Research paper digests

G Callegaro, U Çetin, B D'Auriaa
rXiv, 2026

Cyber attacks rarely arrive at random. Systems that have just been probed or breached tend to attract further activity, and defences periodically "reset" a compromised system back to a clean state. This paper builds a mathematical model that captures exactly this rhythm: a family of processes that drift deterministically between events but are reset at moments inspired by self-exciting attack patterns, where one incident raises the odds of the next.

The framework is deliberately general, so it can describe many practical situations while remaining tractable enough to analyse. Drawing on tools that echo classical insurance ruin theory, the authors study the long-run behaviour of these systems and use it to think about the best moments to intervene, that is, when to reset or reinforce a system to keep it stable and secure. The result is a principled way to reason about optimal defensive responses to clustered, escalating threats.

Access the full paper here.

X Qi, K Ye, C Shi, Y Yang et al. arXiv, 2026

Modern AI image generators have become so sophisticated that the images they produce are often indistinguishable from photographs by the human eye. This poses a real challenge for detecting fake or manipulated content, and most existing detection tools struggle when they encounter images from generators they have not been trained on. This paper proposes a new approach to the problem, borrowing a technique from economics called "difference-in-differences" to identify subtle statistical patterns that distinguish AI-generated images from real ones.

Rather than trying to spot specific visual artefacts, the method looks for systematic differences in how AI images respond to perturbations compared to real photographs. This makes it more robust across different types of generators, including ones the detector has never encountered before, offering a more general-purpose tool for identifying synthetic imagery in the wild.

Access the full paper here.

MG Arouca, A Ribeiro, AM Amorim, ME Barreto et al.
Journal on Interactive Systems, 2026

In many urban communities, especially vulnerable ones, reliable geographic data about health conditions is hard to come by. This paper tackles that gap by proposing a system that uses game-like features, such as points, badges, and challenges, to encourage local residents to contribute information about their surroundings, effectively building detailed maps of Covid-19 spread from the ground up.

The concept is that when people are engaged as active participants rather than passive recipients of public health messaging, the quality and coverage of data improves significantly. By making the mapping process interactive and rewarding, the researchers aim to produce richer, more localised information to support pandemic response, particularly in the kinds of neighbourhoods where top-down data collection tends to fall short.

Access the full paper here.

C Cárdenas-Hurtado, S M Lee, Y Chen, I Moustaki
arXiv, 2025

Diagnostic assessments in education, psychology and healthcare aim to work out which specific skills or attributes a person has, based on how they answer a set of questions. Traditional "cognitive diagnosis models" treat each attribute as either mastered or not, an all-or-nothing view that can be too crude. Newer partial-mastery models allow for degrees of understanding, but they still lean on rigid mathematical assumptions about how mastery translates into answers, which risks misrepresenting the data.

This paper introduces a more flexible approach that keeps the idea of partial mastery but lets the relationship between skill and response be learned from the data rather than fixed in advance. By relaxing those parametric assumptions, the model fits real assessment data more faithfully and gives sharper, more trustworthy estimates of what people actually know, reducing the chance that conclusions are distorted by a poorly chosen model.

Access the full paper here.

SM Lee, Y Chen, T Sit
Journal of the American Statistical Association, 2026

People in contact with the criminal justice system face significant health challenges, yet research into their medical needs has been slow to adopt modern AI tools. This paper conducts a rapid review of how machine learning, deep learning, and natural language processing are being applied in this area, drawing on a systematic search of major medical databases for original research at the intersection of AI and justice health.

The review identified 11 qualifying studies, most focused on machine-learning-based predictions such as suicidality, self-harm, opioid overdoses, and hospital length of stay. Deep learning and NLP applications were rare, and the studies showed limited methodological and geographical diversity. The authors call for greater transparency and rigour in AI-powered justice health research, with the aim of improving health outcomes for one of the most underserved populations in healthcare.

Access the full paper here.

