Hubble: An LLM-Driven Agentic Framework for Safe, Diverse, and Reproducible Alpha Factor Discovery

Abstract: Automated alpha discovery is difficult because the search space of formulaic factors is combinatorial, the signal-to-noise ratio in daily equity data is low, and unconstrained program generation is operationally unsafe. We present Hubble, an agentic factor mining framework that combines LLMs with a domain-specific operator language, an abstract syntax tree (AST) execution sandbox, a dual-channel retrieval-augmented generation (RAG) module, and a family-aware selection mechanism. Instead of treating the LLM as an unconstrained code generator, Hubble restricts generation to interpretable operator trees, evaluates every candidate through a deterministic cross-sectional pipeline, and feeds back both top formulas and structured family-level diagnostics to subsequent rounds. The current system additionally introduces positive/negative RAG, formula-similarity penalties, standardized multi-metric scoring, dual reporting of RankIC and Pearson IC, and persistent diagnostics artifacts for post-hoc research analysis. On a U.S. equity universe of roughly 500 stocks, our main run evaluates 104 valid candidates across three rounds with zero runtime crashes and discovers a top set dominated by range, volatility, and trend families rather than crowded volume-only motifs. We then fix the resulting top-5 factors and validate them on a held-out period from 2025-06-01 to 2026-03-13. In this out-of-sample window, the two range factors and two volatility factors remain positive and several achieve HAC-significant Pearson IC and long-short evidence, whereas the weakest in-sample trend factor decays materially. These results suggest that safe LLM-guided search can be upgraded from a syntax-compliant generator into a reproducible alpha-research workflow that jointly optimizes validity, diversity, interpretability, and family-level generalization.

• The paper introduces Hubble, an LLM-driven agentic framework that combines a domain-specific language, AST validation, dual-channel retrieval, and deterministic evaluation to make alpha discovery safe and reproducible, completing 104 evaluations without parser or runtime crashes.
• The framework’s family-aware scoring reduced concentration in liquidity-volume signals and selected range, volatility, and trend factors; four of five frozen factors showed meaningful out-of-sample evidence across a 195-day holdout, although the trend factor decayed.
• The results remain preliminary because the study uses one unneutralized temporal split, omits transaction costs and key ablations, and may be affected by regime exposure, temporal leakage, and persuasive LLM-generated narratives.

Overview and motivation

Hubble is an agentic framework for automated alpha factor discovery that positions the LLM as a constrained hypothesis engine inside a deterministic evaluation loop, rather than as a free-form code generator (2604.09601). The authors motivate the design by three persistent difficulties in formulaic alpha mining: the combinatorial size of the symbolic search space, the low signal-to-noise ratio of daily equity data, and the operational hazards of unconstrained program generation. Prior LLM-based mining systems such as Alpha-GPT, AlphaAgent, and FactorMiner emphasize generation quality but give comparatively little attention to execution safety, reproducibility, or explicit control over factor crowding (Wang et al., 2023, Tang et al., 24 Feb 2025, Wang et al., 16 Feb 2026). Hubble's stated contribution is to address these jointly.

The framework rests on four design principles: safe generation via a domain-specific language (DSL) with AST validation; structured retrieval through dual-channel retrieval-augmented generation (RAG); family-aware selection that penalizes crowding, similarity, and family concentration; and research-ready outputs that archive RankIC, Pearson IC, bucket returns, long-short spreads, turnover, coverage, and complexity diagnostics for every evaluated candidate.

System architecture

The pipeline is a closed loop of five components. The generator produces candidate formulas from an LLM under a curated operator registry (arithmetic, time-series such as TS_SMA/TS_STD, cross-sectional operators such as CS_RANK/CS_ZSCORE, and logical operators) over raw daily panel fields including OPEN, HIGH, LOW, CLOSE, VOLUME, and VWAP. The parser sandbox validates each candidate in three layers: structural security via an AST node whitelist, complexity control on expression depth and node count, and semantic validity of operator names, arities, and variables. This design is "exec-free" by construction—arbitrary code execution is impossible—and the main run reports zero parser or runtime crashes across 104 evaluations.

The dual-channel RAG module is the paper's most distinctive retrieval design. Positive RAG retrieves representative formulas to encourage exploration of under-covered mechanisms; negative RAG retrieves crowded templates presented explicitly as avoid-like references. The authors argue this converts the reference corpus into an active control signal for exploration rather than passive few-shot exemplars—a materially different use of retrieval than standard RAG practice (Lewis et al., 2020).

The evaluation engine computes both Spearman-based RankIC and Pearson IC per date, along with annualized information ratios for each, quantile bucket returns, top-minus-bottom long-short spreads, turnover (defined on the top-decile set as Jaccard distance between consecutive dates), coverage, and AST-derived complexity measures. Significance is assessed with Bartlett-kernel HAC variance estimation following Newey–West [(0703.0002)-style econometric standard]. The scoring layer first standardizes each metric through a fixed $\tanh$ center-and-scale transform, then applies additive penalties for crowding proximity, similarity to already-selected formulas, and family concentration, plus a small novelty bonus. Notably, the penalty magnitudes are pre-scaled at metadata construction time, so the $\lambda$ decomposition is conceptual rather than explicit runtime multipliers—an implementation detail the authors disclose plainly.

