Algorithmic approach to forecasting rare violent events

• DOI: http://doi.org/10.1111/1745-9133.12476
• Published date: 01 February 2020
• Author: Richard A. Berk,Susan B. Sorenson
• Date: 01 February 2020

DOI: 10.1111/1745-9133.12476

SPECIAL ISSUE ARTICLE

COUNTERING MASS VIOLENCE IN THE UNITED STATES

Algorithmic approach to forecasting rare violent

events

An illustration based in intimate partner violence perpetration

Richard A. Berk Susan B. Sorenson

University of Pennsylvania

Correspondence

RichardA. Berk, Department of Cr iminology,

McNeilHall, University of Pennsylvania,

Philadelphia,PA 19104.

Email:berkr@sas.upenn.edu

Thoughtfulcomments and suggestions were

providedby colleagues Aaron Chalfin, John

MacDonald,Greg Ridgeway, MichaelKearns,

AaronRoth, and two anonymous reviewers.

Research Summary: Mass violence, almost no matter

how defined, is (thankfully) rare. Rare events are difficult

to study in a systematic manner. Standard statistical

procedures can fail badly, and usefully accurate forecasts

of rare events often are little more than an aspiration. We

offer an unconventional approachfor the statistical analysis

of rare events illustrated by an extensive case study. We

report research aimed at learning about the attributes of

very-high-risk intimate partner violence (IPV) perpetrators

and the circumstances associated with their IPV incidents

reported to the police. “Very high risk” is defined as

having a high probability of committing a repeat IPV

assault in which the victim is injured. Such individuals

represent a very small fraction of all IPV perpetrators;

these acts of violence reported to the police are rare. To

learn about them nevertheless, we sequentially apply in

a novel fashion three algorithms to data collected from a

large metropolitan police department: stochastic gradient

boosting, a genetic algorithm inspired by natural selection,

and agglomerative clustering. We try to characterize not

just perpetrators who on balance are predicted to reoffend

but also who are very likely to reoffend in a manner that

leads to victim injuries. Important lessons for forecasts of

mass violence are presented.

Policy Implications: If one intends to forecast mass vio-

lence, it is probably important to consider approaches less

Criminology & Public Policy. 2020;19:213–233. wileyonlinelibrary.com/journal/capp © 2019 American Society of Criminology 213

214 BERK AND SORENSON

dependent on statistical procedures common in criminol-

ogy.Given that one needs to “fatten” the right tail of the rare

events distribution, a combination of supervised machine

learning and genetic algorithms may be a useful approach.

One can then study a synthetic population of rare events

almost as if they were an empirical population of rare

events. Variants on this strategy are increasingly common

in machine learning and causal inference. Our overall goal

is to unearth predictors that forecast well. In the absence of

sufficiently accurate forecasts, scarce resources to help pre-

vent mass violence cannot be allocated where they are most

needed.

KEYWORDS

forecasting, genetic algorithms, intimate partner violence, machine learn-

ing, mass violence, synthetic populations

Forecasts of risk are routinely made in a wide variety of situations. What is the probability that a

hurricane will strike the Gulf Coast in a particular hurricane season? What is t he probability that a

given high school student will be accepted by his or her college of choice? What is the probability

that a particular business firm will declare bankruptcy? Coupled with each probability is the expected

cost should the event of concern occur. For the bankruptcy example, repayment of debt at 10 cents on

the dollar means a loss of 90 cents for every dollar invested. Risk formally is defined as the costs of a

particular event multiplied by the probability that the event will occur.

Forecasts of risk can be useful if they lead to actions that are better informed. For undesirable

outcomes, one hopes that prevention strategies can be implemented or that plans for remedial action

after the fact can be made. This has long been well understood by criminal justice decision makers

in the United States. Indeed, risk assessments have been used to inform criminal justice decisions

since the 1920s (Burgess, 1928). One might wonder, therefore, whether forecasts of risk might be

instructive for contemporary incidents of mass violence. Without good forecasts, scarce prevention

and remedial resources easily can be misallocated. One cannot, for example, place armed guards at

every church, mosque, or synagogue. Likewise, one cannot have grief counsellors at every business

establishment. Currently, those resources are distributed in a haphazard manner. Might legitimate risk

assessments help?

For almost any reasonable definition of mass violence, constructing sufficiently accurate forecasts

is a daunting undertaking. This holds whether one is trying to forecast the likely perpetrator, location,

or timing of an event. One obstacle is that mass violence is heterogeneous. It can include school

shootings, homicides committed by disgruntled employees, brutal hate crimes, systematic execution

of witnesses at a crime scene, fatal assaults by perpetrators of intimate partner violence (IPV), and

other mass violence in which the motives are obscure (e.g., the October 2017 Las Vegas music

festival mass shooting in which 58 people were killed and 851 wereinjured). Alt hough understanding

mass violence in general is an admirable aspiration, in the medium term, different forms of mass

violence might be productively examined separately. Useful forecasts probably will require different