The model is deliberately tuned for high recall — in fraud detection, a missed fraud (false negative) costs far more than a false alarm.