Reach for Change — Predicting Donor Response to Optimize Outreach

A binary classification project to predict whether a potential donor will respond positively to a fundraising campaign, enabling a nonprofit federation to target outreach more precisely and reduce unnecessary contact.

---

Context

The Civic Support Alliance (CSA) is a federation representing multiple humanitarian and social aid programs. As donor fatigue has become an increasing challenge for nonprofits, the organisation sought to move away from blanket outreach and instead identify — in advance — which individuals are likely to donate if contacted.

This project builds a predictive system to answer a single core question: will this person donate if contacted?

---

Dataset

The dataset contains historical records from 13,561 individuals contacted in a previous fundraising campaign, including:

• Demographic features — age, neighbourhood-level socioeconomic indicators (median home value, household income), military affiliation
• Donation history — lifetime gift amounts, recency, frequency, average gift size
• Campaign interaction — number of past promotions, response rates, card response counts
• Target variable — TARGET_B: 1 if the person donated in the last campaign, 0 otherwise

The dataset exhibits significant class imbalance (~25% donors, ~75% non-donors), which drove several modelling decisions.

---

Project Structure

.
├── data/
│   ├── donors_train_target.csv     # Training labels
│   ├── sample_submission.csv       # Submission format reference
│   └── test.csv                    # Test set (no labels)
├── Donors_Predictive_Final.ipynb   # Main notebook
└── README.md

---

Methodology

1. Exploratory Data Analysis & Preprocessing

• Identified and handled missing values through imputation
• Detected and corrected outliers, misclassifications, and data incoherencies
• Analysed class imbalance and its implications for model selection

2. Feature Selection

A multi-method approach was applied to identify the most informative features:

Method	Variable Type
Chi-Square test	Categorical
Variance Threshold	Numerical
Spearman Correlation	Numerical
Decision Tree Importance	Numerical
Recursive Feature Elimination (RFE)	Numerical
Lasso Regularisation	Numerical

3. Model Assessment Strategy

A holdout split was used for model evaluation. Five classifiers were trained and compared using F1 Score (class 1) as the primary metric, given the class imbalance:

Model	Notes
K-Nearest Neighbours	Susceptible to class imbalance; distance metric less optimal for this feature space
Decision Tree	Prone to overfitting; lower stability
Random Forest	Strong in theory but unable to surpass Logistic Regression after tuning
Neural Network (MLP)	Better suited for larger datasets; struggled with imbalance
Logistic Regression ✓	Best performer — selected as final model

4. Optimisation

Hyperparameter tuning was performed via GridSearchCV. The final Logistic Regression configuration achieved:

• Class 1 F1: 0.40
• Class 1 Recall: 0.55
• Kaggle F1 Score: 0.4279

The high recall for class 1 was a deliberate trade-off — correctly identifying real donors matters more than avoiding false positives in this context.

---

Key Results

Logistic Regression outperformed all other candidate models on every relevant metric. The final model correctly identified 373 true donors on the test holdout, compared to 68 for KNN and 240 for Random Forest.

The model's tendency to produce more false positives (non-donors flagged as donors) is an acceptable cost given that the goal is to not miss likely donors rather than to achieve perfect precision.

---

Tech Stack

• Python 3.10
• pandas, numpy, matplotlib, seaborn
• scikit-learn — preprocessing, feature selection, modelling, evaluation
• scipy — statistical tests (Chi-Square, Spearman)
• Google Colab

---

Running the Notebook

1. Clone the repository
2. Ensure the data/ folder contains donors_train_target.csv, test.csv, and sample_submission.csv
3. Run all cells top to bottom — the final cell exports the submission CSV

Note: If using Google Colab, mount your Drive and set lab_root to point to the data/ folder before running.