U Çetin, M Lin

arXiv, 2026

In financial markets, some traders act on private information while others trade for unrelated reasons, and the market makers who provide liquidity have to price that uncertainty into their quotes. Most models assume a fixed, known number of informed traders. This paper studies the more realistic case where that number is itself random: liquidity providers know only the odds that informed traders are present and how many there might be, not the actual situation on any given day.

Working within a single-period market, the authors show that a well-defined equilibrium exists and characterise it precisely. They then examine what happens with very large orders, and find that the resulting price impact follows a "power law" whose shape is governed jointly by the range of possible asset values and the distribution of how many informed traders are active. This helps explain why big trades move prices the way they do, and links that behaviour back to underlying uncertainty about who is trading and why.

Access the full paper here.

J Ouyang, Y Chen, C Li, G Xu
arXiv, 2025

International assessments such as PISA are used to compare and rank education systems across the world, shaping policy debates and reforms. But the questions do not always mean the same thing everywhere: differences in language, culture and curriculum can make an item easier or harder in one country than another. When this "measurement bias" is ignored, the resulting league tables can be misleading, penalising or flattering countries for reasons that have nothing to do with actual student ability.

This paper proposes a new statistical method that corrects for such bias without relying on the restrictive assumptions used by earlier approaches, which typically require researchers to nominate in advance a set of "unbiased" questions or a reference group. The method is computationally efficient and comes with theoretical guarantees that it recovers the true ranking reliably. Applied to PISA 2022 data in mathematics, science and reading, it produces corrected country rankings and reveals where and how measurement bias is creeping into the survey.

Access the full paper here.

F Steele, S Zhang, P Clarke
Journal of the Royal Statistical Society (Series A), 2026

People who live together tend to have correlated health, and researchers have long wanted to measure how much a shared household shapes outcomes over time. The difficulty is that households are not fixed: people move in and out, partnerships form and dissolve, and children grow up and leave. This churn creates a tangled web of overlapping relationships that standard statistical models struggle to handle when data are collected across many years.

The authors develop a multilevel model built around what they call "superhouseholds", clusters that link individuals through shared periods of living together, allowing household influences to be correlated in ways that reflect real relationships. Crucially, how strongly two households are connected can depend on the nature of their link. They also design a careful computational procedure to fit the model reliably and release it as a ready-to-use software package, making the method accessible to other researchers studying health, behaviour or other clustered longitudinal data.

Access the full paper here.

L Calais-Ferreira, ME Barreto, R Borschmann.

Journal of Health Equity, 2026

In fields ranging from healthcare to psychology, researchers often need to track many different outcomes for the same individuals over extended periods of time, producing complex datasets where everything is interconnected. Analysing these datasets is technically demanding: the variables affect each other, measurements may be missing at some time points due to patients dropping out or being censored, and the outcomes may be a mixture of continuous and categorical types. This paper proposes a new statistical model that handles all of these challenges by introducing unobserved "latent factors" that capture the shared patterns driving the observed data.

As well as fitting the model to data, the authors develop tools for drawing reliable inferences about the effects of specific covariates on outcomes and for predicting future measurements. They establish a central limit theorem, which provides a mathematical foundation for hypothesis testing, and introduce a principled method for choosing how many latent factors to include. The result is a coherent and flexible framework for making sense of the kinds of rich longitudinal datasets that are becoming increasingly common in social and health research.

Access the full paper here.

S Gong, E Xu, K Ye, F Quinzan, G Livieri, C Shi
arXiv, 2026

Large language models are increasingly trained to reason using reinforcement learning, where the model tries out answers and is rewarded when they are correct. A central challenge is judging how good each attempt is: methods that estimate this precisely tend to be computationally expensive because they generate many attempts per question, while cheaper methods that use only one attempt are far noisier.

