This research was made possible by the Dill Fund, established through the generosity of G.
Michael Dill, ’71, and by the guidance of Wabash Career Services throughout the
application.
This project traces its roots back to my last summer internship at the Alumni Engagement
office, where I sat inside the machinery of how a college actually stays connected to the
people who once called it home. The thing that stuck with me was how much data Wabash keeps on its alumni, and how little of it looked forward. Every report I saw described what had already happened: who gave, who showed up, who drifted. None of it asked what was coming next. That gap stayed with me long after the internship ended, and it eventually became the question my Dill Fund research set out to answer.
Over twelve weeks, I built and compared three supervised machine learning models, logistic regression, random forest, and XGBoost, to predict whether an alumnus is actively engaged with the College, using Wabash’s own six-tier engagement framework as the target. I pulled together giving history, event attendance, and volunteer records, and engineered time-based and frequency-based features, giving trends across the last four fiscal years, event attendance frequency, days since last gift, the kind of signal a static snapshot never captures.
Before any of that modeling could mean anything, though, I had to confront a quieter problem: a leakage policy. Several fields in the data, like Affinity Score and the Active Giving flag, turned out to be literal components of how “engaged” was already defined. Training on them wouldn’t have been a prediction, it would have been the model reading the answer key. So I excluded them before any model ever saw the data, and rebuilt the feature set entirely around variables that earn their signal rather than borrow it.
The best model, XGBoost, achieved a ROC-AUC of 0.994 on that leakage-controlled feature set, and rather than trust that number on faith, I re-ran the same pipeline, leakage exclusions and all, across ten independent random splits to confirm it wasn’t a lucky draw. It held. I also refused to let one model speak for the whole project. I pulled the same feature rankings out of the logistic regression coefficients and the random forest’s own importance scores and lined all three up side by side. Four variables, years of recent giving, donor loyalty, and giving consistency among them, showed up at the top no matter which model I asked, which told me something more convincing than any single accuracy number could: three structurally different algorithms had independently arrived at the same story about what engagement actually looks like.
I also didn’t take the standard 0.5 decision threshold for granted. I swept it from one end to the other and watched precision and recall trade places, then tested whether weighting the model for class imbalance was even doing anything at all. Turns out, honestly, not much, the imbalance in this data was mild enough that the model didn’t need the extra help, and I’d rather report that plainly than pretend a technique mattered when the numbers said otherwise.
But the result I was looking for didn’t come from the accuracy score at all, it came from interpretability. Using SHAP, both across the whole model and for individual alumni, recent giving history, event attendance, and donor loyalty consistently rose to the top as the strongest predictors, quietly confirming what Wabash’s own Advancement office had long suspected but never formally tested. The real twist showed up at the individual level, though. A person’s specific major or vocation barely moved the needle in the aggregate view, buried under thousands of other alumni, but for one particular graduate it explained almost the entire prediction. Global and local interpretability, I learned, aren’t two ways of saying the same thing, they’re two different questions, and a model that can only answer one of them is only half honest.
As a Computer Science major with a Math minor, this project gave me the clearest picture I’ve had yet of what a career in applied data science really looks like: not chasing the flashiest algorithm in the room, but asking patient, careful questions of real data, and being honest about what it can and cannot tell you. I’m leaving this summer with a project I’m genuinely proud of, and a much sharper sense of the work I want to spend my career doing
