Bernard Frimpong

Data Analyst · Akron, Ohio

I turn messy data into results you can check.

Applied mathematician working in credit analytics, survival analysis and disease forecasting. Every project here is reproducible on GitHub, validated honestly, and written up in plain language.

Now Submitting weekly CDC FluSight forecasts and drafting a paper on turning points →

Bernard Frimpong
Open to data analyst roles
882,533SBA loans analyzed
0.754IPCW C-index, unseen hospital
WeeklyCDC FluSight submissions
0.96056Kaggle ROC AUC

About

I hold an M.S. in Applied Mathematics from the University of Akron and a B.S. in Mathematics from KNUST in Ghana. My training is in modeling, numerical methods and statistics, and I now apply it to business and health data.

I care most about validation that holds up: grouped cross-validation, leakage checks, and scores on data the model has never seen. Three years of university teaching taught me to explain those results so a non-specialist can act on them.

Projects

Commercial Credit Portfolio Analysis

2026
loans approved 2005–08 → 21% of dollars, 54% of losses
Share of loan dollars vs. share of losses by approval year

A bank-style review of 882,533 SBA small-business loans ($170B approved), built in SQL and Python and delivered as an Excel limit monitor, a Power BI dashboard and an executive summary.

  • An early warning indicator from 2-year default rates flagged the 2010 loss peak three years ahead, two years before the unemployment rate did.
  • The Excel limit monitor found three concentration breaches; lender rankings were corrected with an actual-vs-expected benchmark.
  • SQL
  • Python
  • Power BI
  • Excel
View on GitHub →

Leukemia Overall Survival Prediction

Oct 2026
IPCW C-index 0.754 on an unseen hospital · clinical score IPSS-R 0.671
Kaplan–Meier survival curves by predicted risk group

QRT × Institut Gustave Roussy challenge (Challenge Data, ENS Paris): predicting survival of myeloid leukemia patients from blood counts, karyotype text and gene mutations.

  • 91 engineered features, including a parser for free-text karyotypes; penalized Cox, random survival forest and XGBoost-Cox models.
  • Centre-grouped nested cross-validation: the ensemble scored 0.725 ± 0.021 and beat the best single model on 15 of 15 folds.
  • Kaplan–Meier and the IPCW C-index derived by hand and matched to library output; 33-page LaTeX write-up.
  • lifelines
  • scikit-survival
  • XGBoost
  • LaTeX
View on GitHub →

CDC FluSight Influenza Forecasting

Oct 2026 – present · ongoing
trend model: −17% growth-phase error (WIS) vs baseline
Forecast intervals vs. reported flu hospitalizations

Weekly probabilistic forecasts of state-level flu hospitalizations for the CDC FluSight 2026–27 challenge. Every submission passes the hub's automated validation.

  • Baseline, damped-trend and mechanistic SIRS models (RK4 solver, weekly reproduction number), backtested on 2025–26 by epidemic phase.
  • The SIRS model matched the baseline overall with near-zero growth-phase bias and exposed systematic overshoot at peaks.
  • Python
  • ODE models
  • WIS
View on GitHub →

Airline Passenger Satisfaction

Kaggle Playground S6E10 · Oct 2026
ROC AUC 0.95836 → 0.96056 over six rounds · CV within 0.0005 of leaderboard
Cross-validation and leaderboard AUC across six rounds

Six documented rounds of feature engineering, fold-safe target encoding and hill-climbing model blends. A deliberate leakage demo showed leaky encoding inflating CV by 8× the real gain.

  • LightGBM
  • XGBoost
  • CatBoost
View on GitHub →

Research

Do influenza forecasts fail at turning points?

In preparation · target: medRxiv preprint, then journal

A phase-stratified evaluation of the CDC FluSight models and ensemble, 2023–2026: how much worse they do near epidemic peaks, how late they detect the turn, and how far they overshoot.

Computational model of flow near permeable boundaries

Master's thesis · University of Akron · 2025–2026

Numerical simulation of Stokes flow with the method of regularized Stokeslets, validated against analytical solutions with a study of accuracy and convergence.

Writing

Testing a survival model on a hospital it has never seen

23 Oct 2026 · survival analysis

Why I grouped cross-validation by hospital, and what the C-index actually measures.

How a leaky encoding made my CV score lie by 8×

20 Oct 2026 · machine learning

The same feature encoded two ways, and why only one told the truth.

All posts →

Toolkit

Programming & data

Python (pandas, NumPy, SciPy, Matplotlib), SQL (DuckDB), R, MATLAB, Git, LaTeX

Machine learning

scikit-learn, LightGBM, XGBoost, CatBoost; grouped and nested CV, target encoding, blending, permutation importance

Statistics & modeling

Survival analysis, regression, hypothesis testing, time-series forecasting, SIRS/ODE models, numerical methods

Credit & reporting

Vintage analysis, default and loss rates, HHI concentration limits, early warning indicators; Power BI (DAX), Excel

Experience

Mathematics Instructor

T-Squared Honors Academy-North · Aug 2026 – present

Teach Algebra 1, Algebra 2 and Geometry; track assessment results to find learning gaps.

Graduate Teaching Assistant

University of Akron · 2024 – 2026

Recitations, grading and office hours for undergraduate calculus and statistics.

Teaching Assistant

KNUST, Kumasi, Ghana · 2022 – 2024

Tutorials and grading for undergraduate mathematics courses.

Contact

I'm looking for data analyst and credit analytics roles. If you'd like to talk about a role or one of these projects, email me.