Hi, I'm Paul Merica

Senior Data Scientist & Machine Learning Engineer

I transform data into actionable insights and build intelligent solutions that drive business value

View My Work Get In Touch Interview My AI Persona

About Me

Paul Merica

I'm a Senior Data Scientist at Ally Bank specializing in quantitative modeling and credit loss prediction. With a Master's degree in Data Science from Wayne State University and a background in Statistics from the University of Michigan, I bring both theoretical knowledge and practical expertise to complex data challenges.

I develop production-grade models that score billions of dollars in originations, build automated reporting systems, and deliver actionable insights to executive leadership. My work spans machine learning, statistical modeling, data engineering, and business intelligence.

R
Python
SQL
PowerBI
Tableau
AWS

Experience

Senior Data Scientist, Quantitative Modeling Credit Losses

Ally Bank

2023 - Present

Develop structure and methodology for new credit loss scoring models. Designed and implemented model for smaller lease portfolio. Lead documentation and implementation with modeling governance teams.

Data Scientist, Auto Pricing & Credit Losses

Ally Bank

2022 - 2023

Revamped predicted credit losses process scoring $3B+ monthly originations. Model recalibration for improved accuracy. Monthly executive reporting on loss KPIs.

Associate Data Scientist, Auto Pricing

Ally Bank

June 2021 - 2022

Developed and maintained internal Logistic Regression Model for auto loan pricing. Bi-weekly executive presentations using R, SAS, and PowerBI. Automated weekly reporting and data pipeline development.

Analyst, Decision Sciences

Universal McCann

2020 - May 2021

Developed Tableau dashboards enabling automated client reporting. Built data visualization solutions for Fortune 100 marketing campaigns.

Associate, Decision Sciences

Universal McCann

June 2019 - 2020

Ensured accuracy of automated data pipelines using Google DCM data. Performed cluster analysis on 300M+ rows using AWS to create national high-value audiences for major mortgage company. Awarded best marketing campaign pitch for Fortune 100 client.

Education

Master of Science

Wayne State University

Data Science and Business Analytics

December 2023

Bachelor of Science

University of Michigan

Double Major: Statistics & PPE (Political Science, Philosophy & Economics)

May 2020

Featured Projects

Professional Work @ Ally Bank & Universal McCann

A selection of production work I cannot share publicly.

  • Built credit loss models scoring $3B+ in monthly auto loan originations
  • Designed and implemented new model for smaller lease portfolio from scratch
  • Led model documentation and governance with compliance teams
  • Automated executive KPI reporting with weekly data pipeline development
  • Developed internal logistic regression model for auto loan pricing
  • Ran cluster analysis on 300M+ rows in AWS Redshift to build high-value national audiences for Fortune 100 clients
  • Built Tableau dashboards enabling automated client reporting
  • Awarded best marketing campaign pitch for Fortune 100 client
R Python SAS SQL AWS Redshift PowerBI Tableau

AI Agent Playground

I point my AI agent at a problem, give it some direction, and let it run. This repo is where those experiments live. Part collaboration, part chaos, all for fun. Stop by and see what we've been up to.

AI Agents Python Just for Fun
See What It's Building My AI Agent's Portfolio

Get In Touch

Have a project in mind? Let's work together!

Currently

Senior Data Scientist @ Ally Bank

Chat with AI Paul

Ask me anything about my work, experience, or data science!

Hey! I'm AI Paul. Ask me about my work at Ally Bank, my projects, or anything data science related!

Suggested questions:

Big City Recession Analysis with Monte Carlo Power Analysis and Non-Parametric Hypothesis Testing

Stats 406 | University of Michigan

The 2008 recession hit every state differently, and a prevailing theory suggested that states with large metropolitan areas (cities over 500,000 population) would recover faster as jobs concentrated in urban centers. This project tested that theory using Bureau of Labor Statistics U6 unemployment data across all 50 states from 2007 to 2017. The core statistical challenge was choosing the right hypothesis test for small, potentially non-normal samples (n=22 and n=28). Monte Carlo simulation was used to estimate the power of both the t-test and Wilcoxon-Mann-Whitney test under normal and t-distributions, ultimately selecting WMW for its robustness. Permutation testing with 2,000 iterations visualized the null distribution. The result: no statistically significant difference was found (p=0.617), though removing outliers (Michigan, Nevada, New Mexico) dropped it to p=0.398, hinting at a real but underpowered effect. The project highlights how small sample sizes can mask meaningful differences and suggests future work incorporating GDP and population migration data.

