Combating Market Manipulation: A Wasserstein GAN-GP Augmented, Focal-Loss TabNet for Cryptocurrency Pump-and-Dump Detection (AIUB Journal of Science and Engineering (AJSE))

Summary

• On the primary 25-second chunk resolution, the proposed framework reached a precision of 100.0%, a recall of 85.7%, and an F1-score of 92.3% on a stratified held-out test split.p.1
• Under 5-fold stratified cross-validation with fold-wise generator retraining at 25-second resolution, the model recorded a mean F1-score of 0.875 with precision of 0.878 and recall of 0.874.p.1
• The augmentation strategy combines 30,000 synthetic minority rows generated per fold with a thirty-fold replication of genuine minority rows to reach an imbalance ratio of approximately 8.9 majority rows to one minority row.p.1
• A five-configuration ablation demonstrates marginal F1-score contributions of +0.019 from WGAN-GP synthesis, +0.006 from hard-negative mining, and +0.038 from focal loss over the baseline.p.1
• Evaluating the 25-second-tuned hyperparameter configuration without re-tuning on 15-second and 5-second chunk aggregations yields lower cross-validated F1-scores of 0.840 and 0.766.p.1
• The unaugmented dataset at 25-second resolution contains 482,157 rows, of which 317 (0.066%) are labelled as pump-and-dump activity and 481,840 as normal activity.p.4

Key insights

1. The largest single performance gain in the proposed pipeline is attributable to the focal loss function rather than the generative synthetic data generator.p.1
2. Retraining the WGAN-GP generator independently within each cross-validation fold eliminates synthetic-to-validation data contamination.p.2
3. Replicating genuine minority transactions thirty-fold without synthesis increases the classifier's structural weight on real class boundaries.p.6
4. Applying hyperparameters tuned at a 25-second resolution to finer temporal granularities results in monotonic performance degradation down to an F1-score of 0.766 at 5 seconds.p.9
5. The single held-out test set at 25-second resolution contains 96,432 transactions, of which only 63 are pump-and-dump instances.p.5
6. The baseline class-weighted TabNet model without oversampling achieved an F1-score of 0.860 with a precision of 0.961 and recall of 0.778.p.8
7. Cross-validation across five folds showed an F1-score range of 0.826 to 0.952 with a standard deviation of 0.047.p.8