Applied Projects

These projects apply economic models and empirical methods to pricing, forecasting, and spatial analysis. Each project links to a demonstration, case study, or source code so that readers can examine the work in more detail.

Pricing and Promotion Decision Engine

Problem. How can demand estimates inform pricing decisions when costs, inventory, and margins constrain the available choices?

Work. The application uses 6.6 million raw Dominick’s Finer Foods scanner records, yielding a cleaned panel of 4.65 million rows covering 486 UPCs, 93 stores, and 366 weeks. It combines a log-log fixed-effects demand model with price and promotion simulations, constrained profit optimization, and experimental validation plans.

What it demonstrates. Demand estimation, simulation, optimization, power analysis, and deployment of an interactive research application. The observational estimates support conditional scenarios; they are not experimental estimates of causal price response. A/B tests are proposed validation steps, and simulated gains are not measured business outcomes.

Tools. Python, pandas, statsmodels, scikit-learn, Streamlit, and pytest.

Project details · Open the application · Source code

Vegetable Price Intelligence

Problem. How can historical wholesale prices support forecasting and risk assessment when coverage and predictive performance vary across markets?

Work. The system organizes 8.68 million historical records across 117 cities and 30 vegetables into 7.36 million city-day observations with documented mappings and quality flags. It evaluates 7-, 14-, and 28-day forecasts against persistence and seasonal baselines, and studies 14-day price-spike signals.

Evaluation. On the project’s held-out historical test, overall weighted absolute percentage error falls from 18.72% to 17.54%. Ten weak product-horizon groups use a baseline fallback. A proposed directional network was withheld after only 4 of 15 selected edges improved out of sample. These are historical evaluation results, not live forecasts or guaranteed future performance.

What it demonstrates. Data validation, evaluation across time, baseline comparisons, and decisions about when a model is reliable enough to use.

Tools. Python, pandas, Polars, scikit-learn, statsmodels, and pytest.

Case study · Source code · Related market-integration research

China Population Density Atlas

Problem. How can changes in population concentration be compared across places and years?

Work. Annual WorldPop rasters for 2002–2020 are harmonized at approximately 5 km resolution. The interactive atlas presents density, agglomeration, and urban-corridor measures at multiple geographic scales.

What it demonstrates. Geospatial data processing, consistent measurement, and interactive communication of descriptive research.

Tools. Python, GeoPandas, rasterio, NumPy, and Streamlit.

Open the atlas · Companion research

Machine learning competitions

Home Credit — Credit Risk Model Stability

Kaggle silver medal, 2024. As a team contributor, I co-developed data processing with Polars, designed aggregation features, and worked on time stability, feature grouping, and memory efficiency using LightGBM and CatBoost.

ICR — Identifying Age-Related Conditions

Kaggle bronze medal, 2023. As a team contributor, I worked on class imbalance, resampling, model ensembles, and threshold selection using TabPFN, XGBoost, and imbalanced-learn.

Research · Blog · Economist CV