publications

2026

1. Elsevier

   Order flow and cryptocurrency returns

   Alexia Anastasopoulos, Nikola Gradojevic, Fred Liu, and 2 more authors

   Journal of Financial Markets, 2026

2025

1. Wiley

   Deep Learning on Embedded Systems: A Hands-On Approach Using Jetson Nano and Raspberry Pi

   Tariq M Arif

   2025
2. Procedia

   Benchmarking deep learning models on nvidia jetson nano for real-time systems: An empirical investigation

   Tushar Prasanna Swaminathan, Christopher Silver, Thangarajah Akilan, and 1 more author

   Procedia Computer Science, 2025
3. ACM

   Kernel Float: Unlocking Mixed-Precision GPU Programming

   Stijn Heldens, and Ben Werkhoven

   ACM Transactions on Mathematical Software, 2025
4. AJPAS

   Applications of Hidden Markov Models in Detecting Regime Changes in Bitcoin Markets

   Elijah Wanjala Machimbo, Wycliffe Cheruiyot, George Mocheche, and 1 more author

   Asian Journal of Probability and Statistics, 2025

   Abs

   This study explores the identification and assessment of regime shifts in Bitcoin markets through the application of advanced statistical models, namely HMMs, MSMs, and Threshold Models. The analysis utilizes key financial indicators including market capitalization, volatility, trading volume, and historical Bitcoin price data, along with statistical measures such as mean, minimum, and maximum values to enhance the detection of market patterns. Distinctions are made between bullish (sustained price increases exceeding 20%), bearish (sustained price declines exceeding 20%), and neutral (periods of low volatility and sideways movement) market regimes. HMMs provide predictive insights into market transitions, MSMs are employed to capture structural regime changes, and the Threshold Model identifies significant price behaviors.

2024

1. OUP

   The data quality problem (in the European Financial Data Space)

   M Konrad Borowicz

   International Journal of Law and Information Technology, 2024

   Abs

   By creating the European Financial Data Space (EFDS), the European Commission aims to foster the development of data-driven innovation in the financial sector. Artificial intelligence (AI) is the most notable innovation that could benefit from enhanced access to data through the EFDS. However, the current policy framework establishing the EFDS suffers from a major limitation—it neglects the question of data quality or whether the data shared in the EFDS will be fit for the purpose of the development of AI applications. The AI Act and General Data Protection Regulation could play a role in fostering data quality in this area, but the two frameworks are also unlikely to do that to a satisfactory extent. To address the data quality problem, this article proposes policymakers should develop a framework for data quality that encompasses both a conceptual and an organizational dimension aimed at promoting the free flow of high-quality data in the EFDS.

2. Packt

   Efficient Cloud FinOps: A practical guide to cloud financial management and optimization with AWS, Azure, and GCP

   Alfonso San Miguel Sánchez, and Danny Obando Garcı́a

   2024

2023

1. Elsevier

   A Literature Review on Digital Twins in Warehouses

   Marie-Jane Bélanger, Robert Pellerin, and Samir Lamouri

   Procedia Computer Science, 2023

   CENTERIS – International Conference on Enterprise Information Systems / ProjMAN – International Conference on Project MANagement / HCist – International Conference on Health and Social Care Information Systems and Technologies 2022

   Abs HTML

   Due to the high cost of their warehousing operations, many manufacturing companies are looking to increase the efficiency of their warehouse. To this end, Industry 4.0 technologies offer several solutions to address this issue. One of them is the digital twin (DT) technology which is relatively new. This paper presents a literature review on DTs in warehousing. The analysis framework used to classify the papers includes the general description of the DT, the possible uses, and the required data. Our results show that several limitations remain to exploit the full potential of a DT. The limitations mainly concern the characterization of the data required.

2. arXiv

   Report of the DOE/NSF Workshop on Correctness in Scientific Computing, June 2023, Orlando, FL

   Maya Gokhale, Ganesh Gopalakrishnan, Jackson Mayo, and 3 more authors

   Dec 2023

   Abs PDF

   It has been observed in an ongoing study that the floating-point details of GPUs—especially Tensor Cores—can vary in subtle ways with respect to their rounding behaviors, which can hinder direct cross-platform software porting with assurance.

