Researching complex systems through data.
Three independent research projects spanning quantitative finance, multimodal artificial intelligence and urban mobility. Across each project, the focus remains consistent: building interpretable models, controlling experimental conditions and asking what the data can actually support.
Inverse Gaussian Mixture Recovery and Tail Risk under Regime Shifts
A structural break creates a difficult statistical problem: historical observations are plentiful but may represent an obsolete regime, while observations from the new regime are economically relevant but sparse. This study reconstructs the event-regime distribution using an inverse Gaussian-mixture framework and evaluates the resulting Value at Risk and Expected Shortfall.
| Event | Event N | Mean | Volatility |
|---|---|---|---|
| S&P 500 · GFC | 85 | −0.413% | 4.022% |
| S&P 500 · COVID-19 | 62 | −0.165% | 3.883% |
| NVIDIA · ChatGPT | 84 | +0.684% | 3.491% |
| NVIDIA · DeepSeek | 60 | −0.357% | 4.292% |
| CSI 1000 · Stimulus | 44 | +0.740% | 3.175% |
Reduction in reconstruction divergence for the inverse estimator relative to the direct event-window benchmark.
Estimated one-day loss in the extreme tail as a percentage of position value.
The results show why average return alone is insufficient for interpreting a regime shift. A strongly positive event mean can coexist with materially larger reversal risk, while technology events can rotate exposure between the long and short tails rather than simply widening the entire distribution.
Controlled Case Studies of Three Add-on Families in Frozen-CLIP Sarcasm Detection
Multimodal systems often change architecture, features, objectives and external information simultaneously. This study instead freezes the OpenCLIP environment and tests three intervention families independently: shared/private decomposition, generated external context and robustness-oriented training.
Mean performance across seeds 42–51.
Paired macro-F1 differences under the fixed experimental environment.
| Condition | Baseline B4 | Robust R1 | Effect |
|---|---|---|---|
| Image Missing | 0.77625 | 0.79327 | +0.01702 |
| Text Missing | 0.76585 | 0.77766 | +0.01181 |
The controlled comparisons suggest that architectural decomposition can provide a small improvement under fixed features, generated context is not automatically useful, and robustness training mainly becomes valuable when modalities are actually missing. The results highlight the importance of isolating individual design choices instead of attributing the performance of an entire multimodal system to a single module.
UAV-assisted Cyclist Risk Assessment Based on Visual Perception and Rule-based Risk Modeling
Aerial imagery provides a view of traffic interactions that is difficult to obtain from ground-level cameras. This project combines YOLO-based object detection with an interpretable rule-based risk function to identify how exposed each detected rider is to nearby road users.
| Metric / Class | Recall | AP50 | Interpretation |
|---|---|---|---|
| Overall | 0.581 | 0.628 | 5 classes |
| Bicycle | 0.308 | 0.329 | Primary limitation |
| Motor | 0.632 | 0.642 | Stronger rider detection |
| Small Vehicle | 0.858 | 0.909 | Strongest major class |
Validation-stage rider scoring and comparison statistics.
How stable the exposure ranking remains under alternative implementations.
The rule-based risk model produces a stable ordering of rider exposure and remains interpretable because every risk score can be traced back to surrounding road users. The dominant limitation is perception rather than risk aggregation: pedal cyclists are small and comparatively rare in the training data, causing many riders to be missed before the risk model can evaluate them.
Different systems.
One analytical mindset.
From financial markets to multimodal learning and aerial traffic perception, these projects explore the same broader question: how can data be transformed into evidence without losing interpretability, context or uncertainty?