CSI 300 gives a convincing reason to prefer CosmosAlpha's graph search to its own MCTS variant. The gap exceeds the run-to-run noise the authors measured. Its trading returns deserve less weight: the headline mean ranks make those results look easier to interpret than they are.
The factor archive chooses what comes next
Wang, Lang, Li, Lan and Piao build on a familiar loop: an LLM mutates formulas, a backtester evaluates them, and the strongest supply parents for another round. The formulas use a 26-operator language over nine daily fields (open, high, low, close, volume, vwap, turn, free_turn, returns). Windows span 5, 10 or 20 days. The change lies in how the search stores and reuses evaluated factors.
The archive is a graph. Each factor is a node, connected through three kinds of edge. Parent-child links record its ancestry. Structural links join formulas with overlapping canonical subexpressions, requiring Jaccard similarity of at least 0.20. Correlation links join factors whose daily cross-sectional exposures reach |corr| of at least 0.10. For both structure and correlation, a node keeps up to 30 nearest neighbours.
Three scores guide selection. Quality averages normalised |RankIC|, RankIC stability and a direction-aligned Top-20% return. Fertility measures whether a factor produces useful children: it averages the three best gains its children contributed to a Pareto archive of quality. Diversity measures distance from the 10 nearest high-quality references in structure and correlation space. Graph Flow, the authors' name for smoothing scores across neighbours, uses γ=0.75. A formula can therefore receive some credit when a structurally related formula succeeds in a distant branch. NSGA-II sorting selects up to four parents each round without fixing weights for the three scores.
The route to trading performance is indirect. Roughly 150 formulas supply a LightGBM forecast, which drives a long-only TopK-Drop book. Mining and model fitting cover April 2010 to May 2019; validation continues to December 2020. Testing covers 2021 to 2025 on CSI 300 and CSI 500. Across seven metrics, the graph variant has the best mean rank in both markets: 1.43 on CSI 300 and 1.29 on CSI 500.
How much does the MCTS comparison establish?
CosmosAlpha-MCTS (Monte Carlo tree search) is the comparison to watch. It shares the formula language, evaluator, LLM and downstream model. The CSI 300 differences are substantial: IC reaches 0.0324 against 0.0281, and RankIC reaches 0.0494 against 0.0409. ARR is 8.44% against 5.68%, IR is 0.5332 against 0.3296, and drawdown is -25.07% against -31.56%. The twin itself places mid-pack on CSI 300. Its mean rank is 7.14, behind Alpha158 and AlphaForge at 3.86.
The three-run study gives that gap more credibility. Graph averages 8.45% ARR with a population standard deviation of 0.40; its IR averages 0.5398 with 0.0059. MCTS records 5.71% (0.18) and 0.3362 (0.0131). With dispersions below half a point, a 2.7-point ARR difference is difficult to explain as run-to-run variation. The evidence still consists of three runs over one test window.
CSI 500 offers a thinner signal comparison. IC is 0.0360 against 0.0356, and ICIR is 0.3997 against 0.3898. ARR reaches 13.01% against 12.16%, with drawdown of -26.42% against -32.18%. The graph leads across all seven metrics, though an extra run could erase the small IC and ICIR advantages. Its drawdown advantage is 5.76 points, exceeding the 3.65 spread across MCTS's three CSI 300 runs. IR of 0.6689 against 0.6258 differs by about three times MCTS's CSI 300 IR spread of 0.0131. Those comparisons cannot tell us how much CSI 500 varies between runs: the stability study covers CSI 300 alone.
Attributing the result cleanly to topology runs into two problems. The modes receive different design prompts. MCTS directs each refinement toward one target dimension; the graph prompt supplies the parent's utilities, archive gaps and correlation-neighbourhood constraints. Backbone choice also matters about as much as search choice. Graph ARR reaches 8.44% with DeepSeek-V4-Flash, 7.51% with GPT-5.6-Sol and 6.11% with Kimi K3. Kimi K3 ran at temperature 1.0, versus 0.0 for the others. The stronger defence is that Graph beats MCTS and Chain across all seven metrics with every backbone.
Ablations support the same interpretation, within the limits of single runs on CSI 300. Removing the structure relation reduces ARR to 4.37%; setting γ=0 reduces it to 5.90%.
Alpha158 comes close on IC
The handcrafted Alpha158 library nearly matches the graph's CSI 300 signal results. IC is 0.0320 against 0.0324, while RankICIR is 0.5668 against 0.5670. Alpha158 has the higher ICIR, at 0.3474 against 0.3441. It also wins RankICIR on CSI 500, with 0.7898 against 0.7562.
Most of the graph's advantage over this fixed library consequently appears in the portfolio columns. CSI 300 ARR is 8.44% against 6.32%. Those columns need a narrower interpretation.
Reading the 8.44% CSI 300 return
The returns are absolute and measured against a zero benchmark. No risk-free rate is subtracted. ARR multiplies mean daily return by 238, while drawdown is calculated on arithmetic cumulative returns. The portfolio holds 50 names and replaces up to 5 each day. Execution occurs at the next day's open, avoiding look-ahead in the timing. We did not find commissions, stamp duty or slippage in the backtest configuration. Replacing up to 5 of 50 holdings each day permits daily turnover of up to 10% of the book.
Every ARR in the table therefore includes whatever the index returned over 2021 to 2025. Portfolio performance also tracks signal quality loosely. GP has the lowest CSI 500 IC, 0.0213, yet its 12.52% ARR exceeds Alpha158's 11.07%.
IC provides the cleaner evidence.
The authors caution against treating their backtests as financial advice or guaranteed live-trading performance. Their reporting warrants that caution. As far as we read it, the paper also leaves open whether backbone pretraining overlaps the 2021 to 2025 test window. Search receives feedback only from the training period. The models proposing hypotheses may nevertheless have read about the test period.
An architecture worth investigating
Each automated method gets about 150 evaluable factors. This matches output count. The paper does not equate an MCTS iteration with a Chain step or a Graph expansion attempt, so it does not match LLM calls. A graph run with 100-attempt used 6.17 hours of cumulative DeepSeek call time and 3.98M tokens. The three relations have nothing specific to A-shares in their construction, although the paper's tests cover only CSI 300 and CSI 500.
We could not rerun the result. We lack A-share price and volume histories and point-in-time CSI 300 and CSI 500 constituent lists. A US version would constitute a separate experiment.
The search finding warrants action. To move us on the trading claim, the same CSI 300 comparison would need to report excess return over the index, net of costs, alongside the three-run spread.