PARiskRanker: Learn to Rank Risky Investors

🔥 Official implementation of our ACM TOIS 2025 paper: “Learn to Rank Risky Investors: A Case Study of Predicting Retail Traders’ Behaviour and Profitability” by Waylon Li and Tiejun Ma.

📄 Paper DOI: 10.1145/3768623

✨ Highlights

• Profit-Aware Risk Ranker (PA-RiskRanker) reframes risky investor detection as a ranking problem rather than classification.
• Introduces PA-BCE loss to integrate Profit & Loss (P&L) into LETOR training.
• Self-Cross-Trader Attention captures both intra-trader and inter-trader dependencies.

🛠️ Setup

conda create -n pariskranker python=3.10
conda activate pariskranker
pip install -r requirements.txt

📂 Data

1. Download preprocessed data here.
2. Unzip into the project root. Expected structure:

├── data
│   ├── creditcard
│   │   ├── fold1 / fold2 / fold3
│   │   └── creditcard.csv
│   ├── jobprofit
│   │   ├── fold1 / fold2 / fold3
│   │   └── job_profitability.csv
├── evaluation
│   └── metrics.py
...

📦 Pre-trained Model (Optional)

👉 [TODO: Add link] – Place it in the project root as with the data folder.

🎯 Reproducing Results

Run any of the following to benchmark:

# PARiskRanker
python run_pariskranker.py test \
  --model_group_size 100 --test_group_size 100 \
  --dataset creditcard --fold 1 \
  --strategy binary --loss_fn graph

We also provide scripts for Rankformer, LambdaMART, SOUR, and baselines (classification & anomaly detection). See README sections for full commands.

📊 Final evaluation notebooks: notebook/eval.ipynb

🏋️ Train from Scratch

FOLD=1  # choose from 1,2,3
python run_pariskranker.py train \
  --epochs 200 --batch_size 128 \
  --group_size 100 --dataset creditcard \
  --fold $FOLD --strategy binary --pnl 1 \
  --loss_fn graph

Hyperparameters can be customised via --help.

📈 Citation

If you use this code, please cite:

@article{10.1145/3768623,
  author = {Li, Weixian Waylon and Ma, Tiejun},
  title = {Learn to Rank Risky Investors: A Case Study of Predicting Retail Traders’ Behaviour and Profitability},
  year = {2025},
  issue_date = {January 2026},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  volume = {44},
  number = {1},
  issn = {1046-8188},
  url = {https://doi.org/10.1145/3768623},
  doi = {10.1145/3768623},
  journal = {ACM Trans. Inf. Syst.},
  month = nov,
  articleno = {15},
  numpages = {33},
  keywords = {learning to rank, domain-specific application, individual behaviour modelling, risk assessment}
}

🙌 Acknowledgements

This work was conducted at the Artificial Intelligence Applications Institute, School of Informatics, University of Edinburgh.