are set empirically without ablations isolating their individual contributions; in particular, no experiment decomposes how much of the performance gain comes from RMA versus memory-driven routing versus tri-alignment validation. Whether the reported margins persist across seeds, LLM backends, and market regimes remains an open question the paper does not address.

Conclusion

XAlpha formulates alpha discovery as a closed-loop, memory-driven research process integrating report-grounded knowledge absorption, macro-level planning, ex-ante validated code-based factor evolution, and mechanism-level feedback consolidation. On CSI300 it delivers the best overall predictive and held-out portfolio metrics among representative baselines, with a substantial IR advantage, while producing diverse and interpretable elite factors. Its principal contribution is architectural—demonstrating that organizing external knowledge, hypothesis planning, implementation, validation, and reflection into one continuous loop outperforms automating isolated steps—though the generality of this result beyond a single market, horizon, and data contract awaits further evidence.