Open to opportunities

宇祥傅

@0003434

Message

Fraud & credit risk analyst building decisioning rules and Machine Learning models for imbalanced transaction data.

Taiwan

Message

What I'm looking for

I’m looking to build fraud and credit risk decisioning models that improve precision and reduce false positives on imbalanced, large-scale transaction data—backed by rigorous validation and experimentation to drive measurable business outcomes.

I’m a fraud and credit risk analyst with 5 years in fintech, focused on building rules and Machine Learning models that decide who gets flagged and who gets through on imbalanced, large-scale transaction data.

In fraud detection, I lifted precision from 55% to 75%+ at constant recall, while cutting false-positive alerts by roughly 60% with zero cases missed. I also rebuilt a fraud-victim detection model using time-series behavioral deviation and fund-flow patterns across 10M+ transaction records, handling class imbalance with SMOTE + XGBoost.

On the credit-risk side, I built an Automated Credit Limit Evaluation System replacing manual decisions (8,000 → 24,000+ monthly) and contributing to +10% market share. I’ve delivered default prediction improvements validated on a held-out control group (0.25% → 0.03%) and used A/B tests to improve operational efficiency 3x across intervention content, timing, channel, and broker-dealer frequency.

Experience

Work history, roles, and key accomplishments

Business Analyst

Cathay United Bank

Apr 2025 - Jul 2025 (3 months)

Rebuilt the bank's fraud-victim detection model by engineering time-series behavioral deviation and fund-flow features from 10M+ transaction records. Improved precision from 55% to 75%+ while handling class imbalance using SMOTE and XGBoost.

Fraud Models Time Series Analysis Feature Engineering Fund Flows SMOTE XGBoost Class Imbalance Handling

Data Analyst

Cathay Securities

Aug 2020 - Apr 2025 (4 years 8 months)

Built an Automated Credit Limit Evaluation System replacing manual review with model-driven decision rules: 8,000 → 24,000+(3x) and contributing market share +10% .
Built a default prediction model (XGBoost, selected via cross
validation; SMOTE) generating daily high-risk lists for online/offline
intervention. (Monthly default rate: 0.25% → 0.03%)