Sequence-based Target Coin Prediction for Cryptocurrency Pump-and-Dump (SIGMOD 2023)

This is the repo including dataset and code used in the paper "Sequence-based Target Coin Prediction for Cryptocurrency Pump-and-Dump".

Data Science Pipeline

The workflow mainly consists of two parts: data collection and target coin prediction.

Pump-and-dump Activity Logs (Jan. 1, 2019 to Jan. 15, 2022)

The P&D logs includes 1,335 samples and 709 P&Ds that we observed on Telegram. We will periodically update this dataset.

Getting Start

Since we have already collected P&D log dataset and will periodically update it, you can skip the data collection part : )

0. Data Collection

First, we get seed channels verified by PumpOlymp and explore the pump channels.

• /Data/Telegram/Pump_channel1
• /Data/Telegram/Pump_channel2

Step1: get historical messages according to channel

python get_channel_post.py

Step2: Select pump message with keyword filtering

python keyword_filte.py

Step3: Manually label the filtered messages

python pump_message_label.py
# we have already labeled 5000+ samples. 

Step4: Generate features for messages.

python message_fg.py

Step5: Train a detection classifier.

python train_classifier.py

Step6: Use the classifier for detection

python predict_classifier.py

Step7: Aggregate the session based on timestamp

python sess_aggregate.py

Step8: Label the predicted pump session and generate final P&D log

python PD_label.py

Step9: Clean the log dataset

python PDlog_clean.py

1. Target Coin Prediction

1.1 Feature Generation:

Two methods to generate features for Target Coin Prediction.

Method1: Generate features from P&D logs:

HOLD (still organizing this part of code because it involves feature collection from multiple sources,)

Mehod2: Download generated dataset from Good Drive

• train_sample

• test_sample

cd TargetCoinPrediction
tar -xzvf train.tar.gz
tar -xzvf test.tar.gz

1.2 Model Training

Step1: Train SNN model

cd TargetCoinPrediction/SeqModel
python run_train.py  --model=snn \
                     --max_seq_length=8 \
                     --epoch=30 \
                     --batch_size=256 \
                     --learning_rate=1e-4 \
                     --dropout_rate=0.2
                     --do_train=True
                     --do_eval=False \
                     --checkpointDir=xxx 

Parameter	Description
model	Model used for target coin prediction, options (snn, snnta, dnn)
max_seq_length	The maximum length of P&D sequence, options 1~ 50, default=8.
epochs	Number of training epochs, default = 30.
batch_size	Batch size, default = 256.
learning_rate	Learning rate for the optimizer (Adam), default = 5e-4.
dropout_rate	Dropout Ratio for training, default = 0.2.
do_train	Whether to do training or testing, default = True.
do_eval	Whether to do training or testing, default = False.
checkpointDir	Specify the directory to save the checkpoints.

Step2: Evaluate SNN model

python run_eval.py   --model=snn \
                     --max_seq_length=8 \
                     --epoch=1 \
                     --batch_size=256 \
                     --do_train=False
                     --do_eval=True \
                     --checkpointDir=xxx \
                     --init_seed=1234 

Results on current dataset

We periodically update this table according to updated dataset.

Metric	DNN	SNN	SNN_s
HR1	0.203	0.253	0.313
HR3	0.295	0.361	0.445
HR5	0.431	0.471	0.498
HR10	0.463	0.599	0.599
HR20	0.630	0.709	0.714
HR30	0.797	0.824	0.846

In this dataset we generate only 9 statistical features and coin id for pumped coin in sequence. The performance of SNN can be further improved by using more features for sequence.

---

Citation

If you find this repo useful, please cite our paper:

@article{hu2022sequence,
  title={Sequence-Based Target Coin Prediction for Cryptocurrency Pump-and-Dump},
  author={Hu, Sihao and Zhang, Zhen and Lu, Shengliang and He, Bingsheng and Li, Zhao},
  journal={arXiv preprint arXiv:2204.12929},
  year={2022}
}