These projects translate research workflows into decision tools for pricing, market intelligence, machine learning, and regional measurement. Three independent products anchor the portfolio: a retail pricing engine, a historical price-intelligence system with explicit model-governance decisions, and an interactive population-density atlas.

Pricing and Market Intelligence Tools

Pricing and Promotion Decision Engine (2025–Present)

• Live app: pricing-promotion-decision-engine.streamlit.app
• Code: github.com/Li-Shen-Clark/pricing-promotion-decision-engine
• Stack: Python (pandas, statsmodels, scikit-learn), Streamlit, pytest, Git/GitHub
• Narrative role: Industry translation of the estimate–simulate–optimize workflow used in my regional trade and market-integration research.

An end-to-end demand-estimation and decision-support tool built on 6.6M rows of Dominick’s Finer Foods scanner data (cleaned panel of 486 UPCs, 93 stores, 366 weeks). The workflow is the same one I use in my academic work, applied to product-level pricing instead of regional GDP.

• Estimate behavioral response: log-log fixed-effects demand model with Duan smearing retransformation to reduce log-to-level retransformation bias.
• Simulate counterfactuals: price and promotion counterfactual simulations over the estimated demand system.
• Optimize under constraints: constrained profit optimization under cost, inventory, and margin limits; A/B validation plans with power analysis before any candidate action ships.
• Ship: deployed as an interactive Streamlit application backed by a modular Python codebase, a pytest test suite, reproducible requirements, and auto-deploy from GitHub.

Vegetable Price Intelligence — Forecasting, Risk, and Model Governance (2026–Present)

• Portfolio case study: li-shen-clark.github.io/vegetable-price-intelligence
• Code: github.com/Li-Shen-Clark/vegetable-price-intelligence
• Stack: Python, pandas, Polars, scikit-learn, statsmodels, pytest, Git/GitHub
• Narrative role: Historical price forecasting, risk flagging, and evidence-based decisions about which features are reliable enough to publish.

Built an economics-informed price-intelligence system from 8.68M historical wholesale-market records across 117 cities and 30 vegetables. The system produces 7.36M auditable city-day facts with versioned mappings, quality flags, and coverage tiers. It is distinct from the HSR market-integration paper: this project focuses on forecasting and model governance rather than estimating how transport connectivity relates to price dispersion.

• Forecast: rolling-origin 7/14/28-day forecasts against persistence and seasonal baselines reduced overall WAPE from 18.72% to 17.54% on a sealed 16,217-task final test; 10 weak product-horizon groups route to explicit baseline fallback.
• Flag risk: a validation-frozen 14-day price-spike classifier reached PR-AUC 0.2204 versus 0.0775 for the strongest simple baseline, with 7.18 days of average historical lead time for review.
• Withhold weak signals: stress-tested 2,736 predeclared city lead-lag candidates and withheld the directional network when only 4 of 15 frozen edges improved out of sample.

Machine Learning Competitions

Kaggle Silver Medal — Home Credit Credit Risk Model Stability (Feb–May 2024)

• Stack: Python, Polars, scikit-learn, LightGBM, CatBoost
• Role: Statistical modeling and co-developer

Competition focused on model stability over time (Gini Stability metric) rather than raw AUC — a realistic production concern for credit-risk models that retrain on drifting data.

• Co-developed a Polars lazy-execution ETL pipeline for multi-depth relational data, enabling fast iteration on feature definitions across parent-child tables.
• Designed aggregation features (max, last, mean) capturing borrower behavioral trends; these were adopted across the team codebase.
• Optimized directly for the headline Gini Stability metric to prevent performance drift over time; implemented correlation-based feature grouping (Pearson > 0.8) for multicollinearity control and type-downcasting memory optimization (over 50% reduction).

Kaggle Bronze Medal — ICR Identifying Age-Related Conditions (May–Aug 2023)

• Stack: Python, TabPFN, XGBoost, imbalanced-learn
• Role: ML researcher and co-developer

Competition on a small, severely class-imbalanced tabular dataset with auxiliary metadata (Greek) hinting at latent subgroup structure.

• Designed a stratified resampling strategy using the Greek metadata (RandomOverSampler on Alpha-defined latent subgroups) to preserve population structure under severe class imbalance.
• Co-developed a voting ensemble combining parametric (XGBoost) and non-parametric (TabPFN) models under a custom Balanced Log Loss metric.
• Calibrated decision thresholds to prioritize recall in a severely imbalanced classification setting.

Geospatial Measurement Infrastructure

China Population Density Atlas (2022–Present)

• Live app: pop-density-spatial.streamlit.app
• Stack: Python (GeoPandas, rasterio, NumPy), Streamlit
• Companion paper: Population Agglomeration and Spatial Reallocation in China (SSRN preprint)
• Narrative role: Supporting capability — demonstrates geospatial data construction, visual analytics, and regional measurement infrastructure that underpin my regional-economics research.

Interactive geospatial visualization of China’s population geography built from 19 years of 5 km-resolution population-density rasters (2002–2020) harmonized into an annual panel.

• Processed WorldPop rasters covering the full territory of China into grid-cell and province-level summaries.
• Surfaced documented patterns from the underlying research: rising spatial Gini, expansion of dense urban corridors, and a 17 km southeastward shift in the population-weighted centroid.
• Deployed in 2025 for map-based exploration of density, agglomeration, and urban-corridor metrics at multiple spatial scales.

Toolkit

• Programming: Python (pandas, NumPy, scikit-learn, statsmodels, GeoPandas, Polars), SQL, R, MATLAB; Git/GitHub, pytest.
• Machine learning: Gradient boosting (XGBoost, LightGBM, CatBoost), cross-validation, feature engineering, model stacking, regularization.
• Deployment: Streamlit (multiple live apps), reproducible requirements, auto-deploy from GitHub.
• Data at scale: Retail scanner panels, multi-million-row price panels, high-frequency commodity price data, geospatial rasters, multi-regional input-output tables.