FDOR Real Estate Investment Analysis

A machine learning project analyzing 10 years of Florida Department of Revenue tax roll data to predict profitability of single-family home resales.

Project Overview

This project processes 634 GB of Florida property assessment data to build predictive models that forecast whether a single-family home will generate a profit or loss when resold within the same calendar year. The analysis focuses on identifying key factors that drive investment outcomes in the Florida real estate market.

Business Problem

Real estate investors and iBuyer companies need to quickly assess whether a property purchase will be profitable. This project aims to:

• Identify the strongest predictors of property resale profitability
• Build models that can flag high-risk investments
• Provide data-driven insights for investment strategy optimization

Data Source

Florida Department of Revenue (FDOR) Assessment Tax Rolls

• Time Period: 10 years of historical data
• Size: 634 GB across 700+ files
• Scope: All 67 Florida counties
• Records: 10+ million property parcels

Technical Approach

1. Data Engineering

• Challenge: 700 separate files requiring consolidation
• Solution:
  • Command-line automation scripts for file extraction and processing
  • Custom SQL Server database build to bypass import limitations
  • Structured schema design for efficient querying

2. Exploratory Data Analysis

Tools: SQL Server

Key analyses performed:

• Parcel counts and geographic distribution
• Identification of institutional buyers (iBuyers like OpenDoor, Offerpad)
• Transaction volume and timing patterns
• Price trends by region and property type

3. Machine Learning Models

Tools: R, caret package, randomForest

Models Built:

1. Logistic Regression - Baseline binary classifier
2. Random Forest - Advanced ensemble method

Key Techniques:

• Stratified sampling to maintain class distribution
• SMOTE (Synthetic Minority Over-sampling Technique) to address class imbalance
• Cross-validation for model evaluation
• Feature importance analysis

Model Performance:

• Random Forest achieved high recall for loss identification
• Successfully flags high-risk properties for further review

Key Findings

Primary Insight

Geographic location is the single strongest predictor of profitability. The calculated "subdivision risk score" outperformed all property-specific features.

Secondary Insights

• Property characteristics (size, age, condition) have limited predictive power
• Market timing within a calendar year shows minimal impact
• The profit/loss prediction problem is inherently difficult, even with extensive historical data
• iBuyer activity patterns are identifiable in the data

Technologies Used

• Database: SQL Server
• Programming: R
• Key R Packages:
  • caret - Model training and evaluation
  • randomForest - Random Forest algorithm
  • DMwR - SMOTE implementation
  • dplyr - Data manipulation
• Scripting: Bash (command-line automation)

Business Applications

This analysis framework can support:

• Investment Risk Assessment - Flag high-risk properties before purchase
• Portfolio Optimization - Focus acquisitions on high-probability-profit areas
• Market Intelligence - Track competitor (iBuyer) activity patterns
• Strategic Planning - Identify geographic markets with best risk/reward profiles

Future Enhancements

• Power BI Dashboard - Interactive visualization of risk scores and predictions
• Python Migration - Rebuild models using scikit-learn for broader deployment
• Real-time Scoring API - Deploy model as web service for instant predictions
• Expanded Features - Incorporate economic indicators, crime data, school ratings
• Time Series Analysis - Add forecasting for property value appreciation

Video Presentation

For a detailed walkthrough of the project methodology and findings: Watch Project Presentation

Results & Impact

Technical Achievement:

• Successfully processed and modeled 634 GB of complex government data
• Built production-ready database from raw files
• Demonstrated advanced techniques for handling imbalanced datasets

Analytical Achievement:

• Identified geographic location as primary profitability driver
• Created actionable risk scoring system
• Proved concept viability despite problem difficulty

Transferable Skills:

• Government data expertise
• End-to-end analytics workflow (data engineering → modeling → insights)
• SQL database design and optimization
• Statistical modeling and machine learning
• Business problem translation to technical solution

Author

Ruslan Dubas

• GitHub: @RusUsf
• Project Repository: FDOR_Data_and_Models
• LinkedIn: https://www.linkedin.com/in/ruslan-dubas-861a7799/

Acknowledgments

• Florida Department of Revenue for public data access
• Real estate domain experts who provided business context

---