LineageEvo offers a well-designed way to spend an LLM factor-search budget, and I would copy its lineage-balanced parent sampler. Its contribution to the S&P 500 book's 27.00% annualized return (AR) remains unproven. The experiments that isolate the controller's mechanisms all use CSI 500. None of those ablation rows includes repeated runs or a spread.

The search built by Liu, Zhang, Ma and Yang

The system looks for formulas whose cross-sectional ranks predict close-to-close returns from day t+1 to t+2. It writes those formulas in a DSL using daily price and volume data. The LLM, deepseek-v4-flash at temperature 0.7, either mutates a parent formula or crosses two parents. The search controller is the claim: it should extract better formulas from the same data.

Lineage supplies the organizing unit. Older genetic-programming work already tracked it; the new use here is to allocate parents and scope the LLM's prompt notes by lineage. Each valid child enters a family tree. A lineage contains every descendant that inherits the expression backbone of one of 10 LLM-generated seed formulas. The authors put these families to work in two ways.

Lineage-Stratified Parent Selection (LSPS) chooses a family, then a parent within it. The family receives a weight of 0.8 times normalized best quality plus 0.2 times normalized 1/(family size + 1). This keeps small families in the draw. Within the chosen family, sampling is proportional to q(f) = 0.2|train IC| + 0.8|validation IC| minus 0.1 times the train/validation gap.

The other use is family-specific "priors", JSON notes recording which edits helped and which hurt. An LLM revises the notes after a child clears an improvement threshold or falls short by a deterioration threshold. On CSI 500, those thresholds are max(0.003, 30% of parent IC) and max(0.012, 60% of parent IC), respectively. The next prompt combines the family notes with global notes. In the paper's worked example, replacing TsSkew with a close/high-return TsCov lifts validation |IC| from 0.0209 to 0.0369. A sibling branch follows the revised note, makes the same substitution, and rises from 0.0150 to 0.0323.

A run ends after 200 valid factors. A 20-slot elite archive supplies the top 20 by q(f). Validation data determines their signs; each day's values are normalized cross-sectionally and then equal-weighted. Qlib's TopkDropout trades the result with top 50, drop 5, daily rebalancing, and costs of 0.15% on buys and 0.25% on sells. The training sample covers 2015 through 2020. Validation spans January 2021 to April 2022, followed by a test from May 2022 to April 2026.

What does the S&P 500 lead amount to?

The abstract makes prediction the basis of the claim, saying LineageEvo delivers the strongest predictive performance in both markets. The introduction goes further, claiming the strongest portfolio performance in both. The conclusion also credits the method with strong investment returns. Returns therefore carry part of the paper's argument. For S&P 500, the best baseline differs by metric:

IC and portfolio returns tell different stories. GP earns 21.38% with an IR of 1.0958, despite an IC of 0.0055, less than half LineageEvo's. AlphaPROBE ranks second on IC at 0.0119, yet returns 5.86%. Such different rankings leave the return table saying little about factor quality. The predictive margins carry the evidence: 0.0015 in IC and 0.0239 in ICIR. Those appear to come from a single run of a sampler at temperature 0.7.

CSI 500 IC falls from 0.0512 to about 0.038 without lineage priors or LSPS

Attribution belongs in the CSI 500 ablations. The complete system records IC of 0.0512. It falls to 0.0384 without any priors and to 0.0390 when only lineage priors are removed. Keeping all priors while removing LSPS produces 0.0381. Dropping global priors makes little difference, leaving 0.0495.

The individual losses deserve attention. Removing lineage priors costs 0.0122 of IC; removing LSPS costs 0.0131. Removing every prior costs 0.0128. Add the losses from the individual removals and they would be about twice the full system's lead. Strong complementarity could explain this: each component needs the other. Run-to-run noise of about 0.013 IC could explain it too.

The authors interpret the prior table as evidence that lineage-scoped experience supplies most of the useful guidance. With no repeated runs or spreads for any ablation row, the table cannot distinguish the competing readings. Removing LSPS also leaves CSI 500 IC at 0.0381, below EoH's 0.0399.

The diversity figures raise a separate question. Removing LSPS takes active-pool lineage entropy from 2.154 to 0.405, although the setup limits any single lineage to 0.40 of that pool. That cap implies a minimum entropy of about 1.055 using natural log, with shares of 0.4, 0.4 and 0.2. The log base is unspecified. Base 2 would raise the floor to about 1.52, so 0.405 conflicts with the cap under either convention.

The appendix defines lineage shares across the evaluated population state. The Table 3 discussion instead refers to the active factor pool. Either this variant disabled the cap, or the entropy calculation used that other set. The structure-level result changes less: effective structures fall from 41.90 with LSPS to 37.89 without it. That is a 10% decline, compared with 81% for entropy.

Both ablation tables carry the CSI 500 label; no ablation was run on S&P 500. Yet the abstract's LSPS diversity claim names no market. It describes a more balanced lineage distribution and broader factor pool in the same sentence as the predictive claim covering both markets. Only the CSI 500 table supports that diversity result. S&P 500 also uses different rewrite triggers: max(0.0015, 20%) for improvement and max(0.006, 40%) for deterioration. We did not find an explanation of their selection.

Before trusting the S&P 500 annualized return of 27.00%

We are rebuilding only the US half. We cannot run CSI 500 and will not borrow the paper's reported results for it. Our universe selects the top 500 US stocks by market capitalization for each year. It approximates the S&P 500 without reproducing historical index membership. The paper denotes its universe by the date-indexed U_t, but leaves historical S&P 500 membership unspecified. Survivorship in its US test therefore remains an open question. Our evaluation is still running; we report no results from it here.

Selection deserves scrutiny as well. The final ranking gives 0.8 weight to a 16-month validation window, tuning the selected 20 factors to early 2021 through April 2022. The efficiency claim says performance plateaus within about 75 of 200 valid evaluations. Only valid factors enter that count. A slot permits up to five attempts, with invalid and duplicate attempts excluded. Rewriting priors requires additional LLM calls beyond those attempts.

The appendix also defines the plotted curve as validation IC. The test extends to April 2026, and we did not find discussion of possible overlap with the LLM's pretraining.

I would keep the lineage sampler.

Three seeds on S&P 500, removing LSPS and lineage priors separately, would change my view if the across-seed IC spread were well below 0.0015.