Mason Journals (George Mason Univ.)
Not a member yet
3256 research outputs found
Sort by
Disruptive Innovation in Finance: A Cross-Vertical Analysis of Technology-Driven Fintech Firms
No full textFintech startups are reshaping financial services through technologies like AI, blockchain, and big data, often outperforming traditional banks in speed, efficiency, and accessibility. This study examines five leading firms—Chime, Stripe, Ramp, Alloy, and Fireblocks—across key verticals such as neobanking, payments, and digital asset security. Using data from public sources and web scraping, we analyze tech adoption patterns with Python-based tools. Findings show widespread use of AI and big data for fraud prevention, onboarding, and identity verification, while blockchain is applied more narrowly in crypto security. These firms show rapid growth, prompting legacy institutions to respond via partnerships and digital innovation. The study highlights how technology is driving disruption and reshaping competition in finance
How do women access maternal care? Spatial and social dimensions of maternal healthcare visits in Florida
No full textUnequal access to maternal healthcare is a persistent driver of poor maternal and neonatal outcomes, particularly in rural and socioeconomically marginalized U.S. communities. While spatial access is often assessed using proximity or travel time to the nearest facility, such assumptions overlook the realities of where women actually seek prenatal and childbirth care. Overcoming the limitations of traditional access measures, this project examines the spatio-social dimensions of maternal healthcare visit patterns to better understand how women actually access care during their pregnancy and the geographic and social barriers they encounter. The patient visit data includes the total number of prenatal visits between origin-destination (OD) pairs, with origins defined as the women’s residential zip codes and destinations as the healthcare facilities they visited. The data also contains patients’ OD visit patterns disaggregated by age group, race, and health conditions. The OD data were merged with zip code-level demographic, health, and rural-urban indicators, including insurance coverage, income, transportation access, food insecurity, and chronic disease prevalence. Using these zip code-level datasets, we clustered patients’ origin into socioeconomic and health-status groups using hierarchical clustering methods to explore the neighborhood contexts behind these patterns. We performed several exploratory analyses to locate zip codes and hospitals with the highest visit volumes, as well as to examine the differences in visit patterns and distances traveled across racial groups, age categories, and health conditions. Lastly, we applied a negative binomial-based spatial interaction model to explore how geographic, socio-economic, and health factors influence women’s maternal healthcare visit patterns. Our findings reveal that women do not always attend the nearest facility, and access patterns vary substantially by geography, income, and urban-rural classification. This study provides a data-driven foundation for identifying communities experiencing limited access to maternal care. These insights can inform more equitable planning and policy interventions in maternal healthcare infrastructure
Random Number Generators in Modeling and Simulation
No full textRandom number generation is essential to data science, operations research, cryptography, and especially modeling and simulation. Specifically, it is used in seeding experiments in operations research, generating safe keys in cryptography, and characterizing stochastic model behavior in simulations. This study investigates true random number generators (TRNGs), which create non-deterministic outputs based on physical processes like quantum tunneling. While pseudo-random number generators (PRNGs) like the Mersenne Twister (MT) and Linear Congruential Generator (LCG) are frequently used for their accessibility and speed, based on our research, TRNG's utilization in ABM remains unexplored. Our analysis compares PRNGs (MT, LCG) and TRNGs (SwiftRNG Z, SwiftRNG LE, TrueRNG 3, InfNoise) through both statistical testing and simulation-based experimentation. We first employed the NIST SP 800-22 statistical test suite of 15 tests (from frequency to linear complexity), applied on a thousand independent samples each with two million 64-bit binary numbers per RNG. Our best performing TRNG (SwiftZ) and PRNG (MT) showed consistent p-value uniformity (~0.48-0.52) with all tests passing the NIST threshold (≥965/1000), indicating high-quality randomness. The Non-Overlapping Template Test, which ran 148 tests per file, confirmed robust results with pass rates above 98.5 percent and an average p-value of 0.50. Subsequently, we created a Schelling Segregation Model on NetLogo with different parameters, and then built our own segregation model on Python which plugged in data from the TRNGs. Comparing the percent similar over time from both models using a variety of parameters, we found that both NetLogo and the model based on TRNG data produced the same graphs. Our findings from the observation of ABM and the NIST results emphasize the importance of true randomness in hardware generation for enhancing the realism and unpredictability of complex ABM
