Yullie Yang — Quantitative Analyst

About

I work at the intersection of economics, data, and engineering — building reproducible R / Python pipelines, validation and QA workflows, and AI-assisted documentation tooling that turn macro and financial data into versioned, reviewable artifacts.

My focus is production-style research-support code: pulling series from authoritative sources, harmonizing across frequencies, computing derived measures, producing short-horizon forecasts with documented uncertainty, and using AI coding tools responsibly to make recurring QA, code review, and documentation workflows more systematic — packaged so a reviewer can audit every step from Git.

Background

Quantitative Analyst, CoStar Group — Boston, MA (2025 – present)
Data Scientist / Consultant, Guidehouse — McLean, VA (2024 – 2025)
Graduate Intern, Federal Reserve Board, Division of International Finance — Washington, DC (Summer 2023)
M.S. Business Analytics, University of Maryland, College Park (Dec 2023)
B.A. Global Political Economy, Waseda University, Tokyo (2022)

Featured work

Macro forecasting

r-macro-trade-commodity-forecast

Reproducible R pipeline pulling 13 FRED macro / trade / commodity series into a quarterly panel with implicit trade deflators and terms of trade. 8-quarter auto.arima forecasts for net exports, real GDP, and WTI; distributed-lag exchange-rate pass-through regression on U.S. trade prices. CI + Quarto Pages dashboard.

R · forecast · tidyverse · GitHub Actions · Quarto

Stress testing

cre_stress_test

Portfolio-demo stress-testing workflow on 100% public data (FRED + Google Mobility + Boston Zoning). Python package + R / auto.arima companion + SQLAlchemy persistence + Streamlit dashboard. pytest + CI. Personal project; does not reflect any employer's internal data or models.

Python · scikit-learn · SQLAlchemy · Streamlit · R

AI / LLM workflows

llm-research-workflow-assistant

Prompt templates, sample outputs, and a human-in-the-loop checklist for using AI coding tools responsibly in recurring research workflows — data QA, code review, brief review, documentation drafting. CLAUDE.md codifies model behavior.

Markdown · Python · prompt engineering · responsible AI

Applied LLM tooling

cardnews

Claude API + Puppeteer pipeline that renders a topic into a 10-slide 1080×1080 visual deck. Schema-constrained model output, deterministic per-run artifacts, explicit human-review framing.

Node.js · Anthropic SDK · Puppeteer

How I use AI tools

I treat AI coding tools as collaborators on boilerplate, refactoring, and documentation — not as substitutes for analyst judgment. Every repo that uses an LLM in its workflow includes a CLAUDE.md defining how the model should behave, a clear human-review step before outputs are shared, and explicit limitations of what the AI-assisted output represents.

Toolbox

Languages: R (primary), Python, SQL
Data & infra: FRED, SQLite / PostgreSQL / SQLAlchemy, AWS, Azure, Tableau, Power BI
Engineering: Git, GitHub Actions, R Markdown / Quarto, Streamlit, make-driven pipelines, pytest, testthat
Modeling: auto.arima, distributed-lag regression, classification under class imbalance, SHAP explainability
AI / LLM: Claude API, Claude Code, prompt-template design, human-in-the-loop review processes, retrieval-augmented document workflows

Notes

Short methodology notes on the techniques behind the portfolio repos.

Cumulative exchange-rate pass-through: what the coefficient says and what it doesn't — what a distributed-lag pass-through regression actually estimates, and what it cannot.

Contact

yullieyang@gmail.com github.com/yullieyang linkedin.com/in/yullie-yang Resume (PDF)