data-layout="article"

Yifan Li

444 Washington Blvd, Jersey City, NJ, 07310

github.com/yifan-lee

Education

University of Connecticut

PhD in Statistics

Connecticut, USA Sep 2018 - Sep 2023

University of Wisconsin - Madison

Master of Science in Statistics

Wisconsin, USA Sep 2016 - May 2018

Nanjing University

Bachelor of Science in Statistics

Jiangsu, China Sep 2013 - Jun 2017

Professional Summary

Data Scientist and Quantitative Consultant with a PhD in Statistics. Experienced in derivative pricing, CCAR stress testing, multi-agent reporting systems with LangGraph, and credit risk modeling. Strong background in statistical inference, algorithm optimization (American Monte Carlo), and machine learning on imbalanced data.

Work Experience

Ernst & Young LLP (EY)

Senior Consultant | Data Science & Quantitative Strategy

New York, NY Oct. 2023 - Present

  • Dual-Stage Credit Risk Modeling & High-Risk Applicant Screening

    • Architected a two-stage credit risk framework on a dataset of 200,000+ loan applicants: an unsupervised screening engine for instant cold-start risk filtering, and a supervised ensemble model for applicants with repayment history.

    • Implemented Isolation Forest to isolate extreme high-risk outliers based on tree partition path lengths; selected it for its low computational latency and minimal memory footprint, enabling real-time, hardware-constrained screening without requiring historical labels.

    • Aggregated raw payment logs into behavioral metrics, created missing indicator variables to measure data completeness, and applied class-weighted loss to address severe default imbalance.

    • Ensembled Logistic Regression, Random Forest, and XGBoost in a mixture-of-experts style, boosting model AUC by +0.02 and establishing reliable approval cutoffs for risk management.

  • Multi-Agent Financial Reporting System based on LangGraph

    • Built a multi-agent system using LangGraph to automate Ongoing Performance Assessment (OPA) reporting, eliminating heavy manual documentation.

    • Orchestrated specialized agents to execute the end-to-end quantitative workflow: Option Pricing, Validation, Sensitivity & Risk Testing, and Report Synthesis, achieving maximum reusability across diverse report templates sharing common analytical agents.

    • Developed an interactive Streamlit dashboard allowing stakeholders to upload raw data, trigger the pipeline, and visualize validation metrics in real time.

    • Reduced the operational effort from a team of dozens to a small maintenance crew, significantly shortening report delivery time while ensuring high consistency.

  • CCAR Stress Testing for Trading Book (ETD & OTC)

    • Redesigned the CCAR stress testing framework (Comprehensive Capital Analysis and Review) into a modular architecture to simulate P&L and risk impacts across $20B+ in notional exposure for equity, FX, and commodity derivatives under regulatory scenarios.

    • Replaced fragmented manual workflows (spreadsheets, ad-hoc scripts) with a standardized Python pipeline to automatically ingest, clean, and merge data across disparate databases.

    • Implemented a Depth-First Search (DFS) algorithm to traverse complex, nested trade attribute dictionaries, extracting precise contract terms for exact pricing.

    • Shortened the end-to-end stress testing cycle from months to weeks while significantly improving P&L calculation accuracy and audit readiness.

  • Accelerated American Options Pricing via American Monte Carlo (AMC)

    • Resolved the computational bottleneck in American option valuation, replacing computationally prohibitive MCMC-over-MCMC simulations. Implemented the American Monte Carlo (AMC) framework, reducing calculation time from 𝑶(𝒏𝟐)\mathbf{O(n^2)} to 𝑶(𝒏)\mathbf{O(n)}.

    • Incorporated cross-asset underlying correlation structures into the joint simulation, enabling simultaneous valuation of mixed portfolios (e.g., vanilla and binary American options).

    • Slashed portfolio pricing runtime from days to a few minutes while preserving strict pricing precision and convergence standards.

Realmwand: The First Tower (2D RPG Game)

Solo Developer | Indie Game Project

  • Architected and developed a 2D Active Time Battle (ATB) grid-based RPG from scratch, implementing a clean, modular architecture powered by an event-driven signal system and strict data/runtime state separation.

Skills

Product & Business Analytics User/Player Segmentation, Propensity Modeling, LTV Optimization, Anomaly/Outlier Detection

Machine Learning & Statistics XGBoost, Random Forest, Logistic Regression, Isolation Forest, PCA, K-Means, Decision Trees, ViT

Experimentation & Evaluation A/B Testing, Experiment Design, Causal Inference, Hypothesis Testing, ROC Analysis, FP/FN Tradeoffs

Languages & Tools Python, R, PyTorch, SQL, Streamlit, LangGraph, Godot