Pareto-DQN Trades Precision for Recall in Financial Anomaly Tests

A v1 arXiv paper proposes Semantic Pareto-DQN, a multi-objective reinforcement-learning system for financial anomaly detection. Instead of asking one classifier to optimise a single loss, the method represents detection value, false-positive friction, and semantic novelty as separate reward dimensions and searches for policies along a Pareto frontier.

The pipeline first converts tabular transaction or credit records into natural-language narratives, then encodes them as 384-dimensional vectors with the all-MiniLM-L6-v2 sentence transformer. The agent combines each embedding with a rolling false-positive signal and learns whether to approve or block an item.

What the authors tested

The e-commerce experiment uses roughly 150,000 records, while the UCI credit-default dataset contains 30,000 clients and 23 explanatory variables. Both experiments use an 80/20 chronological train-test split. The baselines are raw-feature XGBoost, embedding-based XGBoost, and a single-objective DQN.

On the paper's e-commerce test split, the model reached 0.267 recall at 0.069 precision, producing 371 true positives and 5,026 false positives. Raw-feature XGBoost reported zero recall in the authors' chosen configuration. The Pareto agent therefore found more fraud cases, but it generated about fourteen false alerts for every true alert.

On the UCI credit-default test, Pareto-DQN reported 0.419 recall, 0.563 precision, and a 0.481 F1 score, with 531 true positives and 413 false positives. Raw-feature XGBoost reported 0.353 recall, 0.652 precision, and a 0.458 F1 score. That is a smaller and more operationally plausible trade-off than the e-commerce result.

What the comparison does not establish

The results are point evaluations on one chronological split per dataset. The paper says multi-seed cross-validation and confidence intervals remain future work, so the reported gaps do not yet show how much performance varies with initialisation or data partition. Reward coefficients and friction limits were manually calibrated offline for each dataset, which gives the method access to domain-specific tuning that must be repeated in a new environment.

The supervised baselines also deserve a stronger challenge. A production comparison should sweep class weights and decision thresholds, include calibrated probability models, and compare every method at the same false-positive or review-capacity budget. Without that curve, zero recall from one XGBoost configuration is not evidence that supervised learning inherently collapses.

Narrative encoding adds another open question. In the UCI experiment, embedding-based XGBoost underperformed raw-feature XGBoost, showing that converting numerical histories into text can discard useful structure. The embedding may still help the Pareto novelty reward, but an ablation is needed to separate gains from the representation, multi-objective reward, sequential state, and action-selection method.

LDS analysis: compare policies at a fixed review budget

For fraud teams, recall is not free. Investigators, customer challenges, and blocked transactions impose capacity and financial costs. The fairest evaluation fixes a daily review budget or false-positive ceiling, then compares recovered loss, missed fraud, customer friction, calibration, and stability under temporal drift.

The paper is valuable because it makes that trade-off explicit instead of hiding it inside one scalar score. Its next step should be a reproducible benchmark with released code, repeated seeds, threshold-matched baselines, out-of-time drift tests, and dollar-valued utility. Until then, Semantic Pareto-DQN is a research framework for exploring operating points, not evidence of deployment readiness.