ML_RecessionForecast

Real-time recession risk forecasting using machine learning

---

Overview

RecessionRiskForecast is a Python-based system that predicts the probability of a U.S. recession in the next 6 months using macroeconomic indicators from FRED and machine learning. The system employs walk-forward cross-validation and XGBoost with calibrated probabilities to provide robust, out-of-sample forecasts for rare economic events.

• Precision-Recall AUC: 0.1221 (primary metric)
• ROC-AUC: 0.5522
• Forecast Horizon: 6 months ahead
• Input: Monthly macroeconomic data from FRED
• Output: Probability of recession + visualizations

This project demonstrates practical applications of time series ML, feature engineering, and calibrated probability forecasting for financial decision-making.

---

Features

• Fetch monthly economic series (unemployment, CPI, yields, consumer confidence, industrial production, recession indicator) from FRED

• Engineer lagged and derived features (yield curve, Sahm rule, CPI YoY, industrial production MoM)

• Train XGBoost classifier with Optuna hyperparameter tuning for each walk-forward fold

• Calibrate predicted probabilities using isotonic regression for rare event reliability

• Generate walk-forward out-of-sample predictions and evaluate using precision, recall, F1-score

• Visualizations:

  • Forecasted recession probability over time
  • Precision-Recall curve
  • Sahm Rule indicator chart

---

Installation

1. Clone the repository:

git clone https://github.com/yourusername/RecessionRiskForecast.git
cd RecessionRiskForecast

2. Create a virtual environment (optional but recommended):

python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

4. Create a .env file with your FRED API key:

FRED_API_KEY=your_fred_api_key_here

Key Dependencies:

• Python 3.10+
• pandas, numpy, matplotlib
• FRED API (fredapi)
• scikit-learn (sklearn)
• xgboost (xgboost)
• optuna
• python-dotenv

---

Usage

1. Fetch & Prepare Data

The script fetches U.S. macroeconomic series from FRED and resamples them monthly:

python main.py

Series fetched:

• Unemployment rate (UNRATE)
• CPI (CPIAUCSL)
• 10Y and 2Y Treasury yields (DGS10, DGS2)
• Consumer Confidence (UMCSENT)
• Industrial Production (INDPRO)
• Recession indicator (USREC)

---

2. Feature Engineering

Derived and lagged features include:

• Yield curve: 10Y − 2Y
• Sahm Rule: 3-month unemployment MA − 12-month unemployment min
• CPI year-over-year change
• Industrial production month-over-month change
• Lagged features for 1, 3, 6, 12 months
• Target: any recession within the next 6 months

---

3. Walk-Forward Cross-Validation

Evaluate the model using walk-forward CV to ensure robust, out-of-sample predictions:

• XGBoost classifier with Optuna hyperparameter tuning
• Calibrated probabilities (isotonic or sigmoid)
• Predictions threshold: 0.4 for rare events
• Outputs classification report with precision, recall, F1-score, PR-AUC, ROC-AUC

---

4. Final Model Training

Train the model on the entire clean dataset to generate current recession risk and top predictors.

---

5. Visualization

Generates a single chart recession_forecast.png that combines:

• Forecasted recession probability over time
• Precision-Recall curve
• Sahm Rule labor market stress indicator

python main.py

Example:

---

Results

• Walk-Forward Classification Accuracy: 65.8%
• Precision-Recall AUC: 0.1221 (primary metric for rare events)
• ROC-AUC: 0.5522
• Current recession probability (next 6 months): 3.8%
• 95% bootstrap CI: [3.5%, 4.1%]

Top predictors:

yield_curve_lag12
sahm_rule_lag1
sahm_rule
indpro_mom
CPI
yield_curve_lag6
cpi_yoy_lag12
10Y
CONF_lag1
unrate_min_12

---

Project Structure

RecessionRiskForecast/
├─ main.py                 # Main script: fetch, feature engineer, train, visualize
├─ .env                    # FRED API key
├─ recession_forecast.png  # Combined chart with probability, PR curve, Sahm rule
├─ README.md               # Project documentation
└─ requirements.txt        # Python dependencies

---

Future Work

• Extend to shorter horizon forecasts (1–3 months)
• Add macro sentiment or alternative data features
• Explore interactive dashboards for financial analysts
• Test alternative ML models (LightGBM, TabNet) and ensemble approaches