Download Paper

Predicting Student Performance from Gameplay

Kaggle Competition | Wayne State University - DSA 6100

Can you tell if a student is learning just by watching them play a game? This project tackled a Kaggle competition with 26 million gameplay events across 18 assessment questions. The core challenge was engineering meaningful features from raw clickstream data (hover durations, page navigation patterns, and event sequences) that could predict correctness without any traditional test. Six models were compared (XGBoost, Logistic Regression, and Random Forest, each in balanced and unbalanced variants) using grouped 5-fold cross-validation to prevent data leakage between students. The balanced XGBoost model was selected for its stronger recall on incorrect answers, which mattered most for identifying struggling students before they fall behind.

Download Presentation

Hospice End-of-Life Prediction

NorthStar Care Community | Wayne State University Practicum

NorthStar Care Community needed to know which hospice patients were likely to die within the next 7 or 14 days. Without that early warning, nurses couldn't mobilize resources in time and Medicare/Medicaid reimbursement for end-of-life care went uncaptured. Using 20 years of electronic health records (2.4M+ visit records), 181 features were engineered across rolling weekly time windows capturing visit patterns, PPS scores, clinical measurements, and cost trends. The key modeling decision was optimizing for recall over accuracy: with only 1.78%-2.94% of records representing imminent death, a naive model would miss most at-risk patients. The aggressive XGBoost models used scale_pos_weight tuning to maximize detection, achieving 96.9% AUC. The final deliverable was a patient risk scorecard that nurses could use daily to triage care, ensuring no patient falls through the cracks.

Download Presentation

Twitter Sentiment Analysis

IE 7860 | Wayne State University

Twitter generates hundreds of millions of tweets daily, each carrying signals that brands, analysts, and researchers want to decode. This project tackled classifying tweet sentiment at scale using 1.6 million labeled tweets. Text was cleaned, tokenized, and stemmed before being fed into four modeling approaches: Logistic Regression, Support Vector Machines, a bidirectional LSTM with GloVe word embeddings, and VADER, a rule-based tool built for social media. UMAP and tSNE projections revealed significant class overlap, confirming that positive and negative sentiment in tweets is genuinely hard to separate, largely due to sarcasm and label noise in the data. The simpler models held up best: Logistic Regression hit 78% accuracy and SVM 77%, while the LSTM topped out at 73.8% after 10 epochs despite considerably more compute. The key takeaway: at this scale, classical NLP pipelines can match or outperform deep learning models, especially under real-world compute constraints.

Download Presentation

Novozymes Enzyme Thermostability Prediction

Kaggle Competition | Wayne State University - DSA 6000

Novozymes challenged the Kaggle community to predict enzyme thermostability (Tm) for mutations of a single wildtype protein. The core difficulty: 31,390 training sequences came from multiple inconsistent lab sources, while the 2,413 test mutations all came from one wildtype that was not explicitly identified in training data. The approach began with identifying the wildtype via consensus sequence, then tagging each training sequence by mutation position and amino acid change. Feature engineering leveraged the R bioseq package for amino acid composition weightings, protein sequence clustering, AlphaFold2 PDB b-factor scores to rank structural stability impact of deletion mutations, and BLOSUM62 substitution scores to quantify amino acid biochemical similarity. Three target formulations were tested: raw Tm, delta Tm from wildtype (dTm), and a normalized dTm/ddG ratio. XGBoost outperformed Random Forest on both Tm (MSE 33.9) and dTm (MSE 20.5). The final ensemble combined model rankings with b-factor and BLOSUM scores to handle deletion mutations, finishing in the top 50% of the leaderboard. A reproducible R-based POC notebook was published on Kaggle.

Download Presentation Download Paper