Skip to main content

HMPPS: Risk of Reoffending Tools

Risk of reoffending assessment tools for individuals under the supervision of HMPPS.

1. Summary

1 - Name

HMPPS Risk of Reoffending Tools

2 - Description

The HMPPS Risk of Reoffending Tools are made up of five algorithmic sub-tools. These tools are based on machine learning. Each sub-tool has a distinct predictive focus but all of them provide users with a risk score estimating the likelihood of proven reoffending within 2 years from the point of assessment / calculation (typically the start of a community order, suspended sentence, release on licence, or a review point thereafter) for specific offence categories.

The five sub-tools are:

  • All Reoffending Predictor (ARP) Tool predicting any recordable criminal offence
  • Violent Reoffending Predictor (VRP) Tool predicting violent offences
  • Serious Violent Reoffending Predictor (SVRP) Tool predicting serious violent offences
  • Combined Serious Reoffending Predictor (CSRP) Tool predicting serious offences (covering both serious non-sexual violent offences and sexual offences)
  • Sexual Reoffending Predictor (SRP) Tool predicting specified sexual offences in two categories: Direct-Contact sexual offences (DC-SRP) and Images and Indirect Contact sexual offences against children (IIC-SRP)

The tools are used during sentencing in court, and risk assessment while in prison or on probation, to provide practitioners with an evidence-based risk of reoffending.

The outputs enable evidence based reoffending risk assessments to be utilised to inform risk assessment for sentencing proposals, proposals in parole reports, eligibility for interventions, and setting probation supervision plans for offenders under management in the community.

3 - Website URL

N/A

4 - Contact email

RADSAIPMO@justice.gov.uk or datascience@justice.gov.uk

Tier 2 - Owner and Responsibility

1.1 - Organisation or department

Ministry of Justice / His Majesty’s Prison and Probation Service (HMPPS)

1.2 - Team

Public Protection Group

1.3 - Senior responsible owner

Public Protection Group Director

1.4 - Third party involvement

No

1.4.1 - Third party

N/A

1.4.2 - Companies House Number

N/A

1.4.3 - Third party role

N/A

1.4.4 - Procurement procedure type

N/A

1.4.5 - Third party data access terms

N/A

Tier 2 - Description and Rationale

2.1 - Detailed description

Purpose The HMPPS Risk of Reoffending Tools provide an evidence-based assessment of an individual’s likelihood of proven reoffending within the next two years from the point of assessment / calculation (typically the start of a community order, suspended sentence, release on licence, or a review point thereafter) for five types of offence categories. A distinct sub-tool is used per category:

  • All Reoffending Predictor (ARP) Tool: Predicting any recordable criminal offences
  • Violent Reoffending Predictor (VRP) Tool: Predicting violent offences
  • Serious Violent Reoffending Predictor (SVRP) Tool: Predicting serious violent offences
  • Combined Serious Reoffending Predictor (CSRP) Tool: Predicting serious offences (combining both serious violent offences and specified sexual offences)
  • Sexual Reoffending Predictor (SRP) Tool: Predicting specified sexual offences in two main categories: Direct-Contact sexual offences (DC-SRP) and Images and Indirect Contact sexual offences against children (IIC-SRP)

The tools’ risk scores inform practitioner assessment of risk, supporting decision making across courts, prison and probation services.

Intended Users The primary intended users sit within the two following organisations:

  • His Majesty’s Prison and Probation Service (HMPPS), where risk scores form an important input for 1) pre-sentence reports considered at sentencing in court, 2) Risk of Serious Harm assessment, 3) determining supervision levels for individuals managed in the community and 4) allocating individuals to intervention programmes as part of their sentence.
  • The Police who use the SRP sub-tool as part of their management of individuals with notification requirements in relation to sexual or sexually motivated offences.

Key Aspects and Functions The risk assessment typically takes place when an individual under HMPPS supervision is risk assessed by a probation practitioner as part of their licence or supervision arrangements, by a prison practitioner during their custodial sentence, or when a pre-sentence report is being written about a newly convicted individual.

For each of the above offence categories, the tools produce a numerical probability score and assign it to a band (typically one of low, medium, high, or very high) indicating their estimated risk of committing a new offence of this type within the next two years from the point of assessment / calculation which is eventually brought to justice (through a conviction, formal caution etc.). This outcome is typically referred to as ‘proven reoffending’; for brevity across this document, any reference to ‘reoffending’ does mean proven reoffending and excludes new offences that are not brought to justice.

The tools’ risk scores are derived from different underlying predictive models. For all (ARP), violent (VRP), and serious violent (SVRP) risk of reoffending scores, one of two possible model outputs is used depending on the input data available. For serious risk of reoffending, risk estimates for three different types of serious offences are both presented separately and aggregated into a single combined serious (CSRP) risk score: these include serious violent offences (SVRP), direct contact sexual offences (DC-SRP), and a group comprising images and indirect contact sexual offences with children (IIC-SRP). As mentioned above, the serious violent score uses one of two possible model outputs, whereas the two sexual scores are based on a single model each.

All underlying models were trained on historic offender data, using information about offenders and whether they reoffended within a two-year period from the point of assessment / calculation. We use information relating to proven offending, rather than police reports or other unproven offending, to reduce the potential for bias associated with using allegations or other offending information that has not resulted in a proven outcome. From this training data, the models learned which characteristics were associated with reoffending. Operationally, the tool applies these learned patterns to the offender’s characteristics and offending history at the time of assessment to produce their risk scores.

The tools underwent ethical and equality assessment to identify and assess potential fairness and equality risks and appropriate mitigations (see section 2.5.1).

Scope and Limitations The tools apply only to individuals under HMPPS supervision as well as post-supervision to individuals with notification requirements. Their scope only covers the prediction of proven reoffending for the respective offence categories within a two-year time frame from the point of calculation. Each risk score within the tools does not predict other offence types or longer-term risks. The two-year prediction period was chosen to balance competing considerations. A longer timeframe would capture more potential reoffending but would reduce the practical ability to train and validate the model; a shorter timeframe would improve technical feasibility but underrepresent relevant risk. Finally, the tools only provide predictions of proven reoffending, that is offences proven in court through conviction or proven due to a police caution. The respective sub-tools are not designed to predict unconvicted reoffending. This includes offending that goes undetected by the police, and offences where prosecution is not possible or is discontinued. Additionally, to base the tools on timely evidence, the two-year prediction period is relatively recent – while a minimum of one further year’s “waiting time” is allowed for offences to be brought to justice, some reoffences occurring beyond the waiting period (i.e. more than three years after the start of the prediction period) will not be counted as “proven”.

The risk of sexual reoffending is only calculated for male individuals with a current sexual offence and/or previous history of sexual offending including sexually motivated offences that are not sexual in statute. Females neither receive a DC- nor IIC-SRP risk score or band. However, in the context of CSRP, for females with current and/or previous sexual offending, a base risk rate is added to their overall CSRP risk score because the number of women with proven sexual reoffending outcomes in the development data was insufficient to develop and validate a reliable female-specific model.

