AlphaSAGE: Automated Alpha Mining Framework

• AlphaSAGE is a framework for automated alpha mining in quantitative finance that formulates the search for symbolic expressions as a combinatorial exploration over ASTs.
• It employs a structure-aware RGCN encoder and a GFlowNet objective to overcome reward sparsity, opaque sequential representations, and mode-seeking biases in traditional RL.
• Experimental results on CSI300, CSI500, and S&P500 indicate that AlphaSAGE enhances performance metrics like IC, AR, and SR, enabling the generation of diversified and interpretable trading signals.

Searching arXiv for the cited papers to ground the article and verify titles/ids. AlphaSAGE is a framework for automated alpha mining in quantitative finance that targets the discovery of formulaic alphas as symbolic expressions over historical market data. In the formulation given by "AlphaSAGE: Structure-Aware Alpha Mining via GFlowNets for Robust Exploration" (Chen et al., 29 Sep 2025), an alpha is a deterministic function mapping a daily feature matrix $X_d \in \mathbb{R}^{N \times M}$ to a cross-sectional signal $z_d \in \mathbb{R}^N$. The framework is designed to address three linked deficiencies in prior RL-based alpha generation: reward sparsity, semantically inadequate sequential representations of mathematical expressions, and mode-seeking objectives that are misaligned with the practical requirement for a diverse portfolio of low-correlated signals (Chen et al., 29 Sep 2025). The name should be distinguished from two separate papers centered on systems called SAGE, one for scalable AI governance and evaluation in search systems (Le et al., 8 Feb 2026) and one for trigger-guided LoRA-based self-adaptation in LLMs (Wei et al., 5 Sep 2025), both of which mention “AlphaSAGE” only as a hypothetical extension rather than as the finance framework itself.

1. Definition and problem setting

In AlphaSAGE, automated alpha mining is posed as search over the combinatorial space $\mathcal{X}$ of symbolic formulas constructed from market features and typed operators (Chen et al., 29 Sep 2025). These formulas are represented as programs or expression trees rather than as opaque vectors, which preserves interpretability and auditability and aligns with traditional factor workflows. The paper formalizes an alpha as an Abstract Syntax Tree (AST) $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$, where nodes correspond to operators and base features and edges encode typed parent–child relations (Chen et al., 29 Sep 2025).

The motivating deficiencies of existing frameworks are stated explicitly. First, RL-based alpha mining often evaluates performance only after a full, syntactically valid formula has been completed, so intermediate construction steps receive no signal. Second, linearized representations such as Reverse Polish Notation obscure commutativity, hierarchy, and typed operator–operand relations. Third, standard RL objectives maximize expected return and therefore tend toward a single optimal mode, whereas quantitative portfolio construction requires a library of good but non-correlated alphas (Chen et al., 29 Sep 2025). The paper further states that when alphas are correlated, the covariance matrix $\Sigma$ of alpha returns becomes ill-conditioned, which makes linear combination coefficients high-variance and unstable; this motivates explicit diversity control in the generation procedure (Chen et al., 29 Sep 2025).

A common misconception is that AlphaSAGE is a general label for any future “next-generation SAGE” system. The evidence in the cited papers does not support that usage. In the governance paper, “AlphaSAGE” is described only as something that “would naturally be imagined as a next-generation version of SAGE” (Le et al., 8 Feb 2026). In the LLM adaptation paper, “AlphaSAGE-style systems” is a descriptive phrase for extensions of a different SAGE architecture (Wei et al., 5 Sep 2025). The concrete method named AlphaSAGE is the quantitative finance framework introduced in (Chen et al., 29 Sep 2025).

2. Architectural design

AlphaSAGE consists of two main components: AlphaGenerator and AlphaEvaluator (Chen et al., 29 Sep 2025). The AlphaGenerator treats the current state as a partial alpha represented by an AST, encodes that partial structure with a Relational Graph Convolutional Network (RGCN), and uses a GFlowNet policy to emit probabilities over syntactically valid edits, including operator insertion, operand insertion, and early stopping (Chen et al., 29 Sep 2025). The AlphaEvaluator scores completed alphas by computing their outputs over time and assigning a reward composed of predictive performance, structure–behavior alignment, and novelty relative to an existing alpha pool (Chen et al., 29 Sep 2025).