3. Wiley

   Deep order flow imbalance: Extracting alpha at multiple horizons from the limit order book

   Petter N Kolm, Jeremy Turiel, and Nicholas Westray

   Mathematical Finance, Dec 2023

   Abs

   We employ deep learning in forecasting high‐frequency returns at multiple horizons for 115 stocks traded on Nasdaq using order book information at the most granular level. While raw order book states can be used as input to the forecasting models, we achieve state‐of‐the‐art predictive accuracy by training simpler “off‐the‐shelf” artificial neural networks on stationary inputs derived from the order book. Specifically, models trained on order flow significantly outperform most models trained directly on order books. Using cross‐sectional regressions, we link the forecasting performance of a long short‐term memory network to stock characteristics at the market microstructure level, suggesting that “information‐rich” stocks can be predicted more accurately. Finally, we demonstrate that the effective horizon of stock specific forecasts is approximately two average price changes.

2022

1. O’Reilly

   Fundamentals of data engineering

   Joe Reis, and Matt Housley

   Dec 2022
2. O’Reilly

   Data Quality Engineering in Financial Services

   Brian Buzzelli

   Dec 2022
3. T&F

   How much telematics information do insurers need for claim classification?

   Francis Duval, Jean-Philippe Boucher, and Mathieu Pigeon

   North American Actuarial Journal, Dec 2022

   Abs

   It has been shown several times in the literature that telematics data collected in motor insurance help to better understand an insured’s driving risk. Insurers who use these data reap several benefits, such as a better estimate of the pure premium, more segmented pricing, and less adverse selection. The flip side of the coin is that collected telematics information is often sensitive and can therefore compromise policyholders’ privacy. Moreover, due to their large volume, this type of data is costly to store and hard to manipulate. These factors, combined with the fact that insurance regulators tend to issue more and more recommendations regarding the collection and use of telematics data, make it important for an insurer to determine the right amount of telematics information to collect. In addition to traditional contract information such as the age and gender of the insured, we have access to a telematics dataset where information is summarized by trip. We first derive several features of interest from these trip summaries before building a claim classification model using both traditional and telematics features. By comparing a few classification algorithms, we find that logistic regression with lasso penalty is the most suitable for our problem. Using this model, we develop a method to determine how much information about policyholders’ driving should be kept by an insurer. Using real data from a North American insurance company, we find that telematics data become redundant after about 3 months or 4000 km of observation, at least from a claim classification perspective.

2021

1. IEEE

   Deep Learning’s Diminishing Returns: The Cost of Improvement is Becoming Unsustainable

   Neil C. Thompson, Kristjan Greenewald, Keeheon Lee, and 1 more author

   IEEE Spectrum, Dec 2021

   Abs HTML

   Deep learning is now being used to translate between languages, predict how proteins fold, analyze medical scans, and play games as complex as Go, to name just a few applications of a technique that is now becoming pervasive. Success in those and other realms has brought this machine-learning technique from obscurity in the early 2000s to dominance today. • Although deep learning’s rise to fame is relatively recent, its origins are not. In 1958, back when mainframe computers filled rooms and ran on vacuum tubes, knowledge of the interconnections between neurons in the brain inspired Frank Rosenblatt at Cornell to design the first artificial neural network, which he presciently described as a “pattern-recognizing device.” But Rosenblatt’s ambitions outpaced the capabilities of his era—and he knew it. Even his inaugural paper was forced to acknowledge the voracious appetite of neural networks for computational power, bemoaning that “as the number of connections in the network increases…the burden on a conventional digital computer soon becomes excessive.” • Fortunately for such artificial neural networks—later rechristened “deep learning” when they included extra layers of neurons—decades of Moore’s Law and other improvements in computer hardware yielded a roughly 10-million-fold increase in the number of computations that a computer could do in a second. So when researchers returned to deep learning in the late 2000s, they wielded tools equal to the challenge.

2. Springer

   Software Engineering as an Alchemical Process: Establishing a Philosophy of the Discipline

   Manuel Mazzara, Mirko Farina, Adéla Krylová, and 2 more authors

   In Frontiers in Software Engineering, Dec 2021

   Abs

   Far from being a bizarre pastime, alchemy played a crucially important role in the history of science, being supported and promoted by leading political and scientific figures (such as Rudolf II, Jābir ibn Hayyān, Gerard of Cremona, Adelard of Bath, Roger Bacon, Paracelsus, and even Newton). To understand alchemy, however, one has to approach it from both a material and a spiritual (perhaps philosophical) perspective. On the one hand, alchemists wanted to transform, or better transmute, materials (such as lead into gold). On the other hand, though, alchemists were also aiming at transforming qualities and aspects of themselves. In this paper, we show that Computer Science, and in particular Software Engineering, can be partly understood as alchemical processes. We thus draw analogies and specify points of contact between these two, prima facie, distinct and very distant worlds. In doing so, we also formulate and discuss a number of important questions regarding the nature and metaphysics of computation, that can be of interests to many researchers in computer science.