The tools’ risk scores reflect past patterns of offending alongside an individual’s current personal circumstances where available. Therefore, risk scores’ accuracy is dependent on timely updates should information change. Furthermore, the tools may not capture all aspects of a person’s current circumstances or behaviours relevant for prediction. For this reason, professional judgement and contextual information remain essential parts of any decision informed by the tools.

Further Information HMPPS has used reoffending risk predictors for a number of years, and their use, evaluation and development have been the subject of published Ministry of Justice research, including Revalidation: Risk of recidivism tools (2024) and Risk of Serious Violence of those already known to the Probation Service (2025).

2.2 - Benefits

The tools provide evidence-led and standardised estimates of reoffending risk to inform sentencing proposals, allocation of rehabilitative interventions, and pre- and post-release management. Combined risk scores relating to serious reoffending (serious violent and sexual reoffending) provide a standardised starting point for the practitioner-led Risk of Serious Harm assessment. Risk scores and associated risk bands help identify individuals with different levels of reoffending risk, supporting HMPPS to target supervision, interventions and resources according to assessed risk and need.

2.3 - Previous process

Previous generations of risk of reoffending tools were already in operational use within HMPPS before the current tools were deployed. The current tools replace and update those existing tools rather than introducing algorithmic risk assessment as a new process. Details of the current versions of each model are provided in section 2.4.2.

2.4 - Alternatives considered

For this update of the tools, several algorithmic and non‑algorithmic alternatives were assessed. Logistic regression was selected because it provides a high level of explainability while maintaining acceptable predictive performance. This choice also provides conceptual continuity with previous generations of the reoffending tools, enabling meaningful comparison over time and straightforward integration into existing operational systems.

A tree-based model—specifically Explainable Boosting Machines (EBMs)—was considered as an alternative algorithmic approach. EBMs offered comparable predictive performance, but logistic regression was preferred due to its conceptual continuity with earlier models which allowed straightforward operational deployment within HMPPS’s digital platforms.

A non-algorithmic alternative—relying on practitioner judgement based on business rules—was also considered. However, practitioner assessments can vary across regions and individual practitioners. Algorithmic tools provide a standardised, evidence-based risk estimate, helping to reduce this variability. Machine learning based tools help to address this variability by providing a consistent, evidence-based risk estimate. Nonetheless, when considering Risk of Serious Harm, using an algorithm approach as a standardised starting point supports rather than replaces professional judgement. It allows practitioners to balance the algorithmic assessment with contextual factors not captured by the model.

Tier 2 - Deployment Context

3.1 - Integration into broader operational process

Risk assessment is central to probation and prison practice, and algorithmic tools play an important role in the overall structured professional judgement that risk assessments are based on.

The HMPPS Risk of Reoffending Tools are embedded into the HMPPS digital risk assessment tool and are completed as a standard part of all risk assessments for those subject to a pre-sentence report following conviction, at the start of a community or custodial sentence, and at every assessment review point thereafter. Each sub-tool of the HMPPS Risk of Reoffending Tools provides an estimate of the likelihood of the individual committing any (ARP), a violent (VRP), a serious violent (SVRP), specific sexual (DC-SRP and IIC-SRP) and a serious violent and/or sexual (CSRP) offence respectively within the next two years and subsequently being convicted for that offence.

The risk scores support, rather than automate, practitioners’ individual-level risk judgements.

HMPPS therefore use the ARP, VRP, CSRP and SRP scores primarily, alongside other criteria, for decisions and processes such as allocation of a new case to the right grade of practitioner, targeting and eligibility for specific interventions to address reoffending as well as other rehabilitative interventions such as alcohol services.

CSRP and SRP scores are used to identify individuals on the HMPPS caseload assessed as having higher likelihoods of committing a further serious violent and/or sexual offence.

The CSRP score is therefore calculated at court and provided to an assessor as the first part of the four-step process of risk assessment. This ensures that the practitioner is aware of the percentage likelihood for serious reoffending as well as the factors driving the score, such as alcohol use and temper control for violence. The assessor will then take these factors into account when setting an overall level of risk of serious harm that the individual is assessed as posing as well as what action would be needed to manage that risk.

Assessors may also consider the three distinct components of the CSRP score, and the VRP score – especially in cases (e.g. domestic abuse) where persistent violence could inflict serious harm through cumulative and/or psychological impact rather than a single act of grave physical injury.

The ARP, VRP, SRP, SVRP and/ or CSRP score or risk band can be presented in some formal written reports such as pre-sentence reports, parole reports and recall decisions.

3.2 - Human review

An assessor answers several questions in the HMPPS digital risk assessment tool which provides the calculation of the suite of risk scores (ARP, VRP, SVRP, SRP, CSRP) within the Risk of Reoffending Tools. Guidance is available to support assessors in completing the inputs accurately. The scores are calculated automatically from these inputs, but are used within a practitioner-led assessment process; practitioners remain responsible for the wider assessment and for considering relevant information not captured by the models.

When all questions have been answered, all risk scores can be viewed within the HMPPS digital risk assessment tool interface. Assessors update the information used to calculate the scores when they become aware of relevant changes to an individual’s criminal history or dynamic risk factors. Where applicable, the scores can be recalculated using the updated information. For models that include offence-free time, scores may also change over time to reflect this.

Case management may be subject to internal or external audit or inspection.

3.3 - Frequency and scale of usage

The risk scores within the risk assessment system are calculated at certain points within an individual’s journey typically the start of a community order, suspended sentence, release on licence, or a review point thereafter. This means thousands of prison and probation staff are engaging and routinely calculating these risk scores using the Risk of Reoffending Tools daily. The risk scores are available to inform the overall risk assessment of the Risk of Serious Harm, though the CSRP and SRP risk scores are of particular importance as a baseline for this. Depending on what point it is in the individual’s sentence (or if it is pre-sentence), the risk scores may also be used for other purposes (see 2.3.1).

3.4 - Required training

The Risk of Reoffending Tools are introduced as part of practitioners’ mandatory training on offender risk, which includes an opportunity for skills practice. Score calculations occur as part of the wider assessment of offender needs and Risk of Serious Harm within the HMPPS digital risk assessment tool. Guidance is available to support practitioners in completing the relevant inputs accurately. Additional learning and support are available, including bespoke learning for court staff, briefing materials and short videos.

3.5 - Appeals and review

With respect to sentencing and risk management decisions, the tools provide information, but the decision ultimately sits with the judicial office holder / probation officer respectively. Normal appeals and reviews processes will apply.
Further information on criminal court appeals: https://www.gov.uk/appeal-against-crown-court-verdict https://www.gov.uk/appeal-magistrates-court-decision

Tier 2 - Tool Specification

4.1.1 - System architecture

The Risk of Reoffending Tools are deployed as part of the HMPPS digital risk assessment tool, an application used by HMPPS practitioners to assess an individual’s risks and needs. Assessment data is sourced through integrations with other HMPPS systems and through information entered by users. The application and its data are hosted within a secure MoJ cloud environment with controlled access. Results from the tools are presented to users within the application and shared with authorised HMPPS and MoJ systems through system integrations.

