Yaé Gaba — Data Science Portfolio

13 end-to-end data-science projects on real public data, 5 long-form bilingual case studies (EN/FR), one Streamlit twin.

Domains: forecasting · GLMs / actuarial pricing · survival analysis · hierarchical reconciliation · stochastic optimization / RL · experimental design.

Live site: gabayae.github.io/data-portfolio · Version française

Each project follows the same pipeline: business question → data & EDA → modeling → validation → deployment → business outcome.

---

Projects

#	Project	Domain	Real data source	Headline result
01	Health Supply-Chain Demand Forecasting	Health logistics	USAID PEPFAR SCMS	SARIMA WAPE 0.94
02	Hourly Load Forecasting — PJM (case study)	Energy	PJM Hourly Consumption	GBM MAPE 6.2% (UC PI 99%)
03	Insurance Claims Frequency & Severity	Insurance	freMTPL2 freq + sev	Tweedie Gini 0.310
04	Hierarchical Retail Demand Forecasting	Retail	M5 sample	MinT-OLS reconciliation
05	Stochastic Optimization for Resource Allocation	Health / Ops	Kenya KMPDC + SHA	Q-learning +122% vs manual
06	River Flow Forecasting — Lake Kariba	Hydropower	Kariba Reservoir	GBM RMSE 7 cm
07	Solar Energy Forecasting — Nairobi	Energy	NASA POWER API	GBM MAPE 9.4%
08	Customer Survival Analysis — MTN Nigeria	Telecom	MTN Nigeria Churn	Cox PH + Weibull AFT
09	Flight Demand & Price — Southern Africa	Aviation	SA Flight Prices	SARIMA MAPE 2.0%
10	Property Valuation — Lagos	Real estate	Lagos Housing	GBM R² 0.57
11	Geospatial Farm-Output Forecasting	Agriculture	African Farm Households	GBM R² 0.66
12	Churn Classification — MTN Nigeria	Telecom	MTN Nigeria Churn	XGBoost AUC 0.71
13	A/B Test Framework	Marketing	Marketing Campaign A/B	ANOVA F=21.95, p < 1e-9

---

Case studies

Five projects have a long-form narrative deep-dive — business context, methodology, results, trade-offs, deployment sketch. Each is fully bilingual (EN/FR) with reading time, prev/next nav, breadcrumb schema, and a per-case-study OG card.

Case study	Family	Honest finding
PJM hourly load · FR	Time-series	UC state-space underperforms SARIMA on point error (19.3% vs 14.5% MAPE), but SARIMA's 99% PI coverage is what procurement actually uses.
Lake Kariba river flow · FR	Time-series + exog	7 cm RMSE on a 7 m operational band; turbine discharge as exog is a 2× point-error improvement.
Nairobi solar irradiance · FR	Time-series	Monthly climatology (12.3% MAPE) beats SARIMA (13.8%) — at this latitude the seasonal envelope is most of the predictability.
freMTPL2 pricing · FR	GLMs / actuarial	Tweedie wins on segmentation (Gini 0.310); Poisson + Gamma wins on top-decile lift (2.66×). The choice is actuarial, not technical.
Kenya mobile clinics · FR	Stochastic optimization	Q-learning +122% over manual, but the constraint formulation matters more than the algorithm — capped LP with explicit equity at +39% is more defensible.

Browse all five at /case-studies/ (or /fr/case-studies/).

---

Tech stack

• Languages: Python, SQL
• Time-series: SARIMA, UnobservedComponents (state-space, Kalman filter), Prophet, ETS / Holt-Winters, ARIMA
• ML: scikit-learn (GradientBoostingRegressor, RandomForest, LogisticRegression), XGBoost, LightGBM
• GLM / generalized: statsmodels (Poisson, Gamma, Tweedie, Cox PH, Weibull AFT, OLS)
• Survival: lifelines (KaplanMeier, CoxPHFitter, WeibullAFTFitter)
• Hierarchical reconciliation: MinT-OLS
• Stochastic optimization: Markov Decision Processes, Q-learning, linear programming (scipy.optimize.linprog)
• Experimentation: ANOVA, Welch t-tests with Bonferroni correction, Bayesian A/B (posterior simulation)
• Data engineering: pandas, NumPy, Kaggle CLI, NASA POWER API
• Visualization: matplotlib, seaborn

Domain coverage

• Energy (2): PJM hourly load, Nairobi solar
• Health logistics & operations (2): USAID supply chain, Kenya facility scheduling
• Telecom (2): MTN customer survival, MTN churn classification
• Time-series & forecasting infrastructure (2): hierarchical reconciliation, river flow
• Insurance (1), Aviation (1), Real estate (1), Agriculture (1), Experimentation (1)

How to reproduce

git clone https://github.com/gabayae/data-portfolio.git
cd data-portfolio
make install            # install deps from pyproject.toml
make data               # download all datasets (Kaggle CLI required for some)
make notebooks          # execute every notebook end-to-end

For a single project:

cd <NN-project-name>
pip install -r requirements.txt
python download_data.py
jupyter nbconvert --to notebook --execute notebook.ipynb

Each notebook is self-contained — open it on GitHub and the rendered plots and tables are visible without running anything.

Streamlit twin

portfolio_app.py is an interactive twin of index.html — same content, same color palette, with project filtering and clickable cards. Run it locally:

pip install -r requirements-app.txt
streamlit run portfolio_app.py

To deploy on Streamlit Community Cloud:

1. share.streamlit.io → New app
2. Repository: gabayae/data-portfolio · Branch: main · Main file path: portfolio_app.py
3. Advanced settings → Python requirements file: requirements-app.txt
4. Deploy. Free, public, takes ~1 minute.

portfolio_config.py is the single source of truth for project metadata — edit there and both surfaces update.

Profile photo

Drop a square headshot at profile.jpg in the repo root and the hero avatar on index.html automatically swaps the "YG" monogram for the photo (no code change — handled by onerror="this.remove()" on the <img>). Recommended ≥240×240 px; JPG / PNG / WebP all work.

Contact

• Email: yaeulrich.gaba@gmail.com
• LinkedIn: linkedin.com/in/gabayae
• Personal site: gabayae.github.io
• Google Scholar: profile (h-index 12)
• Book (No Starch Press, 2024): The Shape of Data

License

MIT — see LICENSE.