Charity Donation Prediction with Machine Learning

Project Overview

This project applies machine learning to help a charity organization optimize its direct mail fundraising strategy, reducing cost while maximizing return on investment. The goal was two-fold:

Classify potential donors (who will donate)
Predict donation amounts for identified donors

Model Strategy

Two separate models were created:

1. Random Forest Classifier

Predicts whether a person will donate (donr = 1/0)
Achieved 88.85% accuracy on the test set
Identified 408 donors out of 2007 samples

2. Linear Regression Model

Predicts donation amount (damt) for donr = 1
Achieved R² = 0.56 and MAE = $0.99
Accurate, cost-efficient way to forecast campaign returns

Data Insights

Features grouped by: demographics, wealth, donation history
Strong predictors: lifetime giving (lgif, rgif), property value (avhv), income (incm)
Most monetary values were right-skewed → normalized/scaled during preprocessing
Heatmaps showed low correlation between wealth & donation, confirming need for advanced models

Model Development

Tools: Python, Pandas, scikit-learn
Preprocessing:
- Removed irrelevant features (ID, damt, donr)
- Feature scaling with StandardScaler
- Used separate training/test sets
Pipeline:
- Classification first (filter donors)
- Then regression on predicted donors only

Business Value

💡 Strategy 1: Focus on Top 25% Predicted Donors

102 donors selected
Avg donation: $50.23
ROI: 2,411.5%
Net profit: $4,919.46

💡 Strategy 2: Mail to All 408 Predicted Donors

Avg donation: $12.09
ROI: 504.2%
Net profit: $4,113.12

Traditional blanket mailing: Net loss of $1,099.50

Results

Classification Model: Random forest - due to its ability to handle complex feature interactions and better accuracy
Prediction Model: Linear Regression - simplicity, interpretability, and efficiency in modeling relationships between predictors and the target variable

Model Selection

Predict the donation amounts (damt) for donors identified through the classification model
Provide accurate predictions of donation amounts to optimize resource allocation and enhance the charity’s return on investment

Prediction Model

The proportion of donors in the dataset is low, with the majority of the sample being non-donors (1589) and only a small proportion of donors (418)
Most of the donations are in the lower range ($0-50), with only a few donations in the higher range.

Predict the values for Scoring Data