This paper introduces BASIS, a training method that keeps the low cost of generating a single attempt per question but sharpens its estimates by borrowing information across all the questions in a training batch. In experiments, this cut the error in value estimation by around two-thirds compared with a standard single-attempt baseline, and matched the accuracy of much more expensive multi-attempt approaches while using substantially less training time. The upshot is faster, cheaper training of reasoning models without sacrificing quality.

Access the full paper here.

J Loftus et al
arXiv, 2026

Automated decision systems are now used to guide high-stakes choices in areas such as criminal justice, healthcare triage and student support. The usual assumption is that if you make the underlying predictions more accurate, better decisions and outcomes will follow. Drawing on real-world case studies, this Perspective argues that this assumption is often wrong: introducing predictions into an organisation changes how people work, assess cases and make decisions, in ways that a narrow focus on accuracy completely misses.

The authors propose an integrated framework that treats these systems not as standalone prediction tools but as interventions embedded in social and organisational settings. By shifting attention from prediction accuracy alone to the full downstream consequences of deployment, the framework offers a more realistic basis for designing, evaluating and governing such systems, and for anticipating their real effects on the people and institutions they touch.

Access the full paper here.

H Zhou, K Ye, E Xu, J Zhu, S Gong, C Shi
arXiv, 2026

Group Relative Policy Optimization (GRPO) is a training technique that has become widely used to improve the reasoning abilities of large language models, including the high-profile DeepSeek family of models. Despite its popularity, the mathematical reasons why it works have remained somewhat unclear. This paper provides a rigorous theoretical explanation by showing that the core mechanism driving GRPO is what mathematicians call a U-statistic, a particular type of estimator with well-understood statistical properties.

The finding matters because it allows researchers to better understand, predict, and improve how GRPO behaves during training. Rather than treating the method as a black box, the team gives it a solid theoretical foundation, opening the door to more principled improvements to how AI models are taught to reason through complex problems.

Access the full paper here.

X Wu, Q Wen, Y Zhang, H Zhu, T Li, C Shi
arXiv, 2026

A/B testing, where two versions of something are compared by randomly assigning users to each, is a cornerstone of decision-making at major technology companies. However, running these tests on time series data, where the same users or systems receive different treatments over a sequence of time points, introduces complications that standard A/B frameworks are not designed to handle. The order of treatment, carryover effects, and shifting conditions over time all make it harder to isolate the true impact of a change.

This paper proposes a new approach to designing such experiments, using transformer models combined with reinforcement learning to intelligently determine when and to whom treatments should be assigned in a time series setting. The result is an adaptive experimental design that accounts for temporal dynamics, producing more reliable causal estimates than existing methods and offering a practical framework for companies trying to evaluate policies that unfold over time.

Access the full paper here.

C Shi, H Zhou, J Zhu, Y Yang
arXiv, 2026

With the rise of large language models (LLMs), distinguishing between human-written and machine-generated text has become increasingly important. This paper proposes a method for detecting LLM-generated content that comes with formal performance guarantees, ensuring reliability under specified conditions.

The authors develop statistical techniques that analyse patterns in text to identify subtle differences between human and machine outputs. Unlike heuristic approaches, their method is grounded in theory, providing measurable confidence in detection accuracy. This work has important implications for academic integrity, misinformation detection, and the broader governance of AI-generated content.

Access the full paper here.

R Pita, G Rocha, M Ichihara, K Harron, P Brito, R Carreiro, B Almeida, P Ramos, M Barreto
International Journal of Population Data Science, 2025

Linking large administrative health and social databases often depends on turning messy, inconsistently written addresses into accurate locations, a process known as geocoding. The pipelines used for this are highly sensitive to "noise" in the text: misspellings, abbreviations, missing elements and varying formats can all throw off the matching, which in Brazil's CIDACS system relies on comparing how similar two strings of text are.

This work tests whether language models can clean up and standardise address text before the matching step, smoothing out that noise so the geocoder has a clearer signal to work with. By improving the quality and consistency of the input addresses, the approach aims to raise geocoding accuracy across very large national datasets, which in turn strengthens the reliability of the research and public-health analyses built on top of that linked data.

