Leakage is the general category; look-ahead-bias is its time-series form. Common leaks in machine-learning workflows: fitting a scaler on all data before splitting, imputing missing values using the full series mean, selecting features by their correlation with the target computed on everything, and overlapping label windows across folds.
The symptom is a validation score that is too good and a live score that is ordinary. If your classifier reports 0.72 accuracy on next-day direction, assume leakage before assuming genius; published realistic numbers sit near 0.52.
Prevention is structural. Split first, then fit every transformation inside the training fold only, and use purged-cross-validation with an embargo-period.
Related: look-ahead-bias, purged-cross-validation, embargo-period