The system begins from an empty state $s_0$, iteratively grows an AST, and terminates when a special separator action produces a syntactically valid formula. The trajectory is written as

$$\tau = (s_0 \rightarrow s_1 \rightarrow \dots \rightarrow s_n = \alpha), \quad \alpha \in \mathcal{X}.$$

Invalid actions are masked by the grammar, so the generative process respects syntactic constraints throughout construction (Chen et al., 29 Sep 2025). The action space includes adding a new operator or feature at an open leaf, choosing a unary or rolling operator with a time window, choosing left or right child for non-commutative operators, and a special SEP action for early stopping (Chen et al., 29 Sep 2025).

AlphaSAGE includes an explicit early-stop mechanism. When the current AST is already valid, SEP is sampled with probability

$$p_{\text{stop}(s_t)} = \frac{\text{Len}(s_t)}{\text{MaxLen}},$$

where $\text{Len}(s_t)$ is the number of nodes in the AST and $\text{MaxLen}$ is the maximum token budget, given as 20 in the reported hyperparameter settings (Chen et al., 29 Sep 2025). The paper’s ablation results show that early stopping by itself worsens results, and that it becomes useful only once paired with stronger structural encoding, especially the RGCN (Chen et al., 29 Sep 2025). This suggests that variable-length expression generation is not inherently beneficial unless the state representation preserves semantic structure.

3. Structure-aware representation

A defining feature of AlphaSAGE is its use of an RGCN over ASTs rather than a sequence model over tokenized formulas (Chen et al., 29 Sep 2025). Each alpha’s AST is converted into a relational graph with explicit node and edge semantics. The paper lists six relation types: unary operator–operand, commutative operator–operands, non-commutative operator–left operand, non-commutative operator–right operand, rolling operator–feature operand, and rolling operator–time/window operand (Chen et al., 29 Sep 2025).

Node representations are initialized as learnable embeddings for operator, feature, and constant types. The RGCN updates them via relation-specific transformations: $z_d \in \mathbb{R}^N$0 and the global alpha embedding is obtained by max-pooling over final-layer node states,

$z_d \in \mathbb{R}^N$1

The reported encoder uses 2 layers and hidden dimension 128 (Chen et al., 29 Sep 2025).

The representational claim is specific: this graph formulation captures commutativity, hierarchical nesting, and typed argument roles more naturally than sequence models (Chen et al., 29 Sep 2025). In the ablation study on CSI300, adding the GNN component to the base GFlowNet raises IC from 0.048 to 0.070 and SR from 0.72 to 1.25, which the paper identifies as the largest single improvement among the architectural additions (Chen et al., 29 Sep 2025). Within the evidence given, this establishes the structure-aware encoder as the central representational mechanism.

The grammar over which the model operates is also explicit. Base features include Open, Close, High, Low, Vwap, and Volume; unary operators include Abs, Slog1p, Inv, Sign, Log, and Rank; binary operators include Add, Sub, Mul, Div, Pow, Greater, and Less; and time-series operators include Ref, TsMean, TsStd, TsRank, TsCorr, TsDelta, TsEMA, and others listed in the appendix (Chen et al., 29 Sep 2025). This grammar constrains both the search space and the semantic regularities encoded by the RGCN.

4. GFlowNet formulation and reward design

AlphaSAGE replaces conventional mode-seeking RL with a GFlowNet objective that targets a terminal distribution proportional to reward: $z_d \in \mathbb{R}^N$2 The framework uses forward and backward policies together with a learnable partition function $z_d \in \mathbb{R}^N$3, and is trained with Trajectory Balance loss (Chen et al., 29 Sep 2025): $z_d \in \mathbb{R}^N$4 An entropy term is added,

$z_d \in \mathbb{R}^N$5

and the final objective is

$z_d \in \mathbb{R}^N$6

with entropy coefficient $z_d \in \mathbb{R}^N$7 in the reported experiments (Chen et al., 29 Sep 2025).

The reward itself is explicitly multi-faceted: $z_d \in \mathbb{R}^N$8 The predictive term $z_d \in \mathbb{R}^N$9 is the absolute average cross-sectional correlation between alpha outputs and future returns: $\mathcal{X}$0 The structure-aware term $\mathcal{X}$1 is defined through behavioral distances between normalized alpha outputs and a kernel over nearest neighbors in embedding space, rewarding cases where structurally similar alphas exhibit similar behavior (Chen et al., 29 Sep 2025). The novelty term is

$\mathcal{X}$2

which directly penalizes correlation with known strong alphas (Chen et al., 29 Sep 2025).

