A classification threshold is the cut-off value used to turn the probability produced by a classifier into a decision. A model in machine learning may report a 0.63 probability that a payment is fraudulent or that a share will rise tomorrow. The threshold decides what happens next: at 0.5 the case is labeled positive, at 0.7 it is not. The default of 0.5 is only a convention.
Moving the threshold trades one kind of error for the other. A lower value catches more true cases but raises false alarms; a higher value does the reverse and increases the false negative rate. The right setting depends on what each error costs. A trading model may act only when the predicted probability exceeds 0.6, so that the expected gain covers the transaction cost, and accept fewer trades as a result. A threshold tuned on the same data used to judge the model invites overfitting.