Forecasting Monthly New Construction Permits in Chicago

1. Project Rationale

As part of my current role as a Real Estate Investment Analyst for a Swiss investment firm, I work on office development strategies across Lisbon, Madrid, and London. One of our recurring challenges is identifying the right acquisition timing based on early signals of local market dynamics.

Among these signals, construction permit issuance is a highly valuable indicator. While we usually rely on brokers to obtain this data on a one-off basis, I wanted to explore whether it's possible to go further and build a model that could anticipate future construction trends, providing a proactive edge to the investment process.

After researching data availability across multiple markets (including Paris, Seattle, and NYC), Chicago emerged as the city with the most complete and extensive building permit dataset publicly available. That made it the ideal ground for developing a robust time series forecasting pipeline.

---

2. Objective

Build and evaluate machine learning models to forecast the monthly number of "new construction" building permits in Chicago, and generate interpretable predictions for the coming months.

---

3. Dataset Overview

• Source: City of Chicago Open Data Portal
• Records: 2006 → 2025 (~1M total permits, ~30k for new construction)
• Selected Target: PERMIT - NEW CONSTRUCTION
• Date Field: ISSUE_DATE

---

4. Analytical Pipeline

Data Preparation

• Filtering to new construction only
• Monthly aggregation of permit counts
• Creation of lag, rolling, and calendar-based features

📁 Output: data/processed/monthly_features.csv

Exploratory Visualization

plt.plot(df["year_month"], df["permit_count"])
plt.title("Monthly Permits Issued")

Feature Engineering

monthly["rolling_mean_3"] = monthly["permit_count"].rolling(3).mean()
monthly["lag_1"] = monthly["permit_count"].shift(1)
monthly["month"] = monthly["year_month"].dt.month

Modeling

Three models were tested:

Model	MAE	RMSE
Random Forest	10.75	14.74
XGBoost	11.93	15.51
Prophet	13.93	16.85

Final model selected: RandomForestRegressor for best accuracy and explainability.

Model Interpretation

Feature importance:

feat_importances.plot(kind="barh")

Residual analysis:

plt.plot(test["year_month"], test["residuals"])

Forecasting 2025 (with Prophet)

To visualize a realistic forward-looking forecast, Prophet was used to generate predictions from May → December 2025.

plt.plot(df_prophet["ds"], df_prophet["y"], label="Historical")
plt.plot(forecast_future["ds"], forecast_future["yhat"], label="Forecast")

---

5. Quickstart Guide

Requirements

pip install -r requirements.txt

Run notebooks in order:

notebooks/
├── 01_data_exploration.ipynb
├── 02_feature_engineering.ipynb
└── 03_model_training.ipynb

Predict future months (optional)

You can extend predictions to future months by rerunning the Prophet cell in 03_model_training.ipynb.

---

6. Repository Structure

chicago-construction-ts-modeling/
├── data/
│   ├── raw/                 # Original CSV
│   └── processed/            # Cleaned & aggregated
├── notebooks/             # All Jupyter steps
├── visuals/               # Plots & exportable assets
├── requirements.txt
└── README.md

---

7. Next Steps

• Apply this pipeline to European cities if similar data becomes available
• Build a simple Streamlit dashboard for real-time predictions
• Expand to classify permit types and model spatial distributions

---

8. About Me

This project was developed by Sacha Brouck, a data-driven real estate analyst preparing to join the University of Washington MSBA program. My goal is to create professional, rigorous, and interpretable machine learning projects rooted in real-world investment use cases.

LinkedIn: Sacha Brouck
GitHub: SBrouck