3. Springer

   Deep learning, deep change? Mapping the evolution and geography of a general purpose technology

   Joel Klinger, Juan Mateos-Garcia, and Konstantinos Stathoulopoulos

   Scientometrics, Dec 2021

   HTML
4. SIAM

   A stochastic partial differential equation model for limit order book dynamics

   Rama Cont, and Marvin S Muller

   SIAM Journal on Financial Mathematics, Dec 2021

   Abs

   We propose an analytically tractable class of models for the dynamics of a limit order book, described through a stochastic partial differential equation with multiplicative noise for the order book centered at the mid-price, along with stochastic dynamics for the mid-price which is consistent with the order flow dynamics. We provide conditions under which the model admits a finite-dimensional realization driven by a (low-dimensional) Markov process, leading to efficient methods for estimation and computation. We study two examples of parsimonious models in this class: a two-factor model and a model in which the order book depth is mean reverting. For each model we perform a detailed analysis of the role of different parameters, study the dynamics of the price, order book depth, volume, and order imbalance, provide an intuitive financial interpretation of the variables involved, and show how the model reproduces statistical properties of price changes, market depth, and order flow in limit order markets.

5. Elsevier

   Measuring information flow among international stock markets: An approach of entropy-based networks on multi time-scales

   Peng-Cheng Kuang

   Physica A: Statistical Mechanics and its Applications, Dec 2021

   Abs

   Information flow among stock markets leads to the co-movement of stock prices and even the financial risk contagion globally. Measuring information flow on multi time-scales helps to develop optimal portfolio strategy for market participants and establish targeted financial regulation for policy makers, but is a challenge for traditional econometrics. This study aims to construct the information flow networks on multi time-scales among international stock markets. First, we use the data of 31 stock market indices during 2007 to 2018, and decompose each of stock index series into short and mid-long time-scale components by empirical mode decomposition (EMD). Then, we extend a novel concept in information science, called transfer entropy (TE), to measure information flow and characterize causality structure between two stock markets, and find that information flows inside of EU stock markets only cluster on short time-scale but not on mid-long time-scale. Moreover, both out-node and in-node centrality measurements for developed markets in the entropy-based networks are much higher on short time-scale, indicating that these markets are more dominant but more vulnerable to short-term risk contagion. Finally, based on the minimum spanning tree (MST) model, we extract the most probable risk contagion paths from entropy-based networks. The results show that outflow MSTs on short time-scale follow a star-like structure to precipitate the risk contagion in the international stock markets, especially during crisis period (2007∼ 2010). Overall, we emphasize the financial risk contagion on short time-scale.

6. MDPI

   Synthetic dataset generation of driver telematics

   Banghee So, Jean-Philippe Boucher, and Emiliano A Valdez

   Risks, Dec 2021

   Abs

   This article describes the techniques employed in the production of a synthetic dataset of driver telematics emulated from a similar real insurance dataset. The synthetic dataset generated has 100,000 policies that included observations regarding driver’s claims experience, together with associated classical risk variables and telematics-related variables. This work is aimed to produce a resource that can be used to advance models to assess risks for usage-based insurance. It follows a three-stage process while using machine learning algorithms. In the first stage, a synthetic portfolio of the space of feature variables is generated applying an extended SMOTE algorithm. The second stage is simulating values for the number of claims as multiple binary classifications applying feedforward neural networks. The third stage is simulating values for aggregated amount of claims as regression using feedforward neural networks, with number of claims included in the set of feature variables. The resulting dataset is evaluated by comparing the synthetic and real datasets when Poisson and gamma regression models are fitted to the respective data. Other visualization and data summarization produce remarkable similar statistics between the two datasets. We hope that researchers interested in obtaining telematics datasets to calibrate models or learning algorithms will find our work ot be valuable.

2020

1. MDPI

   The Neural Network Zoo

   Stefan Leijnen, and Fjodor van Veen

   Proceedings, Dec 2020

   Abs HTML PDF

   An overview of neural network architectures is presented. Some of these architectures have been created in recent years, whereas others originate from many decades ago. Apart from providing a practical tool for comparing deep learning models, the Neural Network Zoo also uncovers a taxonomy of network architectures, their chronology, and traces back lineages and inspirations for these neural information processing systems.