Effectiveness of Rule-Based Scripts for Automated Dependency Repair in Legacy Software
No full textResearchers often need to compile older projects to compare new techniques of theirs with techniques from prior work. However, one major challenge researchers face is that there are many issues when compiling older projects. The dependencies in older projects are often underspecified, and even when they are specified, specific versions of a dependency may no longer be available. To help with these problems, our work aims to provide solutions to help researchers automatically compile older projects. Rule-based scripts can help automate dependency repair in large legacy systems. Four rule-based scripts we created were compared on 19 legacy software projects that have compilation problems. The most successful script was to enable and disable plugins related to the compilation, resolving 52.6% of the problems. Of the remaining problems, changing to different versions of Java proved to be the best, resolving 50% of problems. The other two scripts were to remove the text SNAPSHOT and change the occurrence of “http” to “https” in the dependency declaration file. These changes effectively change a snapshot version of a dependency into a stable version of the same dependency and allows the compilation to run securely as newer dependency repositories block insecure http downloads. These scripts resulted in 33.3% of the remaining problems being resolved. Overall, the scripts repaired 89.5% of the compilation problems. These results demonstrate that rule-based scripts are a promising solution to repair a large proportion of dependency-related compilation problems in legacy software. Further research to explore integrating these scripts with intelligent systems could be beneficial to deal with more complex projects
Evaluating Confidential Computing for Research Replication: A Case Study Using Intel TDX
No full textReproducibility in empirical social-science research is essential, yet replicating code and data from academic journals often demands complex, unsecured computing environments that editors cannot easily audit. Confidential computing technologies such as the Intel Trust Domain Extensions (TDX) hold promise to embed the entire workflow in a verifiable enclave, but their practical cost, performance, and compatibility have not been systematically evaluated. To test whether TDX can underpin scale replication, Azure and Google Cloud TDX virtual machines were launched and attempted to reproduce 2025 Management Science articles with a public or provided replication package. Five papers for which complete materials were available were evaluated. Packages spanned Stata, R, SAS, MATLAB/Octave, and hybrid toolchains; some relied on WRDS or other proprietary data sources, and Virtual Machine billing logs alongside wall-clock runtimes were tracked and recorded within the VM. Three packages generated outputs identical to those reported (60%), one failed owing to mission data or irreconcilable code errors (20%), and one remains in extended execution (20%). This yields a replication rate of ~60% at a median cost of 0.30 USD, a median runtime of 186 minutes, and a mean runtime of ~167.4 minutes. While based on a limited sample size, these findings highlight that TDX virtual machines can securely replicate most multi-language research workflows at a negligible incremental cost, positioning confidential computing as a viable pathway for journals to enforce reproducibility without exposing sensitive code or data
Replication Failures in TDX-Confidential VM due to OS Incompatibility and Missing Dependencies
No full textReproducibility is essential to scientific research, but many published replication packages associated with research papers are undocumented, fragile, and even incompatible across various computing environments. This project focuses on evaluating whether confidential computing, specifically Intel’s Trusted Domain Extensions (TDX), can support secure and reliable replication of academic research papers. In this project, three recent Management Science studies were tested in a TDX-enabled confidential virtual machine on Microsoft Azure, but many technical barriers were faced. One package broke under a Waf-based build system with file handling errors. Another was blocked by a missing dataset (Hotels_MW.dta) with no author response. The final one relied on a Windows-only R package containing .dll binaries which made it unusable in a Linux-based confidential virtual machine. Despite immense efforts to troubleshoot these problems such as symbolic linking, environment configuration, and script rewrites, the replication packages could not be run for these papers, which highlights the necessity for platform-agnostic design, complete data sharing, and avoidance of OS-specific dependencies
