Should We Use the RFGV Algorithm in Domestic Abuse?

In recent years, the UK government has stepped up its commitment to tackling violence against women and girls (VAWG), launching a series of measures designed to improve prevention, enforcement, and victim protection. As part of this strategy, the Home Office has introduced new offender-focused policies — including halving the amount if VAWG over the next decade. One such approach, endorsed by national guidance, is the use of perpetrator prioritisation frameworks, like the Recency, Frequency, Gravity, Victims (RFGV) model and the College of Policing recommends RFGV-style tools as part of a broader offender management strategy, suggesting law enforcement should use past offending patterns to identify and target individuals most likely to cause serious harm. But the use of frameworks like RFGV in the context of domestic abuse might not just be inappropriate — it might actively obscure the very harm it’s trying to prevent.

Models like RFGV aim to help police identify high-harm domestic abuse perpetrators by scoring their behaviours numerically. It uses recency of offences, frequency of incidents, gravity of harm and number of victims. Each of these is given a score based on, for example, the number of incidents an offender has been involved in or the number of victims they have offended against over a certain period of time. The scores for each component are then entered into a formula which gives a final score of between 0 and 100.

This offers a simple answer for many forces which are flooded with high-risk cases, often with no clear way to triage which perpetrators to review first. Ranking offenders using RFGV can provide a good starting point by identifying repeat offenders, recent incidents, and high-gravity events. This can help police focus limited resources on individuals who are statistically more likely to reoffend or cause harm - the repeat, high-gravity offenders who shouldn’t fall through the cracks.

However, domestic abuse isn’t just a series of isolated incidents—it’s a pattern of control and manipulation. Some of the most dangerous perpetrators don’t show up with high scores in any RFGV category. They may have only a handful of recorded incidents or have only abused one partner. Many domestic homicides and suicides occur because of perpetrators previously rated as medium risk—people who weren’t being actively monitored and which RFGV would not flag. And herein lies the danger: a tool that looks objective but is blind to the nuances may lull decision-makers into a false sense of clarity.

One of the issues with RFGV is cliff-edge scoring where one perpetrator may be labelled high risk where another is evaluated as medium risk due to scoring either side of a threshold with no real-world difference in behaviour between those two cases (e.g., 20 incidents = 100 points, whereas 21 incidents  = 150 points). Vital context that could add more nuance to RFGV —like psychological abuse, separation triggers, suicidal ideation, or controlling behaviours—is often buried in free-text logs, informal notes, or not recorded at all. In many domestic abuse cases, separate incidents of abuse — new acts of harm or escalation — are not even logged as new records. Instead, they are simply added as narrative to the original case. If an officer determines that a new report is "part of an ongoing matter," it can be entered as a note rather than a new incident. This means the RFGV model has no way to register that added harm, and therefore it does not get included in the calculations. Even if domestic abuse incidents are given their own record those that involve elements like coercive control, and psychological and verbal abuse are likely to be downplayed compared to those involving physical violence because they have low sentences and are therefore lower gravity scores.

This means RFGV is inherently skewed towards perpetrators whose abuse is episodic and clearly delineated, and physically violent—not those whose abuse is persistent, cumulative, or relationally complex. In fact, the more nuanced and interconnected the abuse is, the more likely it is to be invisible to the algorithm.

This suggests that RFGV should be modified to include factors such as non-fatal strangulation, suicidal threats, coercive control indicators, MARAC flags, or recent separation. And some of these elements are now being considered for inclusion in more advanced risk tools. But at this point, you’re no longer using RFGV—you’re building a custom risk model. And that might just be the best way forward.