4.1.2 - System-level input

Some offender details are pulled from other MoJ/HMPPS applications (Probation Delius and Prison NOMIS) into the HMPPS digital risk assessment tool:

  • Date of birth,
  • Gender
  • Details of the most recent offence

Other inputs are entered manually by the user:

  • Offending history
  • Criminogenic need and risk factors such as accommodation, employment and drug and alcohol use.

4.1.3 - System-level output

All, violent, serious violent, combined serious and sexual reoffending risk scores (as percentages) and bands (typically Low, Medium, High and, for some risk scores, Very High) are presented to users on-screen. The outputs may also be shared with authorised HMPPS and MoJ systems and used for reporting.

4.1.4 - Maintenance

Revalidation studies form part of MoJ’s ongoing research and monitoring of model performance. The models do not have a fixed retraining schedule. Retraining is considered where monitoring or revalidation indicates a need, or where there is evidence that an updated model would improve on the operational model.

4.1.5 - Models

All Reoffending Predictor (ARP) Tool: All reoffending risk score predictor models (static/dynamic versions): logistic regression Violent Reoffending Predictor (VRP) Tool: Violent reoffending risk score predictor models (static/dynamic versions): logistic regression Serious Violent Reoffending Predictor (SVRP) Tool: Serious violent reoffending risk score predictor models (static/dynamic versions): logistic regression Combined Serious Reoffending Predictor (CSRP) Tool:

  • Serious violent risk score predictor models (SVRP) (static/dynamic versions): logistic regression
  • Direct contact (DC) - Sexual Reoffending Predictor (SRP): rule-based plus mathematical combination to align to logistic regression model
  • Images and Indirect Contact sexual offences against children (IIC) - Sexual Reoffending Predictor (SRP): baseline-risk model Sexual Reoffending Predictor (SRP) Tool:
  • Direct contact (DC) - Sexual Reoffending Predictor (SRP): rule-based plus mathematical combination to align to logistic regression model
  • Images and Indirect Contact sexual offences against children (IIC) - Sexual Reoffending Predictor (SRP): baseline-risk model

Tier 2 - Model Specification: All Reoffending Predictor (ARP) model - static version (1)

4.2.1. - Model name

All Reoffending Predictor (ARP) model - static version

4.2.2 - Model version

February 2026 release The previously deployed model version was released in 2009 and known as Offender Group Reconviction Scale 3 (OGRS3).

4.2.3 - Model task

This model is designed to predict the two-year risk of proven reoffending for any recordable criminal offences.

4.2.4 - Model input

The model uses static risk factors. The wording ‘static risk factors’ reflects that someone’s characteristics were fixed/ unchangeable at the time of a specific sentence.

The factors can be grouped into the following:

  • Gender
  • Age at effective assessment date by gender
  • Category of index offence
  • Whether current index offence is the individual’s first recorded sanction
  • Whether current index offence is the individual’s second recorded sanction
  • Total count of sanctions
  • Where the current index offence is an individual’s second recorded sanction, the number of years between first and second sanction is given (by gender)
  • Offence free time in months since discharge / non-custodial sentence (zero when in custody)
  • For those with three or more sanctions, the COPAS rate for all offences (by gender)

For more information on the COPAS rate see Copas & Marshall (1998, https://www.jstor.org/stable/2986060) as well as Howard (2015, Chapter 8, https://assets.publishing.service.gov.uk/media/5a7f676fed915d74e33f6380/researchanalysis-offender-assessment-system.pdf)

4.2.5 - Model output

A risk score between 0 and 1, the probability of proven reoffending with any recordable criminal offence within the next two years from a given date.

4.2.6 - Model architecture

Type of model: Logistic regression implemented using the scikit-learn Python library (see https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html).

Default hyperparameters were used except for the ‘C’ parameter (inverse of regularisation strength) which was tuned using stratified K-fold cross-validation with a randomised grid search (https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RandomizedSearchCV.html).

The default regularisation (‘ridge’ or ‘L2’ regularisation) was used.

4.2.7 - Model performance

Model validation was done using 10-fold cross-validation during model training as well as testing on an unseen dataset. Trained models were evaluated using two key metrics:

  1. c-index, a measure of concordance, equivalent here to the area under the receiver operating characteristic curve. It is a measure of how well a model predicts the order of outcomes, in this case reoffending. Imagine two individuals A and B: A reoffends and B doesn’t within two years of starting their community sentence. If the model gives a higher risk score to A compared to B, then A and B are a concordant pair. The c-index is the proportion of all comparable pairs where the model got the order right. 100% (c-index = 1) means perfect ordering. 50% (c-index = 0.5) means random guessing.

  2. Mean miscalibration, which is the accuracy of the estimate of prevalence of reoffending at the population level. The closer to 0 the better calibrated is the model.

Please see Table 1 for details.

4.2.8 - Datasets and their purposes

Each dataset used for model development contained information from criminal justice system administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. Importantly, there are two types of risk assessments that differ in the amount of information they assess. In short, there is an abbreviated risk assessment called Layer 1, and full risk assessment, called Layer 3. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train the model and evaluate it using cross-validation. This dataset included anyone in the caseload regardless of risk assessment layer type.

A flow dataset containing any offender who started the community part of their sentence between 01 July to 31 December 2018 (with information from administrative systems from probation) was used to test the model. Two versions of this dataset were used: the ‘full evaluation dataset’ including anyone in the dataset regardless of risk assessment layer type. The ‘Layer 3 evaluation dataset’ containing only individuals with a Layer 3 risk assessment.

Tier 2 - Model Specification: All Reoffending Predictor (ARP) model - dynamic version (2)

4.2.1. - Model name

All Reoffending Predictor (ARP) model - dynamic version

4.2.2 - Model version

February 2026 release This model is deployed operationally for the first time.

4.2.3 - Model task

This model is designed to predict the two-year risk of proven reoffending for any recordable criminal offences.

4.2.4 - Model input

The model uses relevant static risk factors from the ARP static model. In addition, it uses dynamic risk factors/criminogenic needs, which reflect circumstances that may change over the course of an individual’s sentence.

For risk factors from the static model version please refer to its section above.

The additional dynamic risk factors can be grouped into the following:

  • Accommodation needs
  • Employment needs
  • Relationship needs
  • Drug misuse needs
  • Alcohol misuse needs
  • Behavioural and attitude needs
  • Lifestyle needs
  • Specific drug usage

4.2.5 - Model output

A risk score between 0 and 1, representing the probability of proven reoffending with any type of offence (except breach) within the next 2-years from a given date.

4.2.6 - Model architecture

Type of model: Logistic regression implemented using the scikit-learn Python library (see https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html).

Default hyperparameters were used except for the ‘C’ parameter (inverse of regularisation strength) which was tuned using stratified K-fold cross-validation with a randomised grid search (https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RandomizedSearchCV.html).

The default regularisation (‘ridge’ or ‘L2’ regularisation) was used.