Access the full paper here.

Y Chen, J Qiao, Z Ying
Statistica Sinica, 2026

This paper introduces a new statistical method for uncovering hidden structures in large and complex datasets. Traditional factor analysis methods can struggle when data are high-dimensional or when relationships between variables are layered or hierarchical. The authors propose a “divide-and-conquer” approach that breaks a large problem into smaller, more manageable pieces, allowing researchers to detect patterns more efficiently and reliably.

Importantly, the method comes with strong theoretical guarantees, meaning its performance is mathematically justified under well-defined conditions. This makes it particularly appealing for applications in fields like psychology, genomics, or social sciences, where researchers often deal with many interrelated variables. Overall, the paper offers both a practical tool and a rigorous framework for better understanding complex data structures.

Access the full paper here.

G Livieri, T De Angelis, M Ghio
arXiv, 2026

Imagine many firms each deciding when to expand capacity, where each one's best move depends on what everyone else is doing. "Mean-field games" model exactly this kind of large-population strategic interaction. In such problems the optimal strategy is described by a boundary that divides the situation into a region where a firm should act and one where it should wait. A previous study established that this boundary exists but left open how to describe it precisely or compute it.

This paper closes that gap. The authors prove that the boundary is the unique solution of a particular type of integral equation, and along the way establish that the boundary behaves smoothly and extend a fundamental tool of stochastic calculus (Itô's formula) under weaker conditions than before. They also provide an algorithm to solve the equation numerically and compute the optimal strategies in practice, turning an abstract existence result into something usable.

Access the full paper here.

M Azadkia, H Dette
arXiv, 2026

Measuring how strongly two variables depend on each other is a classic problem in statistics, but most traditional measures either miss nonlinear relationships or require strong assumptions. A measure introduced in 2013 by Dette, Siburg, and Stoimenov offered a more flexible alternative, equalling zero under independence and one when the relationship is perfectly deterministic. This paper develops a practical way to estimate this measure from data using kernel smoothing, a technique that replaces the need for sharp boundaries in estimation with smooth, continuous approximations.

The result has practical implications beyond theory. Because the limiting distribution under independence is now well understood, practitioners can construct reliable significance thresholds for independence tests without resorting to permutation methods. The authors support their findings with numerical experiments demonstrating the estimator's sensitivity to local alternatives, offering a sharper and more principled tool for detecting complex dependencies in real data.

Access the full paper here.

P Roudaki, S Gavioli-Akilagun, F Kalinke, M Azadkia, Z Szabo
arXiv, 2026

Understanding whether two variables are related to each other is one of the most fundamental tasks in statistics, but standard tools often struggle when the relationships are complex or when the data is not simple numerical values (for example, images, functions, or networks). This paper introduces a new statistical measure called the Kernel Integrated R², which extends an existing measure of dependence by incorporating a powerful mathematical framework known as reproducing kernel Hilbert spaces. The result is a tool that can detect a much wider range of relationships, including subtle patterns in the tails of distributions and oscillating dependencies.

The new measure retains attractive theoretical properties: it equals zero when two variables are truly independent and reaches its maximum when one variable can be perfectly predicted from the other. Crucially, it works not just with ordinary numbers but with rich, structured data types, making it applicable to a broad range of modern scientific problems where understanding complex interdependencies is essential.

Access the full paper here.

H Zhou, J Zhu, E Xu, K Ye, Y Yang, C Shi
arXiv, 2026

Large language models like GPT, Claude, and Gemini can now produce text that reads as naturally and fluently as something written by a human. While this has many positive applications, it also raises serious concerns around misinformation, academic dishonesty, and the erosion of trust in written content online. Detecting whether a piece of text was written by a human or generated by a machine is therefore an increasingly urgent challenge, and one that existing tools handle inconsistently.

