PARTIAL · risk gate
Hypothesis
An autoencoder trained on "normal" market states flags anomalous regimes (high reconstruction error) to step aside before disorderly moves.
Math
$$ \text{anomaly}_t = \big\| \mathbf{x}_t - \text{Dec}(\text{Enc}(\mathbf{x}_t)) \big\|^2 $$
Method
Train on calm-regime feature vectors, monitor reconstruction error live, de-risk on spikes.
Results
| Reconstruction error spikes pre-vol | often |
| Lead time | short / coincident |
| As a risk gate | mildly useful |
A mild risk-gate: reconstruction error does rise around regime breaks, but mostly coincident with volatility, so it confirms more than it predicts. Kept as one input to a de-risking ensemble, not a trigger.
Anomaly detectors mostly tell you the present is unusual, not that the future is. Useful for stepping aside, weak for stepping in.