4.2.7 - Model performance

Model validation was done using 10-fold cross-validation during model training as well as testing on an unseen dataset. Trained models were evaluated using two key metrics:

  1. c-index, a measure of concordance, equivalent here to the area under the receiver operating characteristic curve. It is a measure of how well a model predicts the order of outcomes, in this case reoffending. Imagine two individuals A and B: A reoffends and B doesn’t within two years of starting their community sentence. If the model gives a higher risk score to A compared to B, then A and B are a concordant pair. The c-index is the proportion of all comparable pairs where the model got the order right. 100% (c-index = 1) means perfect ordering. 50% (c-index = 0.5) means random guessing.

  2. Mean miscalibration, which is the accuracy of the estimate of prevalence of reoffending at the population level. The closer to 0 the better calibrated is the model.

Please see Table 1 for details.

4.2.8 - Datasets and their purposes

Each dataset used for model development contained information from criminal justice system administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. Importantly, there are two types of risk assessments that differ in the amount of information they assess. In short, there is an abbreviated risk assessment called Layer 1, and full risk assessment, called Layer 3. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train the model and evaluate it using cross-validation. This dataset included only individuals in the caseload with complete Layer 3 risk assessments, as only this dataset contained all necessary feature information required to train the model.

A flow dataset containing any offender who started the community part of their sentence between 01 July to 31 December 2018 (with information from administrative systems from probation) was used to test the model. The evaluation dataset only included individuals with a Layer 3 risk assessment (‘Layer 3 evaluation dataset’).

Tier 2 - Model Specification: Violent Reoffending Predictor (VRP) model- static version (3)

4.2.1. - Model name

Violent Reoffending Predictor (VRP) model- static version

4.2.2 - Model version

February 2026 release This model is deployed operationally for the first time.

4.2.3 - Model task

This model is designed to predict the two-year risk of proven reoffending for nonsexual violent offences.

4.2.4 - Model input

The model uses static risk factors. The wording ‘static risk factors’ reflects that someone’s characteristics were fixed/ unchangeable at the time of a specific sentence.

These factors can be grouped into the following:

  • Gender
  • Age at effective assessment date by gender
  • Category of index offence
  • Whether the current index offence is the individual’s first recorded sanction
  • Whether the current index offence is the individual’s second recorded sanction
  • Total count of sanctions
  • For those for whom the current index offence is someone’s second recorded sanction, the number of years between first and second sanction by gender
  • Offence free time in months since discharge/ noncustodial sentence (zero when in custody)
  • For those with three or more sanctions, the COPAS rate* for violent offences split by gender
  • Whether an individual has been sanctioned once for violent offences?
  • Whether an individual has been sanctioned once for violent offences?
  • Total count of sanctions that included violent offences
  • COPAS rate for violent offences

For more information on the COPAS rate see Copas & Marshall (1998, https://www.jstor.org/stable/2986060), as well as Howard (2015, Chapter 8, https://assets.publishing.service.gov.uk/media/5a7f676fed915d74e33f6380/researchanalysis-offender-assessment-system.pdf)

4.2.5 - Model output

A risk score between 0 and 1, representing the probability of proven reoffending with a non-sexual violent offence within the next 2-years from a given date.

4.2.6 - Model architecture

Type of model: Logistic regression implemented using the scikit-learn Python library (see https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html).

Default hyperparameters were used except for the ‘C’ parameter (inverse of regularisation strength) which was tuned using stratified K-fold cross-validation with a randomised grid search (https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RandomizedSearchCV.html).

The default regularisation (‘ridge’ or ‘L2’ regularisation) was used.

4.2.7 - Model performance

Model validation was done using 10-fold cross-validation during model training as well as testing on an unseen dataset. Trained models were evaluated using two key metrics:

  1. c-index, a measure of concordance, equivalent here to the area under the receiver operating characteristic curve. It is a measure of how well a model predicts the order of outcomes, in this case reoffending. Imagine two individuals A and B: A reoffends and B doesn’t within two years of starting their community sentence. If the model gives a higher risk score to A compared to B, then A and B are a concordant pair. The c-index is the proportion of all comparable pairs where the model got the order right. 100% (c-index = 1) means perfect ordering. 50% (c-index = 0.5) means random guessing.

  2. Mean miscalibration, which is the accuracy of the estimate of prevalence of reoffending at the population level. The closer to 0 the better calibrated is the model.

Please see Table 1 for details.

4.2.8 - Datasets and their purposes

Each dataset used for model development contained information from criminal justice system administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. Importantly, there are two types of risk assessments that differ in the amount of information they assess. In short, there is an abbreviated risk assessment called Layer 1, and full risk assessment, called Layer 3. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train the model and evaluate it using cross-validation. This dataset included anyone in the caseload regardless of risk assessment layer type.

A flow dataset containing any offender who started the community part of their sentence between 01 July to 31 December 2018 (with information from administrative systems from probation) was used to test the model. Two versions of this dataset were used: the ‘full evaluation dataset’ including anyone in the dataset regardless of risk assessment layer type. The ‘Layer 3 evaluation dataset’ containing only individuals with a Layer 3 risk assessment.

Tier 2 - Model Specification: Violent Reoffending Predictor (VRP) model-dynamic version (4)

4.2.1. - Model name

Violent Reoffending Predictor (VRP) model-dynamic version

4.2.2 - Model version

February 2026 release The previously deployed model version was released and 2009 and known as Offender Assessment System (OASys) Violence Predictor (OVP).

4.2.3 - Model task

This model is designed to predict the two-year risk of proven reoffending for nonsexual violent offences.

4.2.4 - Model input

The model uses relevant static (those that cannot be changed over the course of someone’s sentence) risk factors from the VRP static model. In addition, it also uses dynamic risk factors/ criminogenic needs. The wording dynamic risk factors reflects that these are factors which change as the individual passes through their rehabilitative journey of the criminal justice system.

For risk factors in the static model version please refer to its relevant section above.

The additional dynamic risk factors can be grouped into the following:

  • Accommodation needs
  • Employment needs
  • Relationship needs
  • Drug misuse needs
  • Alcohol misuse needs
  • Behavioural needs
  • Lifestyle needs
  • Specific drug usage

4.2.5 - Model output

A risk score between 0 and 1, representing the probability of proven reoffending with a serious non-sexual-violent offence within the next 2-years from a given date.

4.2.6 - Model architecture

Type of model: Logistic regression implemented using the scikit-learn Python library (see https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html).

Default hyperparameters were used except for the ‘C’ parameter (inverse of regularisation strength) which was tuned using stratified K-fold cross-validation with a randomised grid search (https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RandomizedSearchCV.html).

The default regularisation (‘ridge’ or ‘L2’ regularisation) was used.

4.2.7 - Model performance

Model validation was done using 10-fold cross-validation during model training as well as testing on an unseen dataset. Trained models were evaluated using two key metrics:

  1. c-index, a measure of concordance, equivalent here to the area under the receiver operating characteristic curve. It is a measure of how well a model predicts the order of outcomes, in this case reoffending. Imagine two individuals A and B: A reoffends and B doesn’t within two years of starting their community sentence. If the model gives a higher risk score to A compared to B, then A and B are a concordant pair. The c-index is the proportion of all comparable pairs where the model got the order right. 100% (c-index = 1) means perfect ordering. 50% (c-index = 0.5) means random guessing.

