Mai Hakamada, PhD - Credit Growth by RF&XGBoost

Predicting Private Credit Growth (with Lag Structure) Using Machine Learning Models: Random Forest vs. XGBoost

Objective:

This project aims to predict next-year private credit growth (% of GDP) across countries using current-year macroeconomic variables.
By applying a lag structure, the model simulates a forward-looking framework for understanding credit dynamics—a perspective relevant for financial stability analysis and macroprudential oversight.

Data

Data Sources:

World Bank World Development Indicators

Time Period and Coverage:

Years: 1991–2023
Countries: Over 30 countries (adjust depending on the actual number)

Target Variable:

Private sector credit growth in year t+1

This prediction task incorporated a 1-year lag, meaning macro variables from year t were used to predict credit growth in t+1.

This approach aligns with practical applications in macroprudential surveillance, allowing the model to anticipate financial risks with a policy-relevant lead time.

Explanatory Variables (year t):

Real GDP growth
Inflation rate
Unemployment rate
Government expenditure-to-GDP ratio
Trade openness ((exports + imports)/GDP)

Models and Methodology

Random Forest (ntree = 100)
XGBoost (nrounds = 50, objective = "reg:squarederror")
Evaluation Metric: Mean Squared Error (MSE)

Results

Model Fit Comparison: Random Forest vs. XGBoost
Actual vs Predicted Credit Growth Plots:

Random Forest:

Strong alignment with the 45-degree ideal line
Stable performance in both low and high-growth ranges (0–150%)
Robust to outliers, with minimal overfitting

XGBoost:

Moderate alignment in central observations
Greater dispersion in high-growth areas (>150%), suggesting potential overfitting
Tends to underpredict in the mid-to-high range

Model Performance Comparison (MSE):

Random Forest: 207.63

XGBoost: 349.76

Random Forest outperformed XGBoost in predicting next-year private credit growth, suggesting it better captured the underlying nonlinear relationships across countries.

Feature Importance (Top Predictors):

Random Forest:

Each Variable Importance (IncNodePurity)

Trade openness: 2,353,766

Inflation: 2,298,908

Government expenditure: 2,153,340

Unemployment:1,809,099

GDP growth: 1,400,043

XGBoost:

Each Variable Gain

Trade openness: 0.281

Inflation: 0.240

Unemployment: 0.212

Government expenditure: 0.193

GDP growth: 0.075

Notably, trade openness and inflation emerged as consistent top predictors across both models, indicating their forward-looking relevance to credit supply behavior.

Insights

Lagged macroeconomic variables can effectively predict next-year credit growth, confirming the value of a forward-looking approach.
Trade and inflation signals in particular may anticipate changes in credit dynamics.
Random Forest proved more robust than XGBoost, possibly due to better handling of cross-country heterogeneity and noise.
This model structure could support early warning systems or macro-financial surveillance frameworks.