The weights $\mathcal{X}$3 and $\mathcal{X}$4 are annealed over training, with $\mathcal{X}$5 and $\mathcal{X}$6 in the experiments (Chen et al., 29 Sep 2025). Early training therefore emphasizes representation quality and exploration, while later training shifts weight toward predictive performance. The paper characterizes this as a response to reward sparsity: although the rewards remain terminal, the additional terms make feedback smoother than raw IC alone (Chen et al., 29 Sep 2025). A plausible implication is that AlphaSAGE uses auxiliary structural objectives not merely for regularization but as a practical credit-shaping mechanism within a terminal-reward generative setting.

5. Experimental protocol and empirical performance

The experiments use Qlib data on CSI300, CSI500, and S&P500 (Chen et al., 29 Sep 2025). The China splits are Train: 2010-01-01 – 2020-12-31, Validation: 2021-01-01 – 2021-12-31, and Test: 2022-01-01 – 2024-12-31. The US splits are Train: 2010-01-01 – 2016-12-31, Validation: 2017-01-01 – 2017-12-31, and Test: 2018-01-01 – 2020-12-31 (Chen et al., 29 Sep 2025). Baselines include MLP, LightGBM, XGBoost, GP, AlphaGen, AlphaQCM, and AlphaForge (Chen et al., 29 Sep 2025).

The reported metrics cover both signal quality and portfolio behavior. Correlation-based measures are IC, ICIR, RankIC, and RankICIR. Portfolio metrics are Annualized Return (AR), Maximum Drawdown (MDD), and Sharpe Ratio (SR), with backtests defined as long-only top-20% with 20-day holding for CSI300/500 and long–short top/bottom 10% with 20-day holding for S&P500 (Chen et al., 29 Sep 2025).

The main test-set results place AlphaSAGE ahead of the listed baselines across the three markets (Chen et al., 29 Sep 2025).

Market	Selected AlphaSAGE results	Reported comparison
CSI300	IC $\mathcal{X}$7, ICIR $\mathcal{X}$8, RIC $\mathcal{X}$9, RICIR $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$0, AR $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$1, MDD $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$2, SR $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$3	Best on all listed metrics
CSI500	IC $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$4, ICIR $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$5, RIC $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$6, RICIR $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$7, AR $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$8, MDD $$\mathcal{T}_\alpha = (\mathcal{V}_\alpha, \mathcal{E}_\alpha)$$9, SR $\Sigma$0	Best or tied/best-or-near as stated
S&P500	IC $\Sigma$1, ICIR $\Sigma$2, RIC $\Sigma$3, RICIR $\Sigma$4, AR $\Sigma$5, MDD $\Sigma$6, SR $\Sigma$7	Best, with SR slightly better than AlphaForge’s 6.30

The ablation study on CSI300 decomposes the contributions of the major components (Chen et al., 29 Sep 2025). The base GFlowNet without early stopping, GNN, structure-aware reward, novelty, or entropy yields IC $\Sigma$8 and SR $\Sigma$9. Early stopping alone causes AR to become negative and SR to fall to $s_0$0. Adding the GNN raises IC to 0.070 and SR to 1.25. Adding structure-aware reward improves stability metrics and MDD. Adding novelty raises IC to 0.075, RIC to 0.092, AR to 6.77%, and SR to 1.53. Adding entropy yields the full AlphaSAGE with IC $s_0$1, RIC $s_0$2, AR $s_0$3, and SR $s_0$4 (Chen et al., 29 Sep 2025). Sensitivity analysis further shows that performance is robust to a wide region of hyperparameters for the novelty and structure-aware reward weights (Chen et al., 29 Sep 2025).

6. Generated alpha pool, combination strategy, and practical interpretation

The output of AlphaSAGE is not a single formula but a pool $s_0$5 of mined alphas and their embeddings (Chen et al., 29 Sep 2025). This is then combined out of sample using a dynamic linear combiner described as “AlphaForge-style,” which periodically selects effective alphas and learns time-varying linear weights to form a portfolio-level “Mega-Alpha” (Chen et al., 29 Sep 2025). This pool-and-combine design is tightly linked to the framework’s diversity objective: the GFlowNet samples multiple high-reward modes, the novelty reward penalizes redundancy, and the downstream combiner operates over a candidate set that is intended to be both predictive and decorrelated (Chen et al., 29 Sep 2025).