  2. Mean miscalibration, which is the accuracy of the estimate of prevalence of reoffending at the population level. The closer to 0 the better calibrated is the model.

Please see Table 1 for details.

4.2.8 - Datasets and their purposes

Each dataset used for model development contained information from criminal justice system administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. Importantly, there are two types of risk assessments that differ in the amount of information they assess. In short, there is an abbreviated risk assessment called Layer 1, and full risk assessment, called Layer 3. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train the model and evaluate it using cross-validation. This dataset included only individuals in the caseload with complete Layer 3 risk assessments, as only this dataset contained all necessary feature information required to train the model.

A flow dataset containing any offender who started the community part of their sentence between 01 July to 31 December 2018 (with information from administrative systems from probation) was used to test the model. The evaluation dataset only included individuals with a Layer 3 risk assessment (‘Layer 3 evaluation dataset’).

Tier 2 - Model Specification: Serious Violent Reoffending Predictor (SVRP) model - static version (5)

4.2.1. - Model name

Serious Violent Reoffending Predictor (SVRP) model - static version

4.2.2 - Model version

February 2026 release The previously deployed model version was released in 2014 and known as serious-non-sexual violence (SNSV) model – static version.

4.2.3 - Model task

This model is designed to predict the two-year risk of proven reoffending for serious nonsexual violent offences.

4.2.4 - Model input

The model uses static risk factors. The wording static risk factors reflects that someone’s characteristics were fixed/ unchangeable at the time of a specific sentence.

The factors can be grouped into the following:

  • Gender
  • Age at effective assessment date by gender
  • Category of index offence
  • Whether the current index offence is the individual’s first recorded sanction
  • Whether the current index offence is the individual’s second recorded sanction
  • Total count of sanctions
  • For those for whom the current index offence is someone’s second recorded sanction, the number of years between first and second sanction by gender
  • Offence free time in months since discharge/ noncustodial sentence (zero when in custody)
  • For those with three or more sanctions, the COPAS rate* for violence offences by gender
  • Whether an individual has never been sanctioned for violent offences by gender
  • Whether an individual has been sanctioned once for violent offences?
  • Total count of sanctions that included violent offences
  • COPAS rate* for violent offences

For more information on the COPAS rate see Copas & Marshall (1998, https://www.jstor.org/stable/2986060), as well as Howard (2015, Chapter 8, https://assets.publishing.service.gov.uk/media/5a7f676fed915d74e33f6380/researchanalysis-offender-assessment-system.pdf)

4.2.5 - Model output

A risk score between 0 and 1, representing the probability of proven reoffending with a serious non-sexual-violent offence within the next 2-years from a given date.

4.2.6 - Model architecture

Type of model: Logistic regression implemented using the scikit-learn Python library (see https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html).

Default hyperparameters were used except for the ‘C’ parameter (inverse of regularisation strength) which was tuned using stratified K-fold cross-validation with a randomised grid search (https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RandomizedSearchCV.html).

The default regularisation (‘ridge’ or ‘L2’ regularisation) was used.

4.2.7 - Model performance

Model validation was done using 10-fold cross-validation during model training as well as testing on an unseen dataset. Trained models were evaluated using two key metrics:

  1. c-index, a measure of concordance, equivalent here to the area under the receiver operating characteristic curve. It is a measure of how well a model predicts the order of outcomes, in this case reoffending. Imagine two individuals A and B: A reoffends and B doesn’t within two years of starting their community sentence. If the model gives a higher risk score to A compared to B, then A and B are a concordant pair. The c-index is the proportion of all comparable pairs where the model got the order right. 100% (c-index = 1) means perfect ordering. 50% (c-index = 0.5) means random guessing.

  2. Mean miscalibration, which is the accuracy of the estimate of prevalence of reoffending at the population level. The closer to 0 the better calibrated is the model.

Please see Table 1 for details.

4.2.8 - Datasets and their purposes

Each dataset used for model development contained information from criminal justice system administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. Importantly, there are two types of risk assessments that differ in the amount of information they assess. In short, there is an abbreviated risk assessment called Layer 1, and full risk assessment, called Layer 3. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train the model and evaluate it using cross-validation. This dataset included anyone in the caseload regardless of risk assessment layer type.

A flow dataset containing any offender who started the community part of their sentence between 01 July to 31 December 2018 (with information from administrative systems from probation) was used to test the model. Two versions of this dataset were used: the ‘full evaluation dataset’ including anyone in the dataset regardless of risk assessment layer type. The ‘Layer 3 evaluation dataset’ containing only individuals with a Layer 3 risk assessment.

Tier 2 - Model Specification: Serious Violent Reoffending Predictor (SVRP) model - dynamic version (6)

4.2.1. - Model name

Serious Violent Reoffending Predictor (SVRP) model - dynamic version

4.2.2 - Model version

February 2026 release The previously deployed model version was deployed in 2014 and known as serious-non-sexual violence (SNSV) model – dynamic version.

4.2.3 - Model task

This model is designed to predict the two-year risk of proven reoffending for serious nonsexual violent offences.

4.2.4 - Model input

The model uses relevant static (those that cannot be changed over the course of someone’s sentence) risk factors as the SVRP static model. In addition, it also uses additional static as well as dynamic risk factors/ criminogenic needs. The wording dynamic risk factors reflects that these are factors which change as the individual passes through their rehabilitative journey of the criminal justice system.

For risk factors in the static model version please refer to its relevant section above.

The additional static risk factors can be grouped into the following:

  • Whether the current offence involved carrying a weapon
  • Whether the person has previous convictions for offences involving a) homicide, b) wounding/grievous bodily harm, c) kidnapping, d) firearms, e) robbery, f) aggravated burglary, g) weapons other than firearms, h) criminal damage endangering life, and i) arson

The additional dynamic risk factors are:

  • Accommodation needs
  • Employment needs
  • Alcohol misuse needs
  • Behavioural and attitude needs

4.2.5 - Model output

A risk score between 0 and 1, representing the probability of proven reoffending with a non-sexual violent offence within the next 2-years from a given date.

4.2.6 - Model architecture

Type of model: Logistic regression implemented using the scikit-learn Python library (see https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html)

Default hyperparameters were used except for the ‘C’ parameter (inverse of regularisation strength) which was tuned using stratified K-fold cross-validation with a randomised grid search (https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RandomizedSearchCV.html)

The default regularisation (‘ridge’ or ‘L2’ regularisation) was used.

4.2.7 - Model performance

Model validation was done using 10-fold cross-validation during model training as well as testing on an unseen dataset. Trained models were evaluated using two key metrics:

  1. c-index, a measure of concordance, equivalent here to the area under the receiver operating characteristic curve. It is a measure of how well a model predicts the order of outcomes, in this case reoffending. Imagine two individuals A and B: A reoffends and B doesn’t within two years of starting their community sentence. If the model gives a higher risk score to A compared to B, then A and B are a concordant pair. The c-index is the proportion of all comparable pairs where the model got the order right. 100% (c-index = 1) means perfect ordering. 50% (c-index = 0.5) means random guessing.

