By Hannah S. Laqueur & Colette Smirniotis

Using firearm transaction and crime gun recovery records from California (2010–2021), we employ machine learning to identify dealers who sold the largest number and highest fraction of guns recovered in crimes within 1 year of sale. This short “time-to-crime” (TTC) is a well-established indicator of potential illegal activity by dealers or traffickers. We developed two primary prediction models: the first classifies dealer-years in the top 5% of 1-year crime gun sales volume (prediction model 1), the second identifies dealer-years in the top 5% based on the fraction of sales recovered within a year (prediction model 2). Both models demonstrated strong discriminative performance, with areas under the receiver operating curve (AUCs) of 0.95 and 0.86, respectively, and areas under the precision-recall curve (AUC–PRs) of 0.72 and 0.43. By comparison, a random classifier would be expected to achieve an AUC of 0.50 and an AUC-PR of 0.05. Prediction model 1 was particularly effective at identifying the highest risk dealers: Those with predictions exceeding 0.90 consistently ranked in the top 5% across multiple years, averaging 33 1-year crime gun sales annually. The machine learning models generally outperformed simpler regression and rule-based approaches, underscoring the value of data-adaptive methods for prediction. Key predictors included prior-year crime gun sales, the average age of purchasers, the proportion of “cheap” handgun sales, and the local gun robbery and assault rate.

Firearms dealers may engage in behaviors that facilitate the diversion of guns to criminal markets. Combining detailed transaction and recovery records with machine learning could help efficiently identify high-risk retailers for targeted enforcement to disrupt the flow of firearms to gun offenders. Future research is needed to determine whether a high number of short TTC sales as compared to a high fraction is a more reliable predictor of law evasion.

Criminology & Public Policy Volume 24, Issue 3 Aug 2025 Pages307-497

By Jack Nicholls; Aditya Kuppa; Nhien-An Le-Khac

Machine Learning and Deep Learning methods are widely adopted across financial domains to support trading activities, mobile banking, payments, and making customer credit decisions. These methods also play a vital role in combating financial crime, fraud, and cyberattacks. Financial crime is increasingly being committed over cyberspace, and cybercriminals are using a combination of hacking and social engineering techniques which are bypassing current financial and corporate institution security. With this comes a new umbrella term to capture the evolving landscape which is financial cybercrime. It is a combination of financial crime, hacking, and social engineering committed over cyberspace for the sole purpose of illegal economic gain. Identifying financial cybercrime-related activities is a hard problem, for example, a highly restrictive algorithm may block all suspicious activity obstructing genuine customer business. Navigating and identifying legitimate illicit transactions is not the only issue faced by financial institutions, there is a growing demand of transparency, fairness, and privacy from customers and regulators, which imposes unique constraints on the application of artificial intelligence methods to detect fraud-related activities. Traditionally, rule based systems and shallow anomaly detection methods have been applied to detect financial crime and fraud, but recent developments have seen graph based techniques and neural network models being used to tackle financial cybercrime. There is still a lack of a holistic understanding of the financial cybercrime ecosystem, relevant methods, and their drawbacks and new emerging open problems in this domain in spite of their popularity. In this survey, we aim to bridge the gap by studying the financial cybercrime ecosystem based on four axes: (a) different fraud methods adopted by criminals; (b) relevant systems, algorithms, drawbacks, constraints, and metrics used to combat each fraud type; (c) the relevant personas and stakeholders involved; (d) open and emerging problems in the financial cybercrime domain.

IEEE Access ( Volume: 9), 2021, 22p.

By Gian Maria Campedelli, Gianmarco Daniele, Marco Le Moglie

Detection is one of the main challenges in the fight against organized crime. We show that machine learning can be used to predict mafias infiltration in Italian local governments, as measured by the dismissal of city councils infiltrated by organized crime. The model successfully predicts up to 96% of out-of-sample municipalities previously identified as infiltrated by mafias, up to two years earlier, making this index a valuable tool for identifying municipalities at risk of infiltration well in advance. Furthermore, we can identify “high-risk” local governments that may be infiltrated by organized crime but have not been detected by the state, thereby improving the efficacy of detection. We then apply this new time-varying measure of organized crime to investigate the underlying causes of this type of rent-seeking. As criminals infiltrate politics to capture public resources, we study how a positive shock in public spending (European Union transfers), affects this phenomenon. Employing a geographic Difference-in-Discontinuities design, we find a substantial and lasting increase in the predicted risk of mafia infiltration (up to 14 p.p.), emphasizing the unintended effects of delivering aid where criminal organizations can appropriate public funds.

Unpublished paper, 2024. 103p.

By: Felix Soldner, Bennett Kleinberg, and Shane D Johnson

Counterfeits harm consumers, governments, and intellectual property holders. They accounted for 3.3% of worldwide trades in 2016, having an estimated value of $509 billion in the same year. While estimations are mostly based on border seizures, we examined openly labeled counterfeits on darknet markets, which allowed us to gather and analyze information from a different perspective. Here, we analyzed data from 11 darknet markets for the period Jan-2014 and Sep-2015. The findings suggest that darknet markets harbor similar counterfeit product types as found in seizures but that the share of watches is higher and lower for electronics, clothes, shoes, and Tobacco on darknet markets. Also, darknet market counterfeits seem to have similar shipping origins as seized goods, with some exceptions, such as a relatively high share (5%) of dark market counterfeits originating from the US. Lastly, counterfeits on dark markets tend to have a relatively low price and sales volume. However, based on preliminary estimations, the original products on the surface web seem to be worth a multiple of the prices of the counterfeit counterparts on darknet markets. Gathering insights about counterfeits from darknet markets can be valuable for businesses and authorities and be cost-effective compared to border seizures. Thus, monitoring darknet markets can help us understand the counterfeit landscape better.

Crime Science (2023) 12:18