The paper provides illustrative examples of the kinds of formulas AlphaSAGE can generate, such as a volatility-adjusted mean-reversion signal using $s_0$6, $s_0$7, and inverse volatility; a liquidity or volume-based momentum signal using $s_0$8; and a cross-sectional rank of $s_0$9 scaled by $$\tau = (s_0 \rightarrow s_1 \rightarrow \dots \rightarrow s_n = \alpha), \quad \alpha \in \mathcal{X}.$$0 (Chen et al., 29 Sep 2025). The paper explicitly notes that these examples are typical of the grammar rather than a printed catalog of discovered formulas (Chen et al., 29 Sep 2025). Compared with hand-crafted alphas, the generated formulas are described as often combining temporal operators in deeper, non-linear ways, using transforms such as Slog1p, Abs, and Inv, and being selected not only for IC but also for low correlation and stable behavior (Chen et al., 29 Sep 2025).

This suggests that AlphaSAGE is best interpreted not as a symbolic regression system seeking the single highest-IC expression, but as a controlled generative process for constructing a diversified factor library. That interpretation is consistent with the formal objective $$\tau = (s_0 \rightarrow s_1 \rightarrow \dots \rightarrow s_n = \alpha), \quad \alpha \in \mathcal{X}.$$1, the explicit novelty term, and the downstream dynamic linear combination procedure (Chen et al., 29 Sep 2025).

7. Relation to adjacent SAGE systems, limitations, and extensions

The finance framework should be situated against two distinct lines of work that use the name SAGE. "SAGE: Scalable AI Governance & Evaluation" (Le et al., 8 Feb 2026) concerns product-policy evaluation in LinkedIn search systems using a Policy–Precedent–Judge loop, teacher–student distillation, and online monitoring. "A Lightweight Framework for Trigger-Guided LoRA-Based Self-Adaptation in LLMs" (Wei et al., 5 Sep 2025) uses SAGE to denote a trigger-guided, LoRA-based, test-time adaptation system for atomic reasoning subtasks. Neither paper introduces the quantitative-finance AlphaSAGE method, although both invoke “AlphaSAGE” as a hypothetical extension of their own architectures (Le et al., 8 Feb 2026, Wei et al., 5 Sep 2025). This naming overlap can cause confusion, but the technical content is unrelated.

Within automated alpha mining, AlphaSAGE is positioned against genetic programming, RL-based systems such as AlphaGen and AlphaQCM, GAN-based generation via AlphaForge, and more recent LLM-based alpha generation (Chen et al., 29 Sep 2025). The paper’s stated distinction is the combination of AST + RGCN for structure-aware semantic embeddings, GFlowNet training with trajectory balance to sample proportionally to quality, and a dense multi-faceted reward combining IC, structure alignment, novelty, and entropy (Chen et al., 29 Sep 2025).

The limitations discussed or implied in the paper are also specific. There remains risk of overfitting or data leakage, mitigated by long train windows, separate validation and test sets, out-of-sample portfolio metrics, and dynamic combination with rolling selection (Chen et al., 29 Sep 2025). The framework does not explicitly handle regime change through online continual learning; the discovered alphas are found offline, although the dynamic linear combiner adapts weights over time (Chen et al., 29 Sep 2025). Transaction costs and market impact are not explicitly modeled in the reported backtests, and evaluation over many symbolic alphas across long histories is computationally expensive, with training using 10,000 episodes for CSI300 and 20,000 for CSI500 and S&P500 (Chen et al., 29 Sep 2025). The optimizer is not explicitly stated, though the paper notes online-like gradient updates after each complete trajectory (Chen et al., 29 Sep 2025).

The authors’ listed future directions include richer data and asset classes, extended grammar, more advanced multi-objective optimization, online or continual learning, alternative GFlowNet objectives such as Subtrajectory Balance, and integration with LLMs for proposing new operators or seeding initial alphas (Chen et al., 29 Sep 2025). Because these are framed as prospective directions, they should not be conflated with the implemented system. A plausible implication is that AlphaSAGE provides a general search-and-diversify template whose current implementation is finance-specific but whose methodological components—grammar-constrained generation, graph-based formula encoding, and reward-proportional sampling—could be specialized further as operator sets, data modalities, and portfolio objectives broaden.