Hi, I'm Howard Zeng
Quantitative Researcher & ML Engineer
I build models that explain themselves, pipelines that don't break, and results you can defend.
$400M+ resi credit modeled
100M+ loan-level records
End-to-End Prepay · Default · Loss
About
I build and productionize loan-level prepayment / delinquency models and data pipelines across the full RMBS stack — CRT (STACR/CAS), MIR, Non-QM, Jumbo, and HELOC.
I own modules end-to-end — from factor design and diagnostics to daily risk refresh, tracking, and explainable outputs used in investment decisions.
I care about turning messy, high-dimensional data into models that actually ship — and about writing code the next person can read.
R (mgcv/bam GAM) Python (Polars, XGBoost/LightGBM) SQL (Redshift/SQL Server) Jenkins/PowerShell Ray
Previously at
Structured credit · Lending analytics · Asset pricing · Derivatives research
LibreMax Capital Navy Federal Credit Union Gravity Investments Huatai Securities China Securities
Skills
Structured Credit / Modeling
RMBS CRT (STACR/CAS) Non-QM Jumbo HELOC Prepayment Default/Loss Scenario/Stress Testing
ML / Statistics
GAM (mgcv/bam) XGBoost LightGBM Logistic Regression Time Series Imbalanced Classification Walk-forward Validation
Data / Engineering
Python (Polars/Pandas) C++ R (Tidyverse) SQL (Redshift/SQL Server) Parquet Jenkins PowerShell Automation/QC
Systems / Tooling
Git Linux Ray (distributed compute) Azure AWS Teradata Reproducible Configs Model Versioning
Experience
2025
Feb 2025 - Present 
- Production prepay/delinquency/loss models across the full resi stack — CRT (STACR/CAS), MIR, Non-QM, Jumbo, HELOC; factors stable under stress.
- End-to-end pipeline: SQL extraction → feature engineering → GAM/XGBoost fitting → automated QC (Jenkins/Python) on loan-level data.
- Own full model lifecycle — factor design through daily risk refresh — and deliver tracking outputs and risk explanations to PMs.
RMBS CRT Non-QM Jumbo Python SQL
2024
May 2024 - Aug 2024 
- Built executive dashboard for a $6B equity-loan portfolio; surfaced pricing anomalies that informed portfolio strategy.
- SQL/Python pipelines on Azure/Teradata processing 100M+ records with automated anomaly detection.
- End-to-end delivery: KPI design → Plotly visualizations → presentation to senior leadership.
Python SQL Azure Pricing Dashboards
Jan 2024 - May 2024 
- Led 6-person team; built predictive models that outperformed S&P 500 benchmark in backtests.
- Migrated Temporal Fusion Transformers to AWS SageMaker, optimizing 20M+ record workloads.
- Drove full pipeline architecture — feature engineering through deployment — as team lead.
Python PyTorch AWS SageMaker Time Series
2023
Jan 2023 - Aug 2023 
- Reduced dataset 91% while preserving predictive signal across ~9M mortgage records; strong holdout AUC.
- SMOTE resampling + Optuna-tuned XGBoost/RF/Logistic classifiers with cross-validation.
- End-to-end mortgage default modeling — data prep through model selection and evaluation.
Mortgage Python PostgreSQL XGBoost
2022
Nov 2022 - Mar 2023 
- Supported deep demand estimation research with reliable data pipelines and model experiments.
- Large-scale econometric dataset preparation and preliminary model evaluation.
- Independently managed preprocessing and evaluation workflows.
Python Demand Estimation Econometrics
Sep 2022 - Dec 2022 
- Published equity research on database technology and virtual power plant sectors.
- Analyzed 19,424 funds using R and Wind terminal data for sector analysis.
- Owned full analysis from data collection through published research notes.
R Wind Equity Research
Feb 2022 - Apr 2022 
- Improved derivatives strategy Sharpe ratio by 17% through ARIMA/GBDT model refinement.
- 11-year (2010–2021) Chinese futures dataset across multiple asset classes.
- Designed rolling-window and walk-forward validation for cross-regime robustness.
Python Derivatives Time Series
Earlier Experience
iRent
Mobile Full Stack Developer | Founder & Team Leader
May 2019 - Jan 2021
Education
Cornell University
2023 - 2024MS Applied Statistics — Data Science
GPA: 4.08 / 4.3
Relevant Coursework
Large-Scale Machine Learning Deep Learning Natural Language Processing Reinforcement Learning Stochastic Processes Statistical Modeling Sequence Models Convolutional Neural Networks Machine Learning
University of Washington
2019 - 2022BS Economics — Econometrics
Minor: Applied Math & Data Science
GPA: 3.6 / 4.0
Relevant Coursework
Econometric Theory & Applications Causal Inference Data Science for Pricing Financial Economics Database Systems (SQL) Linear Algebra & Numerical Analysis Differential Equations Scientific Computing Data Science in Python
Featured Case Studies
AlphaCycle — Stock Prediction Framework
- Problem
- How to build a reproducible, config-driven research stack for systematic equity signals?
- Approach
- Built a 3-layer framework (data ingestion → feature store → model training → reporting) with walk-forward validation and regime labels.
- Result
- Clean separation of data/model/report layers; config-driven pipelines; evaluation with robust metrics across multiple horizons.
Python XGBoost Pipeline Walk-forward
Read more →
Retail Sentiment Trading Signals
- Problem
- Can retail investor sentiment on social media predict short-term equity price movements?
- Approach
- Built an NLP pipeline with VADER and FinBERT to score Reddit/Twitter posts, then trained classifiers on sentiment-price lag features.
- Result
- Achieved statistically significant predictive signal for 1–3 day returns; backtested strategy returned ~12% annualized alpha on selected tickers.
Python NLP FinBERT Time Series
Read more →
Handling Class Imbalance in Random Forest
- Problem
- Standard Random Forest classifiers degrade on imbalanced datasets common in fraud detection and credit risk.
- Approach
- Systematically compared SMOTE, ADASYN, Tomek links, cost-sensitive learning, and ensemble balancing across multiple imbalance ratios.
- Result
- Cost-sensitive RF with SMOTE achieved 15–20% F1 improvement over baseline on highly skewed datasets.
Python Scikit-learn SMOTE Classification
Read more →