Natanael S. Alpay

Mathematics · Quantitative Research · Machine Learning

Machine Learning and Scientific Computing

I combine mathematical analysis with computational experimentation to develop and evaluate methods in machine learning, statistical modeling, and scientific computing. My projects include drug and vaccine safety signal prioritization, kernel methods, mathematical visualization, and neural hedging with efficient simulation.

I use AI and large language models as conjecture engines to generate candidate formulas, explore algorithms, and develop possible proof strategies. I critically evaluate these ideas through counterexample searches and reproducible computational experiments, and establish mathematical results through rigorous proof. My strength is connecting theory with implementation: understanding how a method works, evaluating its performance, and identifying the assumptions and limitations that matter in practice.

Research and Machine Learning

Memory-Efficient Deep Hedging

Work in progress · Joint work with Emeric Battaglia

We implement neural hedging strategies in Python and PyTorch using Heston market simulations, transaction costs, and CVaR and entropic risk objectives. We study how training data can be generated and used more efficiently while preserving hedging performance.

Our matched three-stock experiment reduced peak training RAM by 63.8%, with closely comparable observed test CVaR. In a separate short benchmark, parallel generation reduced dynamic preparation, training and testing time by a factor of 2.47. These preliminary results motivate further experiments with larger portfolios and longer trading horizons.

A longer-term direction is to study simpler portfolio representations that preserve hedging quality.

Python · PyTorch · Monte Carlo simulation · Risk optimization

References: GitHub repository and notebooks · Project write-up

FAERS: Machine Learning for Drug Safety Signal Prioritization

I developed and evaluated gradient-boosted models to prioritize drug–adverse-event reporting signals in the FDA Adverse Event Reporting System (FAERS). The study combines classical disproportionality measures with report-quality information and examines performance on future quarters, calibration, and the contribution of different feature groups.

2026 · Gradient boosting · Temporal validation · Calibration

VAERS: Early Prioritization of Vaccine–Symptom Signals

Using the Vaccine Adverse Event Reporting System (VAERS), I built a workflow that ranks vaccine–symptom pairs by their likelihood of meeting a reporting-signal criterion in the following quarter. The project compares logistic regression and calibrated gradient boosting with classical statistical scores to prioritize pairs for expert review.

The target is a future statistical reporting signal used for triage, rather than a causal estimate of vaccine risk.

2025–2026 · Logistic regression · Gradient boosting · Forecasting

Complex Stochastic Gradient Descent and Kernel Regression

In joint work with Emeric Battaglia, I study stochastic gradient descent for complex-valued kernel regression. The project combines convergence and directional-bias analysis with numerical experiments in complex reproducing kernel Hilbert spaces, including the recovery of superoscillatory functions and Blaschke products.

2026 · Stochastic optimization · Kernel methods · Numerical analysis

Machine Learning: Theory and Implementation

As an undergraduate researcher working with Mihaela Vajiac, I studied and implemented classification methods, including perceptrons, radial basis function networks, and support vector machines, using Python and MATLAB. The project connects the mathematical foundations of these methods with numerical experiments and visualizations of decision boundaries.

2021–2022 · Python · MATLAB · Classification · Kernel methods

Scientific Computing and Data Analysis

Mathematical Structures in 3D and Virtual Reality

Working with Peter Jipsen, I developed JavaScript visualizations of finite lattices using Babylon.js. The project renders algebraic structures as interactive 3D objects that can be explored in a web browser or with a virtual reality headset.

2020–2022 · JavaScript · Babylon.js · Mathematical visualization

Exploratory Analysis of New York City Traffic Collisions

I analyzed New York City motor vehicle collision records, using data cleaning and visualization to examine recorded crash characteristics and changes around the COVID-19 pandemic. The project focuses on descriptive comparisons and patterns in the reported data.

Data cleaning · Exploratory analysis · Visualization

Additional Programming Projects

Financial Time-Series Modeling

I applied ARIMA and Holt–Winters models in R to financial time series. The project focused on fitting statistical models, examining trends, and producing and visualizing forecasts.

R · ARIMA · Holt–Winters

Interactive Periodic Table Database

I built a Python and SQL web interface for exploring data about the 118 chemical elements. The project connects a relational database with an interface for querying and displaying element properties and their relationships.

Python · SQL · Web development

Binary Search Trees and Red-Black Trees

I implemented binary search trees and red-black trees in C++, studying how tree balancing supports efficient search and updates.

C++ · Data structures · Algorithms

Technical Skills

Programming: Python, MATLAB, R, SQL, C/C++, Java, JavaScript, and Mathematica.

Data analysis and visualization: NumPy, pandas, Matplotlib, Tableau, and Excel.

Methods: Statistical learning, kernel methods, optimization, time-series modeling, and numerical experimentation.

Research Experience and Mentoring

Chapman Brain Institute: Research support involving data collection, cleaning, statistical analysis, and visualization.

DATA@UCI: Mentored an undergraduate research team in data analysis, dimensionality reduction, research design, and presentation.