CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining

summary

The gist

Discovering predictive alpha factors from noisy financial data is a fundamental yet challenging problem in quantitative finance The paper introduces CosmosAlpha…

In short

CosmosAlpha is an LLM-driven framework for finding predictive alpha factors from noisy financial data. It moves beyond simple generation history by representing factors as a multi-relational graph combining lineage, structure, and correlation. Using Graph Flow and Pareto-guided selection, the system assesses factor quality, fertility, and diversity simultaneously to achieve superior performance across multiple metrics compared to traditional search methods.

Key concepts

Multi-Relational Graph
This framework organizes discovered factors as a graph where nodes are factors. The edges represent three distinct relationships: generation lineage (how one factor was created from another), expression structure similarity (how similar the mathematical formulas are), and cross-sectional exposure correlation (how related the factors' performance is across different markets).
Graph Flow Mechanism
This mechanism propagates search utilities—quality, fertility, and diversity scores—across the entire graph. Instead of evaluating a factor in isolation, its utility is informed by relational evidence from its neighbors. This allows each factor to be assessed based on the collective knowledge of related factors during the search process.
Pareto-Guided Selection
This selection method guides expansion by considering three objectives—quality (predictive strength), fertility (ability to produce good descendants), and diversity (redundancy). Instead of picking a single best factor, it uses nondominated sorting to select factors that represent the best possible trade-offs between these competing goals.
Node Utilities
Each factor node is assigned a three-dimensional vector representing its utility: intrinsic quality (predictive strength), fertility (how productive it is), and diversity (its uniqueness relative to existing factors). These utilities are propagated through the graph using Graph Flow to guide the search toward factors that balance these three dimensions effectively.

Terminology used across episodes

This episode discusses

The paper

CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining · Read on arXiv

Zhan-He Wang, Zi-Yang Lang, Yi-Yang Li, Qing-Yu Lan, Yun-Song Piao

Artificial Intelligence and Quantitative Finance Laboratory, University of Chinese Academy of Sciences · School of Physics and Astronomy, Shanghai Jiao Tong University · Hangzhou Institute for Advanced Study, UCAS

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: We're looking at the paper: "CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining".

Warren: Here's the gist of the paper. Discovering predictive alpha factors from noisy financial data is a fundamental yet challenging problem in quantitative finance The paper introduces CosmosAlpha, an LLM-driven framework that organizes evaluated factors as a multirelational search state This framework combines generation lineage, expression structure, and cross-sectional exposure correlation to guide factor mining through Graph Flow and Pareto-guided selection.

