Daily AI Trading Paper

If you’ve ever tried to mine alpha with symbolic formulas, you know the core problem: the search space is enormous, the signals are noisy, and most candidates are weak or redundant. That matters because in quantitative finance, a factor is only useful if it adds predictive information that survives out of sample and improves portfolio construction. A lot of automated alpha mining has focused on generating more candidates, but generation alone is not the same as discovery. The challenge is to search broadly enough to find genuinely new signals, while avoiding endless variants of the same idea. This paper, CosmosAlpha, is about improving that search process for LLM-driven formulaic alpha mining. The setting is familiar to quant researchers: you have market data, you generate explicit mathematical expressions over observable variables, and you evaluate them by how well they rank stocks or support a profitable portfolio. The appeal of formulaic factors is interpretability. You can inspect what the factor is doing, rather than treating it as a black box. But interpretability does not make the search easy. In fact, because these expressions are symbolic, the space of possible combinations is combinatorial, and many factors end up being highly correlated with each other or with the same underlying exposures. The broader landscape has moved from handcrafted libraries, like fixed factor sets, toward automated search. Genetic programming and reinforcement learning have both been used to evolve factor formulas or factor sets. More recently, large language models have entered the picture, bringing semantic reasoning and iterative refinement. Some systems use Monte Carlo Tree Search, some track lineage in a directed acyclic graph, and others evolve complete trajectories or use agentic reinforcement learning. The common theme is that they remember some search history, so they can reuse prior evaluations instead of starting from scratch each time. CosmosAlpha argues that this history is still too narrow if it is organized mainly around parent-child provenance. Knowing who generated whom is useful, but it misses two other important kinds of similarity. First, two factors may share subexpressions even if they come from different branches of the search tree. Second, two factors may behave almost identically in the market, producing very similar cross-sectional exposures across stocks. If you ignore those relationships, exploration can get stuck in redundant regions of the search space. The key idea in CosmosAlpha is to treat the entire evaluated factor pool as a graph, not just a tree or a trajectory. Each node is a complete factor. The graph has three kinds of relations. One is generation lineage: if one factor was mutated from another, that parent-child link is recorded. A second is expression-structure similarity: factors are compared by their canonical subexpressions, so formulas that are structurally close can be linked even if they came from different search branches. A third is empirical exposure correlation: if two factors produce similar cross-sectional stock exposure vectors over time, they are also connected, because they are likely to be redundant from a portfolio-construction perspective. This is an important design choice. Structure similarity tells you whether two formulas look alike. Exposure correlation tells you whether they behave alike. Lineage tells you how they were discovered. CosmosAlpha combines all three, so the search state is not just a record of ancestry; it becomes a relational memory of what has already been tried and how those trials relate to one another. On top of that graph, the authors introduce what they call Graph Flow. The basic idea is to propagate utility information through the neighborhood of each factor. So instead of scoring a candidate parent only by its own current evaluation, the method also uses evidence from nearby factors in the graph. That can help identify promising regions of the search space and reduce overcommitment to isolated candidates. In spirit, this is similar to how graph-based information sharing works in other domains: local neighborhoods can provide context that a single node alone does not reveal. But CosmosAlpha does not rely on one score. It uses Pareto-guided selection over three criteria: quality, fertility, and diversity. Quality is straightforward: how predictive is the factor right now? Fertility asks whether this factor has historically produced strong descendants when mutated. Diversity asks whether the factor sits in a less redundant part of the pool. This is a useful distinction because a factor can be mediocre itself but still be a productive parent, and a high-quality factor can still be a dead end if it keeps generating near-duplicates. Rather than collapsing all of that into one weighted objective, the method keeps the trade-offs explicit and selects non-dominated candidates. The workflow is iterative. The current factor archive is evaluated as a graph. Graph Flow updates utilities over the relations. Pareto-guided selection chooses parents. The LLM then mutates those parents to generate new candidate formulas. Valid children are evaluated and inserted back into the graph. So the archive is continuously updated and becomes the state that drives the next round of search. The experiments are run on two Chinese equity universes: CSI 300 and CSI 500. The paper says the evaluation is done under a common data and downstream LightGBM integration protocol, which means the factors are not just judged abstractly; they are also tested in a standard predictive modeling pipeline. That matters because in practice, a factor’s value is not only its standalone correlation with returns, but also how it contributes inside a downstream model and portfolio process. The authors compare CosmosAlpha-Graph against a controlled CosmosAlpha-MCTS counterpart. That control is important: it helps isolate the effect of the relational graph search itself, rather than confounding it with differences in the factor language, evaluator, LLM backbone, or downstream protocol. They also test across three LLM backbones and three independent runs, which is a good sign that they are checking robustness rather than cherry-picking one lucky configuration. The headline result is that CosmosAlpha-Graph achieves the best overall mean rank across seven predictive and portfolio metrics on both universes. It also outperforms the MCTS-based control on every reported metric. The paper’s claim is not just that one number improved, but that the advantage is consistent across the full metric set. That suggests the graph-based design is helping the search process in a broad and stable way, rather than overfitting to a single evaluation criterion. The ablation studies are also informative. They support the contribution of both relational propagation and multi-objective selection. In other words, the gains are not coming from just one trick. The graph matters because it lets the system reuse information across related factors, and the Pareto selection matters because it preserves useful trade-offs among quality, fertility, and diversity. Taken together, those two pieces appear to be what makes the search more effective than a purely lineage-based approach. What should we make of this as trading researchers? The practical message is that alpha discovery is not just a generation problem; it is a search-and-memory problem. If you only remember ancestry, you may keep walking around in circles, especially when formulas differ syntactically but are economically similar. A multi-relational graph is a natural way to represent both symbolic and behavioral redundancy. And once you do that, selection can be more informed: you are not merely asking “what is best right now?” but also “what is likely to lead somewhere new?” There are still limitations to keep in mind. First, the paper is about factor discovery and offline evaluation, not live trading. So the results show improved search quality and better downstream metrics in the tested setting, but not necessarily production robustness under regime shifts, costs, slippage, or capacity constraints beyond what is encoded in the evaluation. Second, the graph relations depend on the chosen factor language, canonicalization rules, and correlation estimates, so the method is only as good as those design choices. Third, while multi-objective selection is appealing, it still leaves open the question of how to best balance quality, fertility, and diversity in different markets or under different investment mandates. Even so, CosmosAlpha is a strong example of where LLM-driven quant research is heading. The interesting shift is from using an LLM as a formula generator to using it inside a structured search system that remembers what it has learned about the factor space. That combination of symbolic generation, relational memory, and Pareto-aware selection is the real contribution here. It does not guarantee alpha, of course, but it does make the search for alpha more intelligent, less redundant, and better aligned with the realities of quantitative factor discovery.