Causal discovery in microbiome data with identifiable variational autoencoders
Evaluating whether questionnaire-guided iVAE representations support causal discovery in microbiome data with hidden confounding.
Discovering financial factors with language model agents
This project compares language model agents, random search and genetic search for financial factor discovery, isolating the contribution of experimental memory under a common evaluation budget.
Fraud Campaign Detection and Rule Learning for Card Payments
Investigate whether transaction groups and interpretable rules, including rules proposed by language models, improve future fraud detection beyond a LightGBM baseline.
JEPA for Code Generation and Optimization
Investigating whether predictive world models in representation space can guide code generation and compiler optimization.
Relational steering for tabular anomaly detection
Investigating whether context-dependent interventions in tabular models reveal anomalies that violate relationships between variables.
Flow Matching for Credit Card Fraud Detection
Evaluate one-step flow matching for credit card fraud detection, comparing classification quality, training cost and inference speed with gradient boosted trees.
Large language models and news sentiment for algorithmic trading
This project investigates whether financial news sentiment extracted by large language models can improve trading signals, with an emphasis on Apple stock, news timing and rigorous backtesting.
Microbiome analysis, knowledge graphs and causality
Combining literature-derived knowledge graphs with microbiome observations to investigate causal hypotheses in personalized health.
Uncertainty quantification for large language models
This MMF project studies uncertainty estimates for transformer models, initially on tabular classification, with potential applications to the reliability of trading predictions.
Decision trees with complex conditions for trading
Design and compare decision trees with multiple-variable conditions for interpretable prediction on trading or fraud data.