Data leakage occurs when a model learns from information unavailable during real predictions, creating a fraud that excels in training but fails in production. This happens when features contain target data, train/test sets contaminate, or future information leaks into the past. The result is a model that appears brilliant during validation but collapses once deployed.

Common leakage types include target leakage (like using post-diagnosis antibiotics to predict sepsis), train-test contamination (preprocessing on the full dataset), and temporal leakage (using future stock volumes to predict past prices). Each scenario gives the model unfair access to information it won't have at prediction time, creating a false sense of accuracy.

Prevention requires strict discipline: split data before any preprocessing, fit transformers on training data only, and use temporal splits for time-series. Always ask if each feature exists at prediction time. Tools like scikit-learn pipelines help enforce correct workflows. Real-world failures—like fraud detection models that learn from disputed transactions—demonstrate the catastrophic cost of undetected leaks.

The industry standard is to maintain a truly held-out test set collected after model development. Simulate production conditions by predicting on genuinely future data. If your model shows suspiciously high accuracy (e.g., >95%), investigate immediately. Proper validation using tools like TimeSeriesSplit and leak-free target encoding separates reliable models from fraudulent ones.