This paper introduces a new detection framework called Learn-to-Distance, which trains a model to recognise the characteristic "distance" in style and structure between human-written and machine-generated text. Rather than relying on fixed rules or shallow surface features, the method learns flexible representations that can distinguish the two sources even when the generated text has been lightly edited or paraphrased. The approach shows strong performance across different models and writing domains, offering a more robust tool for identifying AI-generated content.

Access the full paper here.

A De Palma
arXiv, 2026

Neural networks can be fooled by adversarial examples, small and often imperceptible changes to an input that cause the model to produce a wrong answer. Two main strategies exist for defending against this: adversarial training, which trains models for empirical robustness against known attacks, and certified defences, which provide formal robustness guarantees. While adversarially-trained models may be vulnerable to stronger and unseen attacks, they display good trade-offs between standard performance and empirical robustness.

This paper explores whether knowledge from empirically robust models can be transferred to improve models that are formally certifiable. By using a well-trained empirically robust network as a "teacher" to guide the training of a certifiable "student" network, the work shows that strictly better trade-offs between certifiable robustness and standard performance can be obtained. The work offers a promising path towards better-performing neural networks that are provably safe by construction.

Access the full paper here.

G Livieri, N Detering, L Galimberti, A Kratsios, A. M Neuman
arXiv, 2026

This paper addresses a fundamental problem in machine learning: how to make reliable predictions when only a single network or graph is observed. The authors focus on “small-world” networks, which are common in real-world systems such as social networks or biological systems.

They develop theoretical guarantees for transductive learning, a setting where the goal is to predict missing information within the observed graph rather than generalising to entirely new data. By leveraging the geometric properties of small-world networks, the paper provides conditions under which accurate predictions are possible, advancing both theory and practice in graph-based learning.

Access the full paper here.

J Kim, M Vojnovic
IEEE/ACM Transactions on Networking, 2026

Many modern systems must decide, on the fly, which job to send to which server, from crowdsourcing and labour platforms matching tasks to workers, to computing clusters assigning workloads to machines. The quality of each match depends on features of both the job and the server, but these payoffs are not known in advance and have to be learned from experience, all while keeping the queue of waiting jobs under control.

This paper develops a scheduling algorithm that learns these job-server payoffs as it goes, using a bilinear model to capture how job and server characteristics interact. It combines a fair allocation rule with a learning method to steer assignments toward high-value matches, and the authors prove that it performs almost as well as an idealised policy that knew all the system parameters in advance, while keeping queues stable and respecting job priorities. Experiments on both synthetic and real cluster data confirm the approach works well in practice, including in large-scale, distributed settings.

Access the full paper here.

J Cribeiro-Ramallo, F Kalinke, Z Szabó
arXiv, 2026

A recurring question in statistics and machine learning is how to measure the disagreement between probability distributions, for tasks such as testing whether two samples come from the same source, checking a model's goodness-of-fit, or testing whether two variables are independent. A family of tools called kernel discrepancies (including MMD, HSIC and KSD) does this, and their fastest estimators are known to converge at a certain rate as more data becomes available. What was not fully understood was whether those rates are the best achievable.

This paper settles that question in broad generality. The authors prove matching lower bounds showing that no estimator can converge faster than the known rate, across very general spaces and under mild assumptions, including cases with unbounded kernels that earlier results could not cover. This confirms that the estimators already in use are optimal and gives a clearer picture of the fundamental limits of these widely used statistical tools.

Access the full paper here.

J Loftus, M Buleshnyi, Sakina Hansen
arXiv, 2026

As machine learning models become more complex, understanding how they make decisions is increasingly important. This paper builds on SHAP (Shapley Additive Explanations), a popular method for interpreting model predictions, by introducing a new approach that provides clearer causal insights.

The proposed method, PE-SHAP, focuses on interpreting how individual features contribute to predictions along specific pathways within a model. By incorporating causal reasoning, it moves beyond simple correlations and offers explanations that are more meaningful and reliable. This is especially valuable in high-stakes applications such as healthcare or finance, where transparency and trust are critical.

