An interactive web visualization and statistical modeling project submitted to the SBCAG Code4Good 2026 Datathon. The project applies open civic data to expose structural barriers to housing production in Santa Barbara County, in the spirit of the Data4Good initiative: using data science for civil and social benefit.

Full analysis report: Code4Good 2026/Datathon.pdf

Santa Barbara County faces a compounding housing affordability crisis. This project investigates two interconnected questions:

1. Where is housing production failing? Using California's mandatory Housing Element reporting data, a Housing Friction Rate metric is derived per jurisdiction — quantifying what share of the housing pipeline (entitlements → permits → certificates of occupancy) never results in completed, habitable units.

2. Can housing output predict cost burden? A predictive modeling analysis in R tests whether income-stratified permit and occupancy data can forecast the average percent of households that are cost-burdened (spending >30% of income on housing).

Located in Code4Good 2026/Datathon.Rmd | Compiled report: Code4Good 2026/Datathon.pdf

Data sources:

Methodology:

The two datasets are merged on jurisdiction and year, filtered to 2017–2021, and split 75/25 for training and testing. Six regression models are trained via 10-fold cross-validation to predict percent_avg_burdened (average cost burden rate) from income-stratified CO counts and aggregate cost burden totals:

All predictors are normalized. Hyperparameters are tuned over regularized grids and evaluated by RMSE.

Results:

KNN outperformed all other models on cross-validation RMSE and was selected as the final model. Final model performance was validated on the held-out test set via last_fit(), with residual diagnostics (residuals vs. predicted, Q-Q plot, actual vs. predicted) confirming reasonable fit with no severe systematic bias.

Key finding: Income-stratified housing production data (particularly very low and low income completions) shows meaningful correlation with cost burden rates, supporting the case that failing to complete affordable housing directly drives cost burden.

A choropleth dashboard rendering the Housing Friction Rate across 28 Santa Barbara County jurisdictions.

frictionRate = 1 − (COs ÷ max(entitlements, permits, COs))

Measures what fraction of the housing pipeline never reached completion. Regions with fewer than 20 total pipeline units, no permits issued, or where COs exceed pipeline start (cross-year artifact) are excluded.

Selected results:

• Click any region or label to view a step-by-step equation breakdown in the sidebar
• Layer toggle between base map and friction rate heatmap
• Expandable description panel explaining the metric and its Data4Good context
• Dynamic region count reflecting only jurisdictions with valid friction rate data
• Region search with highlight

A dynamic backend service providing localized AI recommendations and real-time model inference.

Leverages a Retrieval-Augmented Generation (RAG) pipeline powered by Llama-3.3 (via Groq) to provide context-aware recommendations:

• Where to Build: Analyzes workforce signals and affordability metrics to recommend jurisdiction targets for new affordable housing.
• Worker Matching: Matches specific job titles and income levels to viable Santa Barbara County communities, incorporating live workforce data.

Serves the optimal predictive model (exported as .onnx) from the R statistical analysis, exposing a /predict endpoint that outputs expected cost burden rates given hypothetical housing pipeline metrics.

datathon-project/
├── Code4Good 2026/
│   ├── Datathon.Rmd           # Full statistical analysis (R / Tidymodels)
│   └── Datathon.pdf           # Compiled PDF report
├── scripts/
│   ├── generate_geojson.mjs   # Downloads Census shapefiles, filters to SB County
│   └── parse_pfs.mjs          # Aggregates PFS CSV into per-region friction metrics
├── src/
│   ├── components/
│   │   ├── Dashboard.jsx       # Layout, layer state, region analysis panel
│   │   ├── SantaBarbaraMap.jsx # Choropleth map, annotations, zoom, click handlers
│   │   └── MapTooltip.jsx      # Hover tooltip
│   ├── data/
│   │   ├── santa_barbara_county.json  # County boundary GeoJSON
│   │   ├── santa_barbara_places.json  # 28 places GeoJSON
│   │   └── heatmap_data.json          # Per-region friction metrics (generated)
│   └── index.css              # Dark theme, glassmorphism design system
├── backend/
│   ├── main.py                # FastAPI server and endpoints
│   ├── rag_pipeline.py        # RAG implementation for AI Advisor
│   └── requirements.txt       # Python dependencies
├── burden_model.onnx          # Exported predictive model for inference
├── PFS_Datathon.csv           # Source: ~9,287 rows, CA HCD PFS reports
├── package.json
└── README.md

node scripts/generate_geojson.mjs

Downloads US Census TIGER/Line 2022 shapefiles and filters to Santa Barbara County jurisdictions. Run once; outputs are committed.

node scripts/parse_pfs.mjs

Aggregates PFS_Datathon.csv by jurisdiction, computing total entitlements, permits, and COs. Applies three validity guards before computing a friction rate:

1. Cross-year artifact — COs ≥ pipelineStart excluded (completions arrived from permits issued before the dataset window)
2. No permits issued — entitlements-only regions excluded (construction friction requires permits)
3. Insufficient volume — pipelines with fewer than 20 units excluded as statistically unreliable

Node.js 18+ (visualization); Python 3.9+ (backend); R 4.x with the libraries listed in Datathon.Rmd (analysis).

cd backend
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --reload

Requires a .env file with GROQ_API_KEY. The API runs at http://localhost:8000.

npm install
npm run dev

Open http://localhost:5173.

node scripts/parse_pfs.mjs

npm run build

• Timing lag: The friction rate treats the dataset as a single cumulative pool rather than tracking annual cohorts. COs issued for permits from prior years can distort rates — this is why Goleta is excluded (941 COs from 692 permits). A more precise implementation would track units by permit year.
• Predictive model scope: The R analysis covers 2017–2021 and is limited to jurisdictions present in both the HCD and affordability datasets. Results should be interpreted as exploratory rather than causal.
• Academic basis: Friction rate methodology is consistent with "entitlement-to-construction conversion rate" analysis used by the UC Berkeley Terner Center and Urban Institute.

Private — SBCAG Code4Good 2026 Datathon submission.