Ruth: Pruning redundant reasoning tokens The existing generation-centric and tree-based search strategies primarily organize factors through generation lineage, leaving structural and statistical relationships across search branches underutilized Information remains localized along parent–child trajectories, which can confine exploration to redundant regions CosmosAlpha represents discovered factors as a multi-relational graph encoding generation lineage, expression-structure similarity, and cross-sectional exposure correlation A Graph Flow mechanism propagates search utilities across the graph, allowing each factor to be assessed using relational evidence from nearby factors rather than in isolation Pareto-guided selection jointly considers quality, fertility and diversity, explicitly modeling their trade-offs within a multi-objective search framework Across CSI 300 and CSI 500, CosmosAlpha-Graph achieves the best mean rank over seven predictive and portfolio metrics and outperforms its controlled CosmosAlpha-MCTS counterpart on every reported metric How the training signal is built The search state is modeled as a multi-relational graph G = (V, E gen, E str, E corr), where nodes are evaluated factors The three relations capture complementary notions of factor proximity, including generation relation i → j for provenance, structure relation dstr ij based on canonical subexpression similarity, and correlation relation dcorr ij based on cross-sectional Pearson correlation Cij Node utilities are defined by a three-dimensional vector Ri = (Q self i, Qfert i, Qdiv i), where intrinsic quality qi measures predictive strength, fertility measures whether previous expansions produced high-quality descendants, and diversity measures redundancy relative to a reference set Rr These utilities are propagated through the graph using Graph Flow F(k+1) = (1 − γ)R + γP F(k), where P is the combined transition matrix of generation, structure, and correlation links Where the gains come from The key innovation lies in combining relational evidence with distinct quality, fertility, and diversity criteria through Graph Flow and Pareto-guided selection The search utility propagation allows parent assessment to reflect both local measurements and evidence from related factors Parent selection operates on propagated vectors Fi = (F self i, Ffert i, Fdiv i) by applying nondominated sorting rather than a fixed scalarization of quality, fertility, and diversity This preserves non-dominated trade-offs without collapsing these criteria to a single scalar The results show that CosmosAlpha-Graph achieves the best overall mean rank on both CSI 300 and CSI 500 and outperforms the controlled CosmosAlpha-MCTS counterpart on every reported predictive and portfolio metric Evaluation Protocol The evaluation protocol involves recomputing each method’s factor pool under a common data protocol, using the factors jointly as LightGBM features Predictive performance is measured from held-out LightGBM predictions using IC, ICIR, RankIC, and RankICIR Portfolio performance is measured from the same signal using annualized return (ARR), information ratio (IR), and signed maximum drawdown (MDD) The final MeanRankm summarizes the four predictive and three portfolio metrics for each method CosmosAlpha-Graph achieves the best mean rank on both CSI 300 and CSI 500 under a common data and downstream LightGBM integration protocol The results indicate that the graph variant’s advantage also extends to CSI 500, although Alpha158 retains an advantage on RankICIR LLM Sensitivity Study Across all three backbone models, Graph delivers the strongest overall performance and leads all seven metrics for each backbone The portfolio results particularly favor DeepSeek-V4-Flash in the Graph setting, leading ARR, IR, and signed MDD for Graph The mean IC, ICIR, RankIC, and RankICIR are 0.0326, 0.3467, 0.0501, and 0.5699 for CosmosAlpha-Graph on CSI 300 The same ordering holds for the portfolio metrics: Graph reaches a mean ARR of 8.45%, IR of 0.5398, and signed MDD of −25.10% Across each search mode, the three backbone models remain relatively close on Predictive Performance, whereas the differences between Graph, MCTS, and Chain are more pronounced CosmosAlpha-Graph has the highest mean in all seven metrics when using the three independent runs The mean IC, ICIR, RankIC, and RankICIR are 0.0326, 0.3467, 0.0501, and 0.5699 for CosmosAlpha-Graph on CSI 300 under the three independent runs The same ordering holds for the portfolio metrics: Graph reaches a mean ARR of 8.45%, IR of 0.5398, and signed MDD of −25.10% Ablation Study Insights Every ablation in Table 3 falls below the full model on all seven metrics, supporting the joint use of the quality archive, Flow-level parent selection, all three utility objectives, and all three graph relations The Flow scalarization variant retains the intrinsic-quality Pareto archive but replaces Pareto front and crowding-based parent ranking with an equal-weight score over the propagated utilities (Q, F, D) Removing Structure or Correlation excludes that relation from both Graph Flow and AverageDistance The remaining Flow relation weights are renormalized The results confirm that relational propagation, Pareto-guided selection, utility composition, and the three graph relations are all necessary for optimal performance Factor Language Specification CosmosAlpha restricts candidate generation to a fixed domain-specific language (DSL) defined by a versioned field and operator catalog The DSL includes nine daily market fields as terminals, such as open Oi,t and close Pi,t The operator catalog contains 26 canonical names, including non-windowed operators like add(x,y) and time-series operators like corr(x,y,w) The expression space is constrained by rules such as a maximum AST depth of 15 and an operator count cap of at most 32 call nodes This ensures that the LLM generates expressions within a syntactically valid search space The final factor expression system message instructs the agent to translate one supplied FactorDesign into one CosmosAlpha DSL expression, returning only exactly what is requested The paper provides concrete examples of factors, such as E1 which uses the expression decay linear(div(sub(close,vwap),sub(high,low)),5) The catalog specifies that variables like industry id gi,t are consumed internally by indneutralize and cannot be inserted directly into a generated expression The DSL field set is strictly defined to include only admissible fields from Table 13 The operator catalog details the numerical behavior of each operator, specifying rules for missing inputs and window reducers The syntax normalizer produces deterministic lowercase prefix text but does not perform algebraic simplification or reorder operands, while Graph identity applies a narrower commutative rule to ensure structural consistency A node’s final undirected degree can therefore exceed 30 after combining the three modalities CosmosAlpha introduces an LLM-driven framework that organizes evaluated factors as a multirelational search state connecting generation lineage, expression structure, and cross-sectional exposure correlation It uses Graph Flow to propagate quality, fertility, and diversity through these relations and Pareto-guided selection to guide expansion beyond local generation history CosmosAlpha-Graph achieves the best mean rank on both CSI 300 and CSI 500 and outperforms the controlled CosmosAlpha-MCTS variant on every reported metric The framework demonstrates that relational graph search is superior for LLM-driven factor mining The source code is available at https://github.com/wzhmtldswz/CosmosAlpha References The paper cites numerous prior works in formulaic alpha mining, evolutionary search, and LLM-driven agentic systems Specific related methods include AlphaPROBE which combines retrieval-guided factor proposal with graph-biased evolution, and AlphaSAGE which uses GFlowNets to encourage diverse exploration The paper also compares CosmosAlpha against various LLM backbones including DeepSeek-V4-Flash, GPT-5.6-Sol, and Kimi K3 across different search modes The stability study confirms that CosmosAlpha-Graph has the highest mean in all seven metrics when using the three independent runs Future work will apply CosmosAlpha to additional markets and to the discovery of high-frequency and fundamental factors The paper uses historical market data and does not involve human participants, personal data, or interaction with individual investors This work is an offline research evaluation, and its backtest results should not be interpreted as financial advice or guaranteed live-trading performance The reproducibility statement specifies the rolling-operator lookback window and temporal splits for all experiments The authors acknowledge support from various national science foundations This work is supported by the National Natural Science Foundation of China (Grant No.

