RankGLU

KO | EN

gitlite — search

#javascript #react #android #rust #deep-learning #linux #macos #java #awesome #cli #vue #pytorch #mcp #open-source

(★ 15)

RankGLU research code

RankGLU Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# RankGLU

Residual gated score formation for cross-sectional stock prediction.

RankGLU studies a narrow but important bottleneck in stock ranking models:
after temporal and cross-stock encoders form a stock embedding, the prediction
head decides how that embedding becomes a daily cross-sectional ranking score.
The proposed head keeps a direct linear scoring path and adds a bounded
bottleneck GLU interaction path:

```text
score = Linear(LayerNorm(x))
      + gamma * Linear((Wv LayerNorm(x)) * sigmoid(Wg LayerNorm(x)))
```

This repository contains the core PyTorch implementation, the ranking-aware
training protocol, and the multi-seed experiment runner used for the manuscript
"RankGLU: Residual Gated Score Formation for Cross-Sectional Stock Prediction".

## Main Result

The retained method is RankGLU score formation with the ranking-aware training
protocol:

| Dataset | Model | IC | Best IC | ICIR | RankIC | RankICIR |
|---|---:|---:|---:|---:|---:|---:|
| CSI300 | Original | 0.0654 +/- 0.0052 | 0.0700 | 0.4347 | 0.0696 | 0.4501 |
| CSI300 | Rank-aware | 0.0697 +/- 0.0030 | 0.0745 | 0.4732 | 0.0752 | 0.5028 |
| CSI300 | RankGLU | 0.0727 +/- 0.0037 | 0.0768 | 0.4801 | 0.0814 | 0.5314 |
| CSI800 | Original | 0.0507 +/- 0.0027 | 0.0537 | 0.4305 | 0.0601 | 0.4802 |
| CSI800 | Rank-aware | 0.0502 +/- 0.0018 | 0.0533 | 0.4056 | 0.0642 | 0.4966 |
| CSI800 | RankGLU | 0.0506 +/- 0.0032 | 0.0561 | 0.3975 | 0.0628 | 0.4698 |

On CSI300, RankGLU has the strongest five-seed mean IC among the internally
controlled variants. On CSI800, it remains competitive and has the highest
best-seed IC, but the broader universe gives a more conservative mean
comparison.

## Repository Layout

```text
rankglu/                         Core model and training utilities
scripts/train_rankglu.py          Single-run training entrypoint
scripts/run_multiseed_protocol.py Multi-seed protocol for main, ablation, diagnostic runs
scripts/run_all_multiseed.*       Windows and Linux convenience wrappers
configs/                          Example environment settings
docs/data.md                      Expected processed data layout
docs/reproducibility.md           Reproduction notes
```

Processed data, model weights, logs, and result folders are intentionally not
tracked by Git.

## Setup

```bash
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```

On Windows PowerShell:

```powershell
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
```

## Smoke Test

The smoke test uses synthetic tensors and does not require market data:

```bash
python scripts/smoke_test.py
```

## Data

Place processed pickle files under `data/opensource/`:

```text
data/opensource/csi300_dl_train.pkl
data/opensource/csi300_dl_test.pkl
data/opensource/csi800_dl_train.pkl
data/opensource/csi800_dl_test.pkl
```

See [docs/data.md](docs/data.md) for the expected tensor layout and data notes.

## Run Experiments

Single RankGLU run:

```bash
python scripts/train_rankglu.py
```

Main five-seed comparison:

```bash
python scripts/run_multiseed_protocol.py --sections main --universe csi300
```

Full protocol:

```bash
bash scripts/run_all_multiseed.sh
```

Windows PowerShell:

```powershell
.\scripts\run_all_multiseed.ps1 -Universe csi300 -Gpu 0
```

Outputs are written to `results/<timestamp>/` or
`results/<universe>_multiseed_<timestamp>/`, including `runs.csv`,
`aggregate.csv`, `core_ablation_delta.csv`, logs, and a text summary.

## Configuration

The default single-run entrypoint is the retained RankGLU setting:

```text
CS_NORM=zscore
LOSS_MODE=mse_ic
IC_WEIGHT=0.1
DECODER_TYPE=residual_bottleneck_glu
DECODER_BOTTLENECK=128
S_SCORE_TYPE=dot
S_VALUE_GATE_TYPE=none
```

Relation-path calibration variants are included as diagnostic stress tests, not
as the retained method. This mirrors the manuscript: relation-path changes can
give high single-seed peaks but are less stable under multi-seed evaluation.

## Citation

Paper: [arXiv:2606.08930](https://arxiv.org/abs/2606.08930).

RankGLU studies residual gated score formation for cross-sectional stock
ranking under information-coefficient-oriented evaluation.

```bibtex
@misc{xiao2026rankglu,
  title={RankGLU: Residual Gated Score Formation for Cross-Sectional Stock Prediction},
  author={Xiao, Huixiang and Xu, Jian and Qu, Feiyu and Xie, Zixuan and Li, Xiangyu},
  year={2026},
  eprint={2606.08930},
  archivePrefix={arXiv},
  primaryClass={cs.CE},
  doi={10.48550/arXiv.2606.08930},
  url={https://arxiv.org/abs/2606.08930}
}
```

Related encoder references are listed in [NOTICE.md](NOTICE.md) for
reproducibility context.

## License

MIT. See [LICENSE](LICENSE) and [NOTICE.md](NOTICE.md).

This repository is for research reproducibility and is not investment advice.

RankGLU

RankGLU Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

RankGLU

RankGLU Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README