Research
Open questions, studied carefully.
We investigate how beliefs should be formed and revised as evidence arrives, how models should adapt when the process behind their data changes, and how to act well when the answer remains uncertain. Each area below begins with a question we cannot yet answer to our own satisfaction.
Questions we study
-
Sequential and causal inference
How should beliefs be revised as evidence arrives in order — and which relationships survive intervention?
Filtering, smoothing and change-point detection for data that arrives one observation at a time, together with the harder question of separating predictive association from causal structure in observational data.
- Online Bayesian inference
- Filtering and smoothing
- Change-point detection
- Causal discovery
- Counterfactual evaluation
-
Continual learning
How can a model keep learning from a changing environment without forgetting what still holds?
Non-stationary data, drift detection and the trade-off between adapting quickly and preserving knowledge that remains valid. When to update, how much, and how to tell a regime change from noise.
- Concept and covariate drift
- Catastrophic forgetting
- Online adaptation
- Regime-aware learning
-
Model reliability
When should a model's output be trusted, and how do we know when it should not be?
Calibration, uncertainty quantification, anomaly and out-of-distribution detection, and evaluation protocols designed to resist leakage, selection effects and overfitting to historical tests.
- Calibration
- Conformal prediction
- Anomaly detection
- Out-of-distribution detection
- Evaluation methodology
-
Decisions under uncertainty
How should a system act when outcomes are uncertain and its own actions change the environment it observes?
Sequential decision processes, reinforcement learning and risk-aware objectives — including execution treated as a decision problem shaped by market structure rather than as an afterthought to prediction.
- Reinforcement learning
- Risk-sensitive control
- Market structure
- Execution as decision-making
How results are judged
- The question comes first
- The hypothesis and the test that could refute it are fixed before the data is examined.
- Honest baselines
- Every method is compared against simple alternatives that are difficult to beat.
- Out-of-sample evaluation
- Held-out and walk-forward testing, with deliberate care to prevent information leaking from the future.
- Negative results are kept
- What did not work is recorded, because it constrains what might.
Topic index
Topics that recur across the work, and the question each is studied under.
| Topic | Research area |
|---|---|
| Adaptive systems | Continual learning |
| Anomaly detection | Model reliability |
| Calibrated decision systems | Model reliability |
| Causal reasoning | Sequential and causal inference |
| Continual learning | Continual learning |
| Intelligent execution systems | Decisions under uncertainty |
| Market structure | Decisions under uncertainty |
| Model reliability | Model reliability |
| Probabilistic reasoning | Sequential and causal inference |
| Quantitative research | Decisions under uncertainty |
| Reinforcement learning | Decisions under uncertainty |
| Sequential inference | Sequential and causal inference |
| Time-series modeling | Continual learning |