The challenge
With a positive class of just 0.1%, this is a rare-event problem, similar in shape to fraud detection. A model that predicts "no" for everyone is 99.9% accurate and completely useless. The team needed a way to rank constituents so that gift officers could spend limited time where it would count.
Messy data
The data arrived as multiple flat files: giving history, event attendance and volunteer engagement. I merged and standardized them into constituent-level summary features.
Age, normally a strong signal for planned giving, was missing for 53.9% of records. Rather than impute it, which would have introduced bias, I engineered behavioral proxies that capture donor maturity: giving recency, cumulative giving and engagement scores.
Modeling
- Compared three models: logistic regression (for interpretability), XGBoost and LightGBM, each with a feature set suited to it.
- Built interaction terms that combine capacity, engagement and relationship signals, and quadratic terms for non-linear wealth effects.
- Evaluated with precision-recall AUC and top-K capture rates instead of accuracy, because of the extreme class imbalance.
Choosing the model
LightGBM had the better overall PR-AUC (0.327 vs 0.051). But XGBoost captured 37.7% of likely prospects in the top 1%, against 14.8% for LightGBM, and the top 1% is where Advancement invests most heavily. We chose XGBoost.
The lesson: the best model on paper isn't always the best model for the business problem. Evaluate where the decision actually gets made.
From scores to tiers
Scores only matter if someone can act on them, so we turned them into a three-tier framework with a cultivation playbook for each tier.
| Tier | Constituents | Prospects captured | Suggested approach |
|---|---|---|---|
| Top 1% | 3,182 | 37.7% | High-touch: one-to-one estate planning meetings |
| Top 5% | 15,910 | 67% | Semi-personalized programming |
| Top 10% | 31,820 | 75% | Scalable digital campaigns |
Focusing on the top 10% cut outreach volume by 90% while still reaching three-quarters of likely prospects. We also recommended restructuring gift officer portfolios around these tiers, and refreshing scores quarterly with alerts for "risers" moving between tiers.
What drove propensity
- Spouse-alumni households were the strongest predictor: dual-household ties to the university sharply increased planned-giving propensity.
- Consistent giving beat isolated large gifts. Steady patterns were far more predictive than one big donation, which challenges a common fundraising assumption.
- Geography amplifies engagement. Highly engaged U.S. residents showed elevated likelihood, which makes face-to-face cultivation practical.
- Breadth matters. Donors who supported several designations were significantly more likely prospects than those who only joined campaigns.