Santander: Predicting Customer Transactions

Source codeem-ech/ml-santander-transactions

Santander's dataset gives 200,000 customers, 200 unnamed numeric features, and one question: which customers will make a specific transaction. Only 10 percent do.

LightGBMOptunaNaive BayesKaggle

200,000

training customers

200

anonymous features

0.911

best out-of-fold AUC

100,000

synthetic test rows found

Problem

Understanding the challenge and user needs

Anonymous features remove every shortcut that domain knowledge provides. The model has to find structure in distributions alone, and the test set mixes real customers with synthetic rows that distort any statistic computed across it.

Grid of feature histograms split by target class — Each feature's distribution for customers who transacted and those who did not. The two overlap almost entirely, with small shifts the model has to add up.

Bar chart of the two target classes — Nine customers in ten made no transaction.

Approach

Strategy and methods used to address the problem

A Naive Bayes baseline set the reference, since the features behave almost independently. An analysis of the test set then separated 100,000 real rows from 100,000 synthetic ones, and LightGBM models moved from defaults to Optuna-tuned parameters to a three-seed blend.

Histograms from the synthetic row reconstruction — Tracing unique feature values back to real rows exposes which test rows are synthetic.

ROC curve for the Naive Bayes model — Out-of-fold ROC curve for the final Naive Bayes baseline.

Bar chart of hyperparameter importance from Optuna — Optuna's parameter importance: the fraction of features sampled per tree matters far more than any other setting.

Final Solution

The final design and outcome

Naive Bayes reached a five-fold AUC of 0.890. LightGBM raised that to 0.905 at baseline, 0.910 after tuning, and 0.911 as a three-seed blend.

ROC curve for the LightGBM model — Out-of-fold ROC curve for the LightGBM baseline.

Bar chart of the top twenty features by gain — The twenty most useful features by mean gain across folds. No single feature dominates.

Out-of-fold AUC by model

Model	AUC
Naive Bayes	0.890
LightGBM, default parameters	0.905
LightGBM, Optuna-tuned	0.910
LightGBM, three-seed blend	0.911

Impact & Validation

Results, learnings, and outcomes

The work shows a disciplined order of operations: a simple baseline first, an audit of the data second, and model complexity only after both.