  2. Mean miscalibration, which is the accuracy of the estimate of prevalence of reoffending at the population level. The closer to 0 the better calibrated is the model.

Please see Table 1 for details.

4.2.8 - Datasets and their purposes

Each dataset used for model development contained information from criminal justice system administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. Importantly, there are two types of risk assessments that differ in the amount of information they assess. In short, there is an abbreviated risk assessment called Layer 1, and full risk assessment, called Layer 3. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train the model and evaluate it using cross-validation. This dataset included only individuals in the caseload with complete Layer 3 risk assessments, as only this dataset contained all necessary feature information required to train the model.

A flow dataset containing any offender who started the community part of their sentence between 01 July to 31 December 2018 (with information from administrative systems from probation) was used to test the model. The evaluation dataset only included individuals with a Layer 3 risk assessment (‘Layer 3 evaluation dataset’).

Tier 2 - Model Specification: Direct Contact Sexual Reoffending Predictor (DC-SRP) model (7)

4.2.1. - Model name

Direct Contact Sexual Reoffending Predictor (DC-SRP) model

4.2.2 - Model version

April 2024 release (please note that the model’s name has been changed, it was originally deployed under the name of Offender Assessment System (OASys) Sexual reoffending Predictor for direct contact sexual offences OSP/DC). OSP/DC replaced the OASys Sexual reoffending Predictor for contact sexual offences (OSP-C) which was deployed from 2021 till its replacement.

4.2.3 - Model task

This model is designed to predict the two-year risk of proven reoffending for direct contact sexual offences with children and adults.

4.2.4 - Model input

The DC- SRP model depends on the following risk factors:

  • Gender
  • Sanctions for direct contact adult sexual offences
  • Sanctions for direct contact child sexual offences
  • Sanctions for noncontact sexual offences other than indecent images of children offences
  • Age at start of reoffending follow up
  • Age at last sanction for a sexual offence
  • Any previous criminal history
  • Index sanction includes a direct contact sexual offence with a stranger victim

4.2.5 - Model output

A point score ranging from 0 to 64. Higher scores represent a higher estimated likelihood of proven reoffending on a direct contact sexual offence. The point score is used as the only single predictor in a logistic regression to derive the risk score between 0 and 1, representing the probability of reoffending with a direct contact sexual offence within the next 2-years from a given date. This risk score is used as part of the combined serious risk of reoffending score.

4.2.6 - Model architecture

DC-SRP is a logistic regression model. This model substituted the input features listed above for points scores. This means that the final model has one term only, being the total points score generated by the combination of all the input features.

The DC-SRP model is only calculated for male individuals with a current sexual offence and/or previous history of sexual offending including sexually motivated offences that are not sexual in statute.

When the model is used as part of the CSRP tool, females with current and/or previous sexual offending do not receive an individual DC-SRP score. Instead, a base risk rate is incorporated into their overall CSRP risk score because the number of women with proven sexual reoffending outcomes in the development data was insufficient to develop and validate a reliable female-specific model.

The risk factors used in DC-SRP are scored according to a fixed schema where a defined threshold for a given factor is mapped onto a specific number of points. These points are then summed to give the DC-SRP as a point score.

The exact breakdown of the scoring can be found in Annex A of the following document: “The Actuarial Prediction of Sexual Reoffending Responding to Changing Offending Patterns – Technical Appendix” found at this link https://assets.publishing.service.gov.uk/media/65bbcb1821f73f000de0baf5/actuarial-prediction-sexual-reoffending-annexes.pdf

4.2.7 - Model performance

Model validation was done using on a validation dataset during model training as well as testing on an unseen dataset. Trained models were evaluated using one key metric:

c-index, a measure of concordance, equivalent here to the area under the receiver operating characteristic curve. It is a measure of how well a model predicts the order of outcomes, in this case reoffending. Imagine two individuals A and B: A reoffends and B doesn’t within two years of starting their community sentence. If the model gives a higher risk score to A compared to B, then A and B are a concordant pair. The c-index is the proportion of all comparable pairs where the model got the order right. 100% (c-index = 1) means perfect ordering. 50% (c-index = 0.5) means random guessing.

Please see Table 1 for details.

4.2.8 - Datasets and their purposes

Each dataset used for model development contained information from criminal justice system administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train the model and evaluate it using cross-validation. The DC-SRP model was trained on a subset of cases comprising only male individuals who had a current and or previous history of proven sexual offending.

Tier 2 - Model Specification: Images and Indirect Contact sexual offences against children Sexual Reoffending Predictor (IIC-SRP) model (8)

4.2.1. - Model name

Images and Indirect Contact sexual offences against children Sexual Reoffending Predictor (IIC-SRP) model

4.2.2 - Model version

April 2024 release (please note that the model’s name has been changed, it was originally released under the name of OASys Sexual reoffending Predictor for Images and Indirect Contact sexual offences against children OSP/IIC). It replaced the OASys Sexual reoffending Predictor for Indecent images of children offences (OSP-I) which was deployed from 2021 till its replacement.

4.2.3 - Model task

This model is designed to predict the two-year risk of proven reoffending for indirect contact sexual offences with children and indecent images of children sexual offences.

4.2.4 - Model input

The IIC-SRP depends on the following risk factors:

  • Gender
  • Sanctions for direct contact sexual offences with a child/ children
  • Sanctions for indirect contact sexual offences with a child/ children
  • Sanctions for indecent images of children offences
  • Sanctions for sexual offences that do not involve children (i.e., excluding direct or indirect contact offences with children or indecent image offences). This category includes direct contact sexual offences involving adults and other noncontact sexual offences.

4.2.5 - Model output

A percentage score ranging from 0 to 10.31. Higher scores represent a higher estimated likelihood of proven reoffending for an indirect contact and/or indecent images sexual offence. This risk score is used as part of the combined serious risk of reoffending score.

4.2.6 - Model architecture

The IIC-SRP model is a baseline‑risk model. Individuals are allocated to one of five risk groups using predefined classification rules based on sanction counts. Each group has an associated fixed risk band (low, medium, or high). The model estimates the probability of reoffending with an IIC offence by assigning, as the risk score, the empirical base rates (prevalence) observed for each group in the training data.

For males with a current sexual offence and/or previous history of sexual offending including sexually motivated offences that are not sexual in statute, the model assigns a risk score based on membership in one of five groups:

  • History of sexual offending but no prior direct contact sexual, indirect contact sexual or indecent images of children sanctions
  • 1 sanction for a direct contact sexual offence with a child/ children
  • 2 or more sanctions for direct contact sexual offences with a child/ children
  • 1 sanction for an indecent images of children offence or indirect contact sexual offence with a child/children
  • 2 or more sanctions for indecent images of children offences or indirect contact sexual offences with a child/ children

For females and males with no current or previous history of proven sexual offending, the IIC model assigns no risk score or band. This is due to a) the small number of female sexual offenders on the one hand and b) based on the principle that a risk score is only assigned to those with a current and/ or previous history of proven sexual offending on the other hand.

