A machine learning prediction is the output that a trained machine learning model produces when it is given new data. The model first learns patterns from historical examples during training; it is then used in a separate step, called inference, to score cases it has not seen. The output may be a class (fraud or not), a probability or a number such as a forecast.
Finance uses such predictions widely: real-time fraud scoring of card and crypto transactions, credit default estimates, churn models and trading signals. Their quality depends on the data and on whether the future resembles the past. A model that fits history too closely suffers from overfitting and performs badly on live data, which is why results are checked on out-of-sample data and through backtesting. Price forecasts are especially fragile, because markets adapt and patterns fade once traded on. Banks treat these models as a source of model risk that needs validation and monitoring.
Ref 1300000329