Warren: 12475064), the National Key Research and Development Program of China (Grant No.

Ruth: 2021YFC2203004), and the Fundamental Research Funds for the Central Universities The AI use statement notes that Generative AI tools were used primarily for English-language editing, grammar correction, and local improvements in clarity and organization The ethics statement confirms that this work uses historical market data and does not involve human participants, personal data, or interaction with individual investors The reproducibility statement specifies the rolling-operator lookback window and temporal splits for all experiments The released artifacts will identify the exact code revision used for each experiment Page 11 --- The paper cites Emmanuel Bengio et al.

Paper summary: Ruth: So we're looking at this paper now called CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining. It tackles this really fundamental problem in quantitative finance, which is finding predictive alpha factors from messy financial data.

Warren: Right. The core idea here is that using a large language model to find these formulaic alpha factors is attractive because you get interpretable expressions, but the search methods we have now are mostly just focused on generation lineage.

Iris: That's what the abstract points out, Warren. Existing methods organize things by how they were generated, which means structural and statistical relationships across different search branches are left underutilized.

Ruth: Exactly. So what CosmosAlpha claims is that instead of just looking at where a factor came from in the generation process, it uses a graph structure to connect factors based on three different kinds of relationships.

Warren: It encodes generation lineage, expression structure similarity, and cross-sectional exposure correlation into this multi-relational graph. That means when you look at one factor, you can instantly see its structural neighbors and how it correlates with other factors across the market.

Iris: The framework uses something called a Graph Flow mechanism to spread search utilities across this graph so that each factor gets assessed using evidence from nearby factors instead of being judged in isolation.

Ruth: And on top of that, they use Pareto-guided selection to choose which parents to expand next by jointly considering quality, fertility, and diversity.

Warren: They are explicitly modeling the trade-offs between those three things—quality, fertility, and diversity—within a multi-objective search framework so the system doesn't just optimize for one thing.

Iris: That’s interesting because quality measures present predictive strength, fertility looks at whether past expansions led to good results, and diversity checks if a factor is in a less redundant part of the pool.

Ruth: So what does this mean for us right now? The paper says that across the CSI three hundred and CSI five hundred datasets, CosmosAlpha-Graph achieves the best mean rank over seven predictive and portfolio metrics.

Warren: It also outperforms its controlled CosmosAlpha-MCTS counterpart on every single reported metric, which suggests that this relational graph search approach gives a consistent edge in factor mining with LLMs.

Paper summary: Iris: That’s solid evidence showing that when you combine those three specific dimensions—generation, structure, and correlation—with Pareto guidance, you get a better outcome than relying on just one or two of those signals alone.

Ruth: It seems the main implication is that for anyone doing automated alpha mining with LLMs, moving beyond just lineage-based search to a multi-relational graph approach provides a much richer way to explore the space.

Warren: And it suggests that simply generating more factors isn't enough; you have to manage redundancy and structure simultaneously, or you end up exploring the same noisy regions over and over.

Iris: I think the key is how they defined those relations, especially how they measured structural similarity using Jaccard similarity and cross-sectional correlation by averaging daily Pearson correlations.

Ruth: That detail matters because it shows them mapping very concrete financial concepts—like expression structure and market behavior—onto the graph edges.