Females receive no IIC-SRP score because of insufficient female outcomes to develop/validate a reliable model; males without a current/previous proven sexual offending history receive no score because they are outside the model’s intended population.

The exact breakdown of the scoring can be found in Annex A and more details on the modelling in Annex K of the following document: “The Actuarial Prediction of Sexual Reoffending Responding to Changing Offending Patterns – Technical Appendix” found at this link https://assets.publishing.service.gov.uk/media/65bbcb1821f73f000de0baf5/actuarial-prediction-sexual-reoffending-annexes.pdf

4.2.7 - Model performance

Model validation was done using a validation dataset during model training as well as testing on an unseen dataset. Trained models were evaluated using one key metric:

c-index, a measure of concordance, equivalent here to the area under the receiver operating characteristic curve. It is a measure of how well a model predicts the order of outcomes, in this case reoffending. Imagine two individuals A and B: A reoffends and B doesn’t within two years of starting their community sentence. If the model gives a higher risk score to A compared to B, then A and B are a concordant pair. The c-index is the proportion of all comparable pairs where the model got the order right. 100% (c-index = 1) means perfect ordering. 50% (c-index = 0.5) means random guessing.

Please see Table 1 for details.

4.2.8 - Datasets and their purposes

Each dataset used for model development contained information from criminal justice system administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train the model and evaluate it using cross-validation. The IIC-SRP model was trained on a subset of cases comprising only male individuals who had a current and/ or previous history of proven sexual offending.

2.4.3. Development Data

4.3.1 - Development data description

Each dataset used for model development contained information from criminal justice system (probation, prisons, police) administrative data including data from the offender assessment using HMPPS’ digital risk assessment tool. Importantly, there are two types of risk assessments that differ in the amount of information they assess. In short, there is an abbreviated risk assessment called Layer 1, and full risk assessment, called Layer 3. The datasets were also augmented with data from the Police National Computer database to determine whether individuals had indeed reoffended for a given offending category within two-years of starting in the community.

A snapshot dataset of the HMPPS community caseload as of 30 June 2018 (with information from administrative systems from probation) was used to train and evaluate the models contained within the tool.

A flow dataset containing any offender who started the community part of their sentence between 01 July to 31 December 2018 (with information from administrative systems from probation) was used to test the models. There were two versions of the flow dataset: 1) the ‘full evaluation dataset’ including anyone regardless of risk assessment layer type and 2) the ‘Layer 3 evaluation dataset’ including only the subset of individuals who had a Layer 3 risk assessment.

The static models for ARP, VRP and SVRP were trained and cross-validated using the full caseload dataset, regardless of layer type. They were evaluated using both versions of the flow dataset: the ‘full evaluation dataset’ and the ‘Layer 3 evaluation dataset’. This was to allow fair comparison with their equivalent ARP, VRP and SVRP dynamic models which can only be evaluated on this subset of data.

The dynamic models for ARP, VRP and SVRP were trained and cross-validated using individuals in the caseload snapshot dataset with a Layer 3 risk assessment as only this dataset contains all necessary feature information required to train the model. As a result, for the flow dataset, the models were only evaluated on the ‘Layer 3 evaluation dataset’ version.

The DC and IIC SRP models were trained on a subset of cases from the 30 June 2018 community caseload, comprising only male individuals who had previous history of proven sexual offending.

4.3.2 - Data modality

Tabular

4.3.3 - Data quantities

Training data

Static model training dataset: N = 189,097 (raw dataset) N = 184,980 (after NA removal due to NA in features) N = 180,044 (ARP model, after removal of censored individuals) N = 175,873 (VRP model, after removal of censored individuals) N = 170,146 (SVRP model, after removal of censored individuals)

Dynamic model Layer 3 training dataset: N = 127,265 (raw dataset) N = 123,795 (after NA removal due to NA in features) N = 119,248 (ARP model, after removal of censored individuals) N = 115,609 (VRP model, after removal of censored individuals) N = 110,625 (SVRP model, after removal of censored individuals)

The training dataset was not internally split into training and test sets; model testing used the separate evaluation datasets described below.


Test data

Static model full evaluation dataset: N = 81,258 (raw dataset) N = 77,870 (after NA removal due to NA in features) N = 70,240 (ARP model, after removal of censored individuals) N = 66,371 (VRP, after removal of censored individuals) N = 62,010 (SVRP after removal of censored individuals)

Static and dynamic model Layer 3 evaluation dataset: N = 44,684 (raw dataset) N = 42,269 (after NA removal due to NA in features) N = 36,927 (ARP models, after removal of censored individuals) N = 34,039 (VRP models, after removal of censored individuals) N = 30,696 (SVRP models, after removal of censored individuals)


Sexual reoffending predictor dataset (men with history of proven sexual offending): N = 22,231 (raw dataset) N = 20,631 (DC-SRP model, after removal of censored individuals) N = 20,670 (IIC-SRP model, after removal of censored individuals)

The raw dataset was split into a training, validation and test dataset.

4.3.4 - Sensitive attributes

The development data contains personal and sensitive information, including model features relating to personal characteristics, offence details, previous offending history, individual risk and needs factors, and offence-free time. The model features are described in detail in the relevant model specification sections (2.4.2.4). The protected characteristics used as model features are age and gender.

4.3.5 - Data completeness and representativeness

The data had some missingness in individual features which led to cases being removed. Equally, individuals were removed from the dataset due to so-called censoring events (numbers illustrated above). Censoring means an event (e.g. recall to prison) took place which meant an individual’s outcome (reoffending: yes/no) over the full duration of 2 years from their start in the community could not be observed. This is necessary as the reoffending predictors predict risk of reoffending within the next 2-years, thus including a censored individual would be out-of-scope of the tool’s statistical definition.

The training data reflects the HMPPS caseload at the snapshot date. The evaluation data reflects individuals starting in the community, representing the population at the point of initial assessment. In operational use, risk scores are calculated both at initial assessment and when existing cases are reassessed.

4.3.6 - Data cleaning

The development datasets underwent extensive data cleaning, pre-processing and feature engineering to prepare them for model training and evaluation. This included checking and addressing missing, inconsistent or invalid values; standardising and combining information from different administrative data sources; and deriving model features from raw data.

Feature engineering included the construction of variables relating to offending history, offence types, time since previous offending and other relevant risk factors. Transformations were applied as appropriate to ensure the data could be used consistently across the different models.

4.3.7 - Data collection

The datasets were created by the team based on probation, prison and Police National Computer administrative datasets. The datasets were originally used to revalidate operational risk prediction tools (see https://www.gov.uk/government/publications/revalidation-risk-of-recidivism-tools-an-evaluation-of-the-actuarial-instruments-developed-to-assess-recidivism-risk-in-england-and-wales).

4.3.8 - Data access and storage

Only team members working on the project have access to the development data, which is securely stored within the Ministry of Justice’s controlled cloud environment.