Access the full paper here.

S Lee, Y Chen, X Li
IEEE Transactions on Information Theory, 2026

Sensor networks are used in many settings, from environmental monitoring to industrial systems, to track whether conditions have shifted significantly over time. Detecting such "change points" as quickly as possible is important, but rushing to raise an alarm also increases the risk of false positives. This paper tackles the problem in a particularly realistic version of the setting: one where the network itself is reconfigurable, meaning sensors can be added or removed over time, and where sensors may influence each other rather than acting independently.

The authors propose new detection procedures based on "e-values," a modern statistical concept that allows rigorous control of the false discovery rate at all points in time, not just in hindsight. This means that at any given moment, the proportion of false alarms among all alarms raised is kept below a specified threshold. Numerical experiments confirm that the approach performs strongly under realistic conditions, advancing the state of the art in online monitoring for complex, evolving sensor systems.

Access the full paper here.

A Kratsios, G Livieri, AM Neuman
arXiv, 2026

As AI systems increasingly use chain-of-thought reasoning, where a model works through a problem step by step before producing an answer, researchers have grown interested in understanding what is actually happening inside the model at each stage. "Reasoning probes" are tools designed to test this, essentially asking whether the model has reached certain conclusions at intermediate points. However, little was known about the theoretical conditions under which such probes can be trusted. This paper addresses that gap by analysing probes in a carefully defined mathematical model of looped reasoning, based on circuits that feed their own outputs back as inputs.

The authors establish formal statistical guarantees for when and how well reasoning probes can capture the internal state of such a system. This is an important step towards building interpretability tools for AI that come with rigorous reliability guarantees, rather than being evaluated only on an ad hoc empirical basis.

Access the full paper here:

D Itkin, B Binkert, P Mangers Bastian, J Teichmann
arXiv, 2026

This paper explores how randomness—or stochasticity—can influence long-term growth in dynamic systems. While traditional models often focus on average behaviour, the authors show that incorporating random fluctuations can significantly improve performance in certain settings, particularly when systems evolve over time under uncertainty.

The concept of ergodicity plays a central role, referring to situations where long-term averages can be inferred from single trajectories. By carefully accounting for stochastic factors, the authors propose strategies that lead to more robust and stable growth outcomes. The results have potential applications in finance, economics, and any domain where uncertainty is a key feature.

Access the full paper here.

J Cribeiro-Ramallo, A Aich, F Kalinke, A Baran Aich, Z Szabo
AISTATS, 2026

When working with statistical models, a fundamental question is how closely a proposed probability distribution matches the true distribution generating the data. Kernel Stein Discrepancy (KSD) is a widely used tool for measuring this kind of mismatch, with applications in areas such as model criticism and sampling algorithms. Until now, researchers knew how fast existing KSD estimators converge as data increases, but it was not known whether those rates were actually the best possible.

This paper closes that gap by proving minimax lower bounds for KSD estimation, that is, the fundamental limits on how quickly any estimator can converge, no matter how clever. The authors present two complementary results covering different settings, including a finding that the difficulty of estimating KSD may grow exponentially with the number of dimensions. This confirms that existing estimators are already optimal, and provides a clearer picture of the inherent difficulty of the goodness-of-fit testing problem in high-dimensional spaces.

Access the full paper here.

A Kratsios, G Livieri, P Schmocker
arXiv, 2026

Many problems in finance, stochastic control and reinforcement learning require representing quantities that evolve randomly through time while respecting a crucial constraint: decisions at any moment can depend only on what has happened so far, not on the future. Classical mathematical tools for describing such processes are powerful in theory but awkward in practice, requiring enormous dictionaries of components and complicated high-order calculations.