Warren: And before we go deeper into the specific search mechanics, I want to quickly address how this works in practice for someone running a system. The paper says they restrict candidate generation to a fixed domain-specific language or DSL.

Iris: That’s important because it means the LLM isn't just free-form writing; it has rules about what fields you can use and what operators you can combine, keeping the generated expressions syntactically valid.

Ruth: So, the system enforces a specific vocabulary and syntax before letting the LLM generate something that looks like a factor expression.

Warren: And they specify things like an AST depth limit of fifteen call nodes and an operator count cap of thirty-two call nodes to keep the search space manageable for the model.

Iris: That constraint helps prevent the LLM from just generating incredibly complex, uninterpretable mathematical monstrosities that we wouldn't be able to use for alpha.

Ruth: It seems like they’ve built a very controlled sandbox where you can mine factors systematically, rather than just throwing everything at the wall and hoping something useful comes out.

Warren: It shows they are trying to make the LLM generation process more grounded in what actually constitutes a usable financial expression.

Paper summary: Iris: And as we look at their evaluation protocol, they use a common data protocol and integrate the factors into a LightGBM model for predictive performance measurement using things like IC and RankICIR.

Ruth: That’s smart because it tests the final integrated signal, not just individual factors in isolation, which is how you actually trade or use alpha in production.

Warren: And for portfolio performance, they look at annualized return or ARR, information ratio or IR, and signed maximum drawdown MDD using the same signal.

Iris: That’s a comprehensive way to measure success because it checks both the predictive power of the signals and how those signals behave when you actually try to manage a portfolio.

Ruth: So, in conclusion for CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining, they are proposing this graph framework as superior because it captures more relationships than previous generation-centric methods.

Warren: It’s about using the graph to connect structure and correlation alongside lineage, and then using those connections to guide the search through quality, fertility, and diversity trade-offs.

Iris: The authors argue that this relational propagation allows parent assessment to reflect both local measurements and evidence from related factors in a way that isolates promising regions better than just looking at the immediate parent.

Ruth: It’s essentially saying that exploring based on structural similarity and cross-sectional correlation is a necessary step if you want the LLM mining process to be effective in finding high-quality, diverse factors.

Warren: The paper suggests that this relational graph search approach is the better way to do it when using LLMs for alpha mining compared to older tree-based or generation-only strategies.

Iris: The limitation they mention is that while the graph captures these relations well, they haven't explicitly shown how this framework scales perfectly when moving to even larger market datasets beyond CSI five hundred.

Ruth: It’s a realistic point; the performance is proven on CSI three hundred and five hundred but we still need to see if it holds up under more extreme conditions.

Warren: So, for the listener tuning in, the big picture here is that sophisticated financial discovery with AI needs to move past simple generation and start modeling how different ideas relate to each other across time and structure.

Conclusion: Ruth: So, to wrap up, CosmosAlpha is this new framework that uses a graph structure to connect factors based on generation lineage, expression structure similarity, and cross-sectional exposure correlation to guide the AI factor mining process.

Warren: Right. It’s essentially taking all those different ways factors relate—how they were made, how they look mathematically—and weaving them into one big map.

Iris: And this isn't just a simple map; it uses a mechanism called Graph Flow to spread search utility across the graph so each factor gets assessed using evidence from its neighbors instead of being in isolation.

Ruth: It then uses Pareto-guided selection to choose the next factors to explore by looking at quality, fertility, and diversity all at once.

Warren: That’s what the authors are claiming is a significant step forward because it models those trade-offs between quality and diversity explicitly within a multi-objective search framework.

Iris: They show that this approach achieves the best mean rank across seven predictive and portfolio metrics on both the CSI three hundred and CSI five hundred datasets.

Ruth: It also beats their controlled version, CosmosAlpha MCTS, on every single reported metric, which shows a consistent advantage for this relational graph search method in LLM driven factor mining.

Warren: So what does this mean for the world? It suggests that when you’re using AI to find alpha factors from financial data, just generating more ideas isn't enough; you have to model how those ideas interact structurally and statistically across different market conditions.

Iris: The limitation they mention is that while it performs great on CSI three hundred and five hundred they haven't shown how this framework scales perfectly when you move to much larger market datasets.

Ruth: Exactly. It’s a solid result for the datasets they tested, but we still need to see if this relational graph search holds up under more extreme conditions with bigger data.

Warren: We'll look at those scaling issues next, so keep your eyes peeled for that discussion on how this system handles massive amounts of financial data.

More episodes

← Home