Empirical results

In-sample discovery

The discovery run uses 501 S&P 500 stocks over 840 trading days (2022-01-01 to 2025-05-31), three rounds, batch size 20, and the OpenRouter-backed Nemotron-120B model. Across three rounds the system processed 60 primary candidates and completed 104 full evaluations with zero crashes; the only errors were 16 duplicate detections. Best composite score peaked in Round 2 (5.622) rather than Round 1, which the authors interpret as evidence that feedback remains productive after initial exploration.

Factor	Family	Score	RankIC	IC	LS Ret.	Turnover
Range-1	range	5.622	0.0059	0.0127	0.000810	0.113
Range-2	range	5.486	0.0057	0.0130	0.000724	0.161
Volatility-1	volatility	5.475	0.0057	0.0133	0.000546	0.174
Volatility-2	volatility	5.272	0.0047	0.0124	0.000587	0.137
Trend-1	price trend	4.609	0.0115	0.0110	0.000350	0.368

The headline qualitative claim is that the final top set spans range, volatility, and trend families rather than collapsing into crowded liquidity-volume motifs—which earlier versions of the pipeline did despite generating diverse candidate pools. This is the paper's central evidence that family-aware scoring transmits diversity from the candidate pool into the winners.

Out-of-sample validation

The five factors were frozen before a held-out window of 195 trading days (2025-06-01 to 2026-03-13) with no reranking or regeneration. The results are family-specific:

Factor	OOS RankIC	OOS IC	IC $t$ (HAC)	LS $t$ (HAC)
Range-1	0.0242	0.0391	2.98**	2.17*
Range-2	0.0245	0.0385	3.01**	2.31*
Volatility-1	0.0275	0.0131	1.68	2.60**
Volatility-2	0.0232	0.0204	2.47*	2.27*
Trend-1	0.0083	−0.0034	−0.39	−0.87

Both range factors strengthen materially out-of-sample, with Pearson IC $t$-statistics near 3 and significant long-short spreads. Volatility factors retain significance on rank or portfolio metrics. Trend-1—the lowest-ranked factor in-sample—is the only one that clearly decays, reversing sign on Pearson IC. The authors draw two implications here: first, that the in-sample composite ranking is not arbitrary, since the last-ranked factor also generalizes worst; second, that generalization appears to operate at the family level rather than uniformly across all winners. They are careful to caveat that the OOS window may simply favor range- and volatility-sensitive signals, framing the result as preliminary holdout evidence rather than unconditional alpha claims.

Backend robustness

A second run substituting Hunter-Alpha as the backend yields 107 evaluations, 15 errors, a slightly lower peak score (5.460), but a qualitatively identical family profile. The authors' claim is narrower than full robustness: the design does not depend on a single backend to escape the liquidity-volume collapse. They concede this check is lightweight and that explicit ablations (RAG on/off, standardized scoring on/off) are absent.

Limitations

The paper's limitations section is unusually candid and deserves emphasis. The most consequential concession concerns neutralization: no market, sector, or style neutralization is applied, so reported ICs may embed systematic risk exposures. The authors specifically flag this as a likely contributor to the elevated OOS IC magnitudes for range and volatility factors during a period of elevated broad-market volatility in 2025–2026. This materially tempers the strongest numerical results in the paper.

Additional limitations include: a single temporal split rather than walk-forward retraining; a short 195-day OOS window exposed to regime risk, consistent with the broader literature showing post-discovery attenuation of published predictors [(Davoudiasl et al., 2023)-analogous citations to McLean–Pontiff and Hou–Xue–Zhang]; residual thematic bias in the curated RAG corpus; absence of transaction-cost-aware backtesting; and two LLM-specific risks the authors articulate carefully. First, temporal leakage: the DSL prevents encoding historical return series but cannot prevent the model from having absorbed documented findings about factor effectiveness published after the IS window begins, since pretraining corpora necessarily postdate it. A rigorous temporal firewall would require a generator trained strictly on pre-discovery text, which is not feasible with public models. Second, narrative bias: because LLMs produce fluent, coherent rationales, every discovered factor arrives with a ready-made financial story, and a spurious signal paired with a plausible mechanism is structurally harder to reject. The authors explicitly instruct practitioners to treat generated hypotheses as prompts for scrutiny, not corroboration.

Conclusion

Hubble demonstrates that LLM-guided alpha search can be upgraded from a syntax-compliant generator into a reproducible research workflow in which deterministic components adjudicate validity while diversity constraints shape what gets discovered. Its strongest empirical claims—that family-aware scoring changed the identity of winning factors and that four of five winners survive holdout evaluation—are supported, but rest on a single unneutralized temporal split under conditions favorable to the discovered families. The open questions the paper leaves are concrete: whether the family-level generalization pattern persists under neutralization and walk-forward protocols, whether explicit ablations isolate the contribution of negative RAG versus standardized scoring, and how the framework performs once transaction costs are priced in.