RankGLU: Residual Gated Component Boosts Cross‑Sectional Stock Ranking Performance by 11%

Cross‑sectional stock prediction is fundamentally a ranking problem, where investment decisions depend on the relative ordering of assets each trading day. Common quantitative metrics such as Information Coefficient (IC), RankIC, and their normalized forms (ICIR, RankICIR) emphasize ordering rather than absolute return magnitude.

Existing models focus on improving stock representations via graph, hypergraph, attention, or market‑condition Transformers, but treat the final prediction head as a minor component. This leads to a methodological gap: overly flexible heads may over‑fit unstable returns, while overly linear heads cannot exploit cross‑feature interactions, harming IC‑oriented performance.

Problem Definition

For a trading day d, let S_d be the set of tradable stocks with size N_d = |S_d|. Each stock u has a historical feature window X_{u,d} = [x_{u,d-\tau+1}, \dots, x_{u,d}] (\tau is the look‑back length, F features) and a market‑state vector m_d. The future return label is r_{u,d+h} (or r_{u,d} when the horizon is implicit). The learning task is to estimate a scalar score \hat{r}_{u,d}=f_{\theta}(X_{u,d}, m_d, S_d) whose cross‑sectional ordering aligns with r_{u,d}. The IC for day d is IC_d = corr(\hat{r}_d, r_d), and the overall goal is to maximize the average IC across test days.

Method: RankGLU Prediction Head

RankGLU replaces the final linear mapping with a residual bottleneck GLU module. The module retains a direct linear scoring path ( W_v e_u) and adds a bounded multiplicative branch ( \gamma \cdot GLU(W_g e_u)), where e_u is the temporally aggregated stock embedding, \gamma is a scaling factor, and the hidden dimensions of W_v and W_g are set by the decoder bottleneck. This design allows controlled non‑linear interaction while preserving a stable ranking pathway.

The overall architecture follows the five‑stage time‑relation encoder used in prior work: market‑condition gating, intra‑stock aggregation, inter‑stock aggregation, temporal aggregation, and the RankGLU head. The market state first modulates feature importance; intra‑stock aggregation extracts temporal patterns per asset; inter‑stock aggregation transfers information among correlated assets; temporal aggregation consolidates the sequence; finally, RankGLU maps the aggregated embedding to a score.

Experimental Setup

Experiments are conducted on two Chinese A‑share datasets: CSI300 and CSI800. Daily stock features are drawn from the Alpha158 factor set, with a look‑back window of 8 trading days and a prediction horizon of 5 days. Training data span Q1 2008–Q1 2020, testing data Q3 2020–Q4 2022, and Q2 2020 serves as a temporal buffer.

Baselines cover two groups. The first includes XGBoost, LSTM, GRU, TCN, Transformer, GAT, DTML, and the prior time‑relation attention model. The second group consists of internal controlled variants: the original backbone, a ranking‑aware backbone, and RankGLU.

Evaluation metrics focus on ranking quality (IC, RankIC, ICIR, RankICIR) and portfolio‑oriented measures (annualized excess return AR, information ratio IR). All experiments are repeated over five random seeds; results are reported as mean ± standard deviation and best‑seed performance.

Results

On CSI300, RankGLU raises the average IC from 0.0654±0.0052 (original backbone) and 0.0697±0.0030 (ranking‑aware backbone) to 0.0727±0.0037, an ≈11.2% gain over the original and ≈4.3% over the ranking‑aware variant. On CSI800, RankGLU achieves the highest best‑seed IC of 0.0561, while its average IC remains comparable to the original backbone.

Statistical Significance

Two‑sided paired tests at both seed and day levels confirm that RankGLU’s improvement over the original backbone is statistically significant. The additional gain over the ranking‑aware backbone loses significance after accounting for daily IC autocorrelation.

Ablation Study

Removing the GLU prediction head causes the most pronounced performance drop, whereas eliminating the relational path component does not affect the mean across seeds. This indicates that the GLU head is the robust component driving gains, while the relational path can yield high single‑seed peaks but suffers from seed variance.

Discussion

The prediction head is more than a read‑out layer; it determines whether learned embeddings translate into stable rankings. RankGLU’s residual branch preserves a direct scoring route, and its bounded multiplicative branch introduces controlled non‑linear interaction, explaining the observed IC improvements on CSI300. Seed variance analysis shows that the relational‑path variant attains the highest best‑seed IC on CSI300 but exhibits larger standard deviation, suggesting weaker initialization robustness. Across datasets, RankGLU maintains competitive average IC on the noisier CSI800, indicating partial transferability of the scoring‑head mechanism despite increased tail noise and ranking‑stability sensitivity.