An Assessment & Analysis of Current Cryptocurrency Exchanges and Their Regulation & Behavior in the Current Market Landscape
No full textCryptocurrency assets have risen in popularity extremely quickly, with nearly 4 trillion USD worth of money currently in the global cryptocurrency market (Forbes, 2025). In this extremely large, ever-growing market, however, there remains much to be desired with regard to the availability of information regarding cryptocurrency exchanges, including their overall size, volume traded, products offered, past regulatory incidents, and operational licenses/lack thereof. This paper attempts to accurately compile a lump-sum report on most, if not all, currently available cryptocurrency exchanges, as well as their parent corporations when applicable, for the purposes of accelerating due diligences for investors and individuals alike, so they may quickly regard whether or not they would like to place their money in the hands of an exchange
Comparing Learning Rate and Batch Size Transfer Across Different Network Widths for Standard and Maximal Update Parameterization
No full textIn deep learning, hyperparameters for models are commonly trained on smaller models and transferred to larger models to reduce the cost of learning. However, the scaling behavior of certain hyperparameters from smaller to larger network widths is not well understood. Maximal update parameterization is a new practice for deep learning models that aims to achieve hyperparameter stability across all model widths. This project compares the training dynamics for the optimal learning rate on neural networks trained using standard parameterization and maximal update parameterization on the MNIST dataset. The standard parameterization networks showed 0.001 as the optimal learning rate over 0.01 and 0.1 for network widths 128, 256, and 512 but the optimal batch size was shown to vary from network width to network width, which displays hyperparameter instability. 
Verifier-Guided Reinforcement Learning for GSM8K Math Reasoning
No full textThis project explores Reinforcement Learning from Verifier Rewards (RLVR) as a technique for improving multistep math reasoning in language models. However, despite advances in finetuning, existing instruction tuned LLMs still produce arithmetic errors in multistep reasoning tasks, lacking an inherent mechanism to verify and correct intermediate calculations. We apply RLVR to grade school word problems from the GSM8K dataset, using the flanT5base model for its instruction following capabilities. A Sympy based verifier checks each numeric prediction and issues a reward of 1.0 for exact matches and 0.0 otherwise. In a supervised finetuning baseline, our model achieved 3.07% exact match accuracy on a held out 10% test split. We then finetune the model with RLVR using the PPO algorithm in Hugging Face’s TRL framework, which raises exact match accuracy to 18%. These preliminary results show that verifier guided reinforcement learning can yield significant gains in LLM based math problem solving. Future work will investigate richer reward structures, model scaling, and additional verifier designs to further enhance reasoning performance
Species Distribution Modeling of Costa Rica’s National Bird Under Four Climate Futures
No full textTurdus grayi, the clay-colored thrush and the national bird of Costa Rica, is the most frequent species in the country’s eBird database, with more than 345,599 occurrence records. A robust Species Distribution Model (SDM) was built using this dataset to evaluate Turdus grayi’s present and future habitat suitability under climate change. Four algorithms for modeling: MaxEnt, Generalized Linear Models (GLM), Random Forest (RF), and XGBoost were used to make the prediction of species distributions, with MaxEnt giving the highest AUC of 0.84. The most influential environmental variables shaping the distribution of Turdus grayi were bio-7 (Annual Temperature Range) and bio-11 (Mean Temperature of Coldest Quarter). Future projections were made under four Shared Socioeconomic Pathways (SSPs) representing increasing levels of greenhouse gas emissions and climate change. Maps of environmental novelty were also prepared based on Multivariate Environmental Similarity Surfaces (MESS) to locate areas of extensive climatic change. Under all SSPs, the models indicate a gradual withdrawal and latitudinal shift in suitable habitats, with severe losses under high-emission futures (SSP 585). The results are important for long-term conservation planning and highlight the need for climate mitigation to preserve Costa Rica's avian biodiversit