2. AAAI

   Energy and Policy Considerations for Modern Deep Learning Research

   Emma Strubell, Ananya Ganesh, and Andrew McCallum

   Proceedings of the AAAI Conference on Artificial Intelligence, Apr 2020

   Abs HTML PDF

   The field of artificial intelligence has experienced a dramatic methodological shift towards large neural networks trained on plentiful data. This shift has been fueled by recent advances in hardware and techniques enabling remarkable levels of computation, resulting in impressive advances in AI across many applications. However, the massive computation required to obtain these exciting results is costly both financially, due to the price of specialized hardware and electricity or cloud compute time, and to the environment, as a result of non-renewable energy used to fuel modern tensor processing hardware. In a paper published this year at ACL, we brought this issue to the attention of NLP researchers by quantifying the approximate financial and environmental costs of training and tuning neural network models for NLP (Strubell, Ganesh, and McCallum 2019). In this extended abstract, we briefly summarize our findings in NLP, incorporating updated estimates and broader information from recent related publications, and provide actionable recommendations to reduce costs and improve equity in the machine learning and artificial intelligence community

3. arXiv

   Yolov4: Optimal speed and accuracy of object detection

   Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao

   arXiv preprint arXiv:2004.10934, Apr 2020

   Abs PDF

   There are a huge number of features which are said to improve Convolutional Neural Network (CNN) accuracy. Practical testing of combinations of such features on large datasets, and theoretical justification of the result, is required. Some features operate on certain models exclusively and for certain problems exclusively, or only for small-scale datasets; while some features, such as batch-normalization and residual-connections, are applicable to the majority of models, tasks, and datasets. We assume that such universal features include Weighted-Residual-Connections (WRC), Cross-Stage-Partial-connections (CSP), Cross mini-Batch Normalization (CmBN), Self-adversarial-training (SAT) and Mish-activation. We use new features: WRC, CSP, CmBN, SAT, Mish activation, Mosaic data augmentation, CmBN, DropBlock regularization, and CIoU loss, and combine some of them to achieve state-of-the-art results: 43.5% AP (65.7% AP50) for the MS COCO dataset at a realtime speed of  65 FPS on Tesla V100.

2019

1. MIT

   Training a single AI model can emit as much carbon as five cars in their lifetimes

   Karen Hao

   Apr 2019

   Abs HTML

   The artificial-intelligence industry is often compared to the oil industry: once mined and refined, data, like oil, can be a highly lucrative commodity. Now it seems the metaphor may extend even further. Like its fossil-fuel counterpart, the process of deep learning has an outsize environmental impact.

2. Elsevier

   Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

   M. Raissi, P. Perdikaris, and G.E. Karniadakis

   Journal of Computational Physics, Apr 2019

   Abs HTML PDF

   We introduce physics-informed neural networks – neural networks that are trained to solve supervised learning tasks while respecting any given laws of physics described by general nonlinear partial differential equations. In this work, we present our developments in the context of solving two main classes of problems: data-driven solution and data-driven discovery of partial differential equations. Depending on the nature and arrangement of the available data, we devise two distinct types of algorithms, namely continuous time and discrete time models. The first type of models forms a new family of data-efficient spatio-temporal function approximators, while the latter type allows the use of arbitrarily accurate implicit Runge–Kutta time stepping schemes with unlimited number of stages. The effectiveness of the proposed framework is demonstrated through a collection of classical problems in fluids, quantum mechanics, reaction–diffusion systems, and the propagation of nonlinear shallow-water waves.

3. Springer

   Financial cybersecurity risk management: Leadership perspectives and guidance for systems and institutions

   Paul Rohmeyer, and Jennifer L Bayuk

   Apr 2019
4. O’Reilly

   Cloud FinOps: collaborative, real-time cloud financial management

   JR Storment, and Mike Fuller

   Apr 2019

2018

1. O’Reilly

   Designing event-driven systems

   Ben Stopford

   Apr 2018
2. Elsevier

   Analyzing volatility transmission using group transfer entropy

   Thomas Dimpfl, and Franziska J Peter

   Energy Economics, Apr 2018

   Abs

   We analyze the transmission of volatility between the oil, stock, gold, and currency markets using transfer entropy. Our approach, which we denote by group transfer entropy, measures the overall sensitivity of a process to lagged realizations of a group of other processes. We supplement our measure by a suitable block bootstrap approach which allows to conduct inference. Our empirical analysis uses daily data of OVX, VIX, GVZ, and EVZ from 2008 to 2017. The results show that oil and stock market volatility are the most affected by past volatility changes in the other markets, which highlights the importance for investors to hedge against this predictable risk transmission. Furthermore, we show that group transfer entropy can unveil nonlinear dynamics beyond the standard vector autoregressive methodology and provides a useful basis for forecast averaging.

2017

1. ACM

   ImageNet classification with deep convolutional neural networks

   Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton

   Commun. ACM, May 2017

   Abs HTML PDF

   We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0%, respectively, which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implementation of the convolution operation. To reduce overfitting in the fully connected layers we employed a recently developed regularization method called "dropout" that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.

