Predictive analytics is the practice of using historical data, statistical models and machine learning to estimate what is likely to happen next. A predictive model takes known facts about a case, such as a borrower's payment history or a share's recent price behaviour, and produces a forecast or a probability: the chance of default, the expected demand, the likely price range.
It is one part of a wider field. Descriptive analytics summarises what has happened, predictive analytics estimates what will happen, and prescriptive analytics recommends an action. In finance the predictive kind is long established: credit scoring, insurance pricing and economic forecasting were statistical exercises decades before the term became common. What has changed is the volume of data and computing power available, which has widened both the methods and the range of uses.
Methods
Most predictive models fall into a few families.
- Regression models. Regression analysis relates an outcome to a set of explanatory variables. Logistic regression, which predicts the probability of a yes-or-no event, remains the standard tool for credit scorecards because its workings are easy to explain.
- Tree-based models. A decision tree splits cases into groups by successive rules. Ensembles of many trees, such as random forests and gradient boosting, are widely used on tabular data.
- Neural networks. A neural network can capture complex, non-linear patterns in large data sets, including text and images, at the cost of being harder to interpret.
- Time-series models. Time series forecasting projects a variable from its own past values, seasonality and trend.
Building a model follows a common sequence: defining the outcome, collecting and cleaning data, selecting variables, fitting the model on one portion of the data and testing it on another that it has not seen. Techniques such as cross-validation are used to check that the performance is not an accident of one sample.
Uses in finance
Lending is the oldest application. A credit scoring model estimates each applicant's probability of default, and banks use similar models to set limits, price loans and calculate loan-loss provisions.
Payment companies and card issuers apply predictive models to fraud detection, scoring each transaction in milliseconds for the likelihood that it is unauthorised. Compliance teams use them in transaction monitoring to rank alerts for possible money laundering.
In investment management, models forecast returns, volatility, earnings and trading volume. They feed algorithmic trading systems and risk reports, and they are tested on past data through backtesting before use. Insurers use them to price policies and estimate claims, and retail financial firms use them to predict which customers are likely to leave, respond to an offer or fall into arrears.
Model risk and common failures
A predictive model describes the past, and its forecasts hold only while the future resembles the data it learned from. Several failures recur.
Overfitting occurs when a model learns the noise in its training data and so performs well on history but poorly on new cases. It is a particular danger in financial markets, where the amount of genuinely independent data is small and many researchers test many ideas on the same history. Data leakage, in which information unavailable at the time of the decision slips into the training set, produces the same false confidence.
Relationships also shift. Customer behaviour, fraud methods and market conditions change, so a model's accuracy decays and it must be monitored and refitted. Events outside the range of the historical sample, such as a financial crisis or a pandemic, can make forecasts unreliable at the moment they matter most. A model may also find a correlation that has no causal basis and that disappears once it is acted upon.
Financial institutions manage these weaknesses through model risk management: independent validation, documentation, limits on use and regular performance reviews. United States banking supervisors set out their expectations in guidance issued in 2011, and supervisors elsewhere apply comparable standards.
Regulation and fairness
Where predictions affect individuals, legal limits apply. In the United States the Equal Credit Opportunity Act prohibits discrimination in lending and requires lenders to give specific reasons when credit is refused, which constrains the use of models that cannot be explained. A model that excludes protected characteristics can still discriminate if it relies on variables closely linked to them.
In the European Union the General Data Protection Regulation restricts decisions based solely on automated processing that have significant effects on a person, and the EU's Artificial Intelligence Act, adopted in 2024, treats systems used to assess the creditworthiness of individuals as high-risk, with duties on documentation, data quality and human oversight.
Securities regulators have also examined the use of predictive tools by brokers and advisers to steer customer behaviour, on the ground that a model tuned to increase trading or product sales may place the firm's interest ahead of the client's.