This paper introduces NeuralChaos, a neural-network architecture designed to represent these time-evolving, forward-looking processes efficiently, using only a finite number of observations of the underlying randomness while automatically preserving the "no peeking into the future" property. The authors prove that the method can approximate any process in the relevant class and that it achieves the best possible efficiency for well-behaved cases, offering a practical bridge between deep learning and continuous-time stochastic modelling.

Access the full paper here.

Y Chen, S M Lee
arXiv, 2025

When we rank players, teams or products from head-to-head comparisons, standard models such as Bradley-Terry assume a hidden global ordering: if A tends to beat B and B tends to beat C, then A should tend to beat C. Reality is often messier. In games and contests involving multiple skills or strategies, these "rock-paper-scissors" style cycles are common, and models that assume a clean ranking can predict poorly.

This paper proposes a flexible family of models that drop the transitivity assumption, capturing intransitive patterns through a compact mathematical structure. The authors develop practical estimation methods that work even when only a small fraction of possible matchups have been observed, and prove that their approach achieves the best possible accuracy given the sparsity of the data. Simulations and real-data analyses show it outperforms the classical Bradley-Terry model across a wide range of settings.

Access the full paper here.

G Livieri, M Leocata, S Morlacchi, F Corvino, F Flandoli, A Pirni
Technological Forecasting & Social Change (also on arXiv), 2026

This paper investigates how households decide whether to adopt solar panels, using a mathematical model based on Markov processes. The model captures how individuals transition between different states—such as being unaware, considering adoption, or actively installing solar technology—over time.

By simulating these dynamics, the authors explore how social influence, economic incentives, and policy interventions affect adoption rates. The findings shed light on how renewable energy uptake can be encouraged more effectively, offering insights for policymakers aiming to accelerate the transition to sustainable energy systems.

Access the full paper here.

A Pavlis, U Cetin
arXiv, 2026

The Cox-Ingersoll-Ross (CIR) process is a workhorse of quantitative finance, widely used to model interest rates and volatility. Because it cannot usually be simulated exactly, practitioners rely on approximation schemes, and it matters how quickly these approximations converge to the true answer when pricing financial derivatives. Pinning down that convergence speed for the CIR process has historically required rather heavy mathematical machinery.

This paper studies a well-known approximation scheme in the regime where the process stays strictly positive and proves that it converges at a clean, predictable rate as the number of steps increases. Notably, the proof uses only elementary techniques and avoids the more complicated tools used in earlier work, and the authors expect their approach to carry over to a wider class of models, making it a useful and generalisable result for computational finance.

Access the full paper here.

U Çetin, M Lin, G Livieri
arXiv, 2026

In markets with asymmetric information, prices adjust as traders reveal information through their orders, and a large trade is often read as a signal that someone knows something. But large orders can also simply reflect ordinary liquidity needs. This paper asks what happens to that inference when the size of uninformed trading is "heavy-tailed", meaning occasional very large orders arrive that have nothing to do with private information.

The authors show that when big uninformed trades are common enough, large orders stay plausibly innocent over a much wider range, so the market treats them as less informative. This flattens the relationship between trade size and price impact and slows down how quickly prices absorb information, though sufficiently extreme trades can still eventually move the market. The work sheds light on how the statistical shape of liquidity demand influences price discovery and the risks that come with unusually large trades.

Access the full paper here.

S Geneletti, E Guilfoyle, J Pina-Sánchez, A Morales
CrimRxiv, 2026

This study examines whether sentencing outcomes differ across ethnic groups within the justice system, focusing on both Magistrates’ Courts and Crown Courts. By analysing detailed case-level data, the authors aim to determine whether disparities exist after accounting for legally relevant factors such as offence severity and prior records.

The findings contribute to an ongoing and important debate about fairness and bias in judicial decision-making. By comparing different court settings, the paper highlights how institutional context may shape sentencing outcomes. The research provides valuable evidence for policymakers and practitioners seeking to ensure equitable treatment across the legal system.

Access the full paper here.