4.3.9 - Data sharing agreements

N/A

Tier 2 - Operational Data Specification

4.4.1 - Data sources

Probation Delius (API calls), Prison NOMIS (API calls), user entry.

4.4.2 - Sensitive attributes

Please refer to 2.4.3.4 Sensitive attributes in the section 2.4.3 Development data

4.4.3 - Data processing methods

Data is validated to identify missing values that would prevent a valid calculation of the risk score and band. Where a dynamic model is available the dynamic score will be returned if all required static and dynamic fields are present. Where there is no dynamic model, or in the case where there is such a model, but any required dynamic values are missing the static model score will be returned. In all cases, if a static risk factor is missing, no score is returned. There are checks for implausible values in open-response fields such as the number of past sanctions.

4.4.4 - Data access and storage

Inputs and outputs are stored within the individual’s operational offender/assessment record. Access is restricted through role-based access controls. The tool does not have a separate retention period: its inputs and outputs are retained in accordance with the retention rules applying to the underlying offender/assessment record.

Data retention rules are applied:
These are generally 6 years with no further assessment for all cases except lifers or indeterminate sentences where it is 99 years. We retain cases of special interest, outstanding legal action, historical interest, outstanding inquiry.

4.4.5 - Data sharing agreements

Data from the HMPPS digital risk assessment tool is shared internally within the Ministry of Justice for authorised analytical, research and statistical purposes. Data may also be shared with other government departments or external organisations where there is a lawful basis on which to share operational requirement, subject to appropriate data-sharing arrangements where required.

For example, under Multi-Agency Public Protection Arrangements (MAPPA), Police, Probation and Prison Services work with other agencies to assess and manage the risks posed by individuals convicted of serious violent, sexual and terrorism offences. Risk assessment data may be shared between relevant agencies for this purpose. In particular, the police use SRP risk scores as part of their management of individuals subject to notification requirements following conviction for a sexual offence.

Tier 2 - Risks, Mitigations and Impact Assessments

5.1 - Impact assessments

The tools underwent ethical and fairness assessment to identify and assess relevant risks and appropriate mitigations during design, development and deployment. This included consideration of model performance across relevant demographic groups.

Data protection impact assessments (DPIAs) were obtained for the data used to train and evaluate the models within the tool.

  • DPIA ID 12980 (covering probation, prison and risk assessment data)
  • DPIA ID 8359 (covering probation, prison and risk assessment data augmented with Police National Computer Data)

An equality analysis for the tool and associated models was completed to assess potential equality impacts. The assessment rated the identified impacts as ‘neither or limited’ across the protected characteristics considered.

5.2 - Risks and mitigations

Risk: Machine learning models may perpetuate or reinforce existing biases present in their training data, potentially resulting in unfair treatment of protected groups.

  • Mitigation: We use convicted reoffending data rather than unconvicted offending data sources such as police interactions or arrest records. Convicted reoffending records provide a clearer, legally validated account of an individual’s reoffending history and therefore reduce the likelihood of introducing systemic bias into our models.

Risk: Models may perform unevenly across demographic groups, leading to inequitable treatment.

  • Mitigation: Our equalities impact analysis shows that model performance is appropriate across key relevant groups, including age, gender, ethnicity, and disability. This reduces the risk of disproportionate errors for specific populations.

Risk: Machine learning models can lack transparency, making their outputs difficult to interpret and explain.

  • Mitigation: We use logistic regression, one of the simplest and most interpretable modelling approaches. This enables clear explanations at both a global level and for individual predictions. Practitioners also receive extensive user guidance.

Risk: Using personal data without consent in machine learning models can raise privacy concerns.

  • Mitigation: Our models are only used and trained on data from convicted offenders. All data processing complies with the UK’s data protection framework (UK GDPR, Data Protection Act 2018) and is conducted under lawful public interest basis. Its use has been fully assessed through two DPIAs and strong governance and data‑handling controls are in place as safeguards.

Risk: Data and model storage carry security risks, including the possibility of access by hostile actors.

  • Mitigation: Data is stored within controlled MoJ environments with access restricted to authorised users in accordance with applicable security and data-handling controls.

Risks: Over-reliance on model outputs could result in relevant contextual information not captured by the models being given insufficient weight.

  • Mitigation: The risk posed by offenders on probation is determined by probation officers using all the information available to them and is not set automatically from risk scores.

Risk: Unintended consequences may arise if practitioners misinterpret the outputs of our models or fail to adopt the tool.

  • Mitigation: Guidance and learning materials explain the model inputs, outputs, limitations and appropriate interpretation of risk scores.

Annex: Table 1 - Model performance metrics overview for all models

Model Version Dataset C-index Calibration
ARP Static Full model training dataset 0.8 [CI: 0.797 - 0.802] 0 [CI: -0.002 - 0.002]
VRP Static Full model training dataset 0.801 [CI: 0.799 - 0.804] 0 [CI: -0.002 - 0.002]
SVRP Static Full model training dataset 0.778 [CI: 0.77 - 0.787] 0 [CI: 0 - 0.001]
ARP Dynamic Layer 3 model training dataset 0.811 [CI: 0.808 - 0.813] 0 [CI: -0.002 - 0.002]
VRP Dynamic Layer 3 model training dataset 0.812 [CI: 0.809 - 0.815] 0 [CI: -0.002 - 0.002]
SVRP Dynamic Layer 3 model training dataset 0.787 [CI: 0.778 - 0.796] 0 [CI: -0.001 - 0.001]
ARP Static Full model evaluation dataset 0.803 [CI: 0.799 - 0.806] 0.011 [CI: 0.008 - 0.014]
VRP Static Full model evaluation dataset 0.788 [CI: 0.785 - 0.792] 0.015 [CI: 0.012 - 0.017]
SVRP Static Full model evaluation dataset 0.777 [CI: 0.762 - 0.792] 0.002 [CI: 0.001 - 0.003]
ARP Static Layer 3 evaluation dataset 0.795 [CI: 0.79 - 0.799] -0.014 [CI: -0.019 - -0.01]
ARP Dynamic Layer 3 evaluation dataset 0.802 [CI: 0.797 - 0.806] -0.01 [CI: -0.014 - -0.005]
VRP Static Layer 3 evaluation dataset 0.769 [CI: 0.764 - 0.774] -0.001 [CI: -0.005 - 0.004]
VRP Dynamic Layer 3 evaluation dataset 0.778 [CI: 0.773 - 0.783] 0.002 [CI: -0.003 - 0.006]
SVRP Static Layer 3 evaluation dataset 0.756 [CI: 0.737 - 0.774] 0 [CI: -0.001 - 0.002]
SVRP Dynamic Layer 3 evaluation dataset 0.769 [CI: 0.75 - 0.787] 0.001 [CI: 0 - 0.003]
DC-SRP Static Men with sexual history dataset 0.761 [CI: 0.724 - 0.797] N/A
IIC-SRP Static Men with sexual history dataset 0.765 [CI: 0.747 - 0.782] N/A