2. My take on Ali Rahimi’s ‘Test of Time’ award talk at NIPS

   Yann LeCun

   Facebook, https://www. facebook. com/yann. lecun/posts/101 54938130592143, May 2017

   PDF
3. O’Reilly

   Designing data-intensive applications: The big ideas behind reliable, scalable, and maintainable systems

   Martin Kleppmann

   May 2017

2016

1. IEEE

   You Only Look Once: Unified, Real-Time Object Detection

   Joseph Redmon, Santosh Divvala, Ross Girshick, and 1 more author

   In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2016

   Abs HTML PDF

   We present YOLO, a new approach to object detection. Prior work on object detection repurposes classifiers to perform detection. Instead, we frame object detection as a regression problem to spatially separated bounding boxes and associated class probabilities. A single neural network predicts bounding boxes and class probabilities directly from full images in one evaluation. Since the whole detection pipeline is a single network, it can be optimized end-to-end directly on detection performance. Our unified architecture is extremely fast. Our base YOLO model processes images in real-time at 45 frames per second. A smaller version of the network, Fast YOLO, processes an astounding 155 frames per second while still achieving double the mAP of other real-time detectors. Compared to state-of-the-art detection systems, YOLO makes more localization errors but is less likely to predict false positives on background. Finally, YOLO learns very general representations of objects. It outperforms other detection methods, including DPM and R-CNN, when generalizing from natural images to other domains like artwork.

2. Wiley

   Price discovery in futures and options markets

   Naomi Boyd, and Peter Locke

   Journal of Futures Markets, Jun 2016

   PDF

2015

1. SNL-NM

   Getting the right answer despite incorrect hardware.

   Mark Frederick Hoemmen

   Jun 2015

2013

1. Wiley

   Paul Wilmott on quantitative finance

   Paul Wilmott

   Jun 2013

2010

1. Wiley

   Security engineering: a guide to building dependable distributed systems

   Ross Anderson

   Jun 2010
2. Springer

   A multi agent model for the limit order book dynamics

   Marco Bartolozzi

   The European Physical Journal B, Jun 2010

   Abs

   In the present work we introduce a novel multi-agent model with the aim to reproduce the dynamics of a double auction market at microscopic time scale through a faithful simulation of the matching mechanics in the limit order book. The agents follow a noise decision making process where their actions are related to a stochastic variable, the market sentiment, which we define as a mixture of public and private information. The model, despite making just few basic assumptions over the trading strategies of the agents, is able to reproduce several empirical features of the high-frequency dynamics of the market microstructure not only related to the price movements but also to the deposition of the orders in the book.

2009

1. Wiley

   Market Risk Analysis

   Carol Alexander

   Jun 2009
2. OUP

   A dynamic model of the limit order book

   Ioanid Roşu

   The Review of Financial Studies, Jun 2009

   Abs

   This paper presents a model of an order-driven market where fully strategic, symmetrically informed liquidity traders dynamically choose between limit and market orders, trading off execution price and waiting costs. In equilibrium, the bid and ask prices depend only on the numbers of buy and sell orders in the book. The model has a number of empirical predictions: (i) higher trading activity and higher trading competition cause smaller spreads and lower price impact; (ii) market orders lead to a temporary price impact larger than the permanent price impact, therefore to price overshooting; (iii) buy and sell orders can cluster away from the bid-ask spread, generating a hump-shaped order book; (iv) bid and ask prices display a comovement effect: after, e.g., a sell market order moves the bid price down, the ask price also falls, by a smaller amount, so the bid-ask spread widens; (v) when the order book is full, traders may submit quick, or fleeting, limit orders.

2007

1. Springer

   Information security economics–and beyond

   Ross Anderson, and Tyler Moore

   In Annual international cryptology conference, Jun 2007

2005

1. SIAM

   SRT division algorithms as dynamical systems

   Mark McCann, and Nicholas Pippenger

   SIAM Journal on Computing, Jun 2005

2000

1. APS

   Measuring information transfer

   Thomas Schreiber

   Physical review letters, Jun 2000

   Abs

   An information theoretic measure is derived that quantifies the statistical coherence between systems evolving in time. The standard time delayed mutual information fails to distinguish information that is actually exchanged from shared information due to common history and input signals. In our new approach, these influences are excluded by appropriate conditioning of transition probabilities. The resulting transfer entropy is able to distinguish effectively driving and responding elements and to detect asymmetry in the interaction of subsystems.

1995

1. IEEE

   Computational aspects of the Pentium affair

   Tim Coe, Terje Mathisen, Cleve Moler, and 1 more author

   IEEE Computational Science and Engineering, Jun 1995