Welcome to the user guide page
Before exploring our prediction models, we recommend watching our explanatory talks and read useful materials that are collected here. Our models have been developed for criminal justice professionals and healthcare staff. They are intended to compliment, not replace, professional judgement, and individual factors need to be considered. We encourage the use of our models for research purposes.
About the prediction models
The prediction models on this website are statistical tools that use information about a person or situation to estimate the likelihood of future outcomes.
For example, the OxRec tool predicts the risk of violent reoffending within one year. A predicted probability of 10% means that, among people with similar characteristics, about 1 in 10 are expected to reoffend within a year. This is similar to the chance of rain shown in a weather app: a 50% chance of rain means that, on days with similar conditions, it rains about half the time.
To assess the performance of the models it is important to understand two key metrics that you will come across in our publications: discrimination and calibration.
Discrimination (also called the c-statistic) refers to how well a model distinguishes between those who will experience the outcome and those who will not. A model with good discrimination assigns higher predicted risks to people who go on to experience the outcome than to those who do not.
Calibration refers to how close the predicted probabilities are to what actually happens in reality. A well-calibrated model gives risk estimates that match observed outcomes—for example, if 100 people are each given a 10% risk, about 10 of them should experience the outcome.
A model can have good discrimination but poor calibration. For example, suppose a model predicts a 10% chance of rain on Saturday and 70% on Sunday. If, over many similar Saturdays, it actually rains 50% of the time, and over similar Sundays it rains 90% of the time, then the model correctly ranks Sunday as riskier than Saturday (good discrimination), but its probability estimates are inaccurate (poor calibration).
In most of the prediction models here you will find Area Under the Curve (AUC) as discrimination metric and calibration plot or Brier Score as a calibration metric.
We have created an intuitive explanation about AUC
Clinical implications of the new risk assessment research
A new classification framework for risk models
Discussion paper (May 2026)
Is suicide prediction possible? Tackling myths
The content of the OxRisk website (including the risk calculators) is accurate according to the opinions of the authors of the calculators. The information and directions may or may not apply to specific individuals considered for assessment. The calculators provide a ‘best guess’ of likely outcomes based on current knowledge, but cannot provide an accurate prediction for each individual. Determination of appropriate application and scoring of the OxRisk calculators to specific cases and in specific jurisdictions requires the judgement of a clinician or criminal justice professional. OxRisk calculators are intended to be an adjunct to clinical decision-making.
Materials on this website are protected by copyright law. Access to the information on this website should be for the sole purpose of personal, educational and research use only. Where appropriate a single print out of a reasonable proportion of these materials may be made for personal education, research and private study. Materials should not be further copied, photocopied or reproduced, or distributed in electronic form. Any unauthorised use or distribution for commercial purposes is expressly forbidden.