About me.

Data in the real world is distributed through time. Time series modelling isn’t a new idea, dating back to the era of Galton (1880s), but still remains a significant challenge in deep learning. I enjoy working on unique and impactful problems within time series, where traditional deep learning paradigms fail. For example, many time series within healthcare (e.g., ECG, blood glucose measurements, sleep studies) are variable in length, yet, most of the current literature focuses on finite-context methods, which can only process sequences of fixed length. This need for robust variable-length time series classification (VSTC), is present not only in healthcare, but in many other real-world scenarios where there is no predetermined sequence length (e.g., finance, weather). We recently released a paper addressing this, called Stochastic Sparse Sampling: A Framework for Variable-Length Medical Time Series Classification—for more info on this, see my projects page.

While traditional deep learning methods provide point estimates—such as a single value or vector—they lack measures of uncertainty, limiting their usefulness in high-risk domains such as healthcare or finance. For example, incorporating uncertainty measures can be highly advantageous in clinical contexts, as it not only enhances patient outcomes but also improves clinician trust and promotes the adoption of models in the clinic. Currently, I am focusing on developing conformal prediction (CP) methods for time series forecasting of electromagnetic radiation to reduce radiation exposure. Conformal prediction is advantageous over other uncertainty quantification methods, as it provides a guarantee that prediction intervals contain the true outcome at a specified confidence level, regardless of data distribution or base model choice.

Currently I'm working on OpTrade, a framework for high-frequency forecasting of alpha term structures in American options markets using specialized deep learning architectures. My project aims to analyze market microstructure dynamics across various options contracts and eventually translate these insights into actionable trading signals. I'm particularly excited about my approach to position sizing based on model uncertainty using conformal prediction, which views alpha through the lens of uncertainty rather than traditional excess returns. This perspective allows me to capture more complex risk scenarios that standard volatility metrics might miss, potentially enhancing trading strategy performance in these sophisticated markets.

The rapid rise of AI in healthcare promises significant advancements, but there's a growing need for methods specifically designed to address clinicians' practical challenges and workflow. Currently, I am working on time series classification in seizure onset zone (SOZ) localization—the task of identifying seizure-inducing regions within the brain for epilepsy patients. Traditionally, neurosurgeons and epileptologists visually inspect intracranial electroencephalography (iEEG) signals to determine the SOZ. However, this challenge isn't so easy to apply any off-the-shelf machine learning algorithm. The signals can be extremely long (due to high sampling rate), variable in their length (for different recordings or patients), and the number of channels can vary among different iEEG setups. Furthermore, there are currently no clinically validated markers of the SOZ, and therefore there is a need of understanding the link between signal characteristics and pathology.
We just released a preliminary version of our work surrounding this at the NeurIPS 2024 Workshop on Time Series in the Age of Large Models, with the full version on arXiv and under review. Our approach, called Stochastic Sparse Sampling (SSS), can handle not only variable-length and long sequences, a varying number of channels, but also provides post-hoc model interpretability with respect to local signal regions, which are directly tied to the probability of the SOZ. For more information on this, see my projects page.