Module temporal_ml

Differentiable temporal feature extraction.

Defines the TemporalOps trait for computing fixed-size feature vectors from variable-length embedding trajectories. The AnalyticBackend implementation uses closed-form computations (no ML framework needed).

Future backends (BurnBackend, TorchBackend) will implement the same trait with learnable parameters for end-to-end training.

Feature Vector Layout

The output is a fixed-size feature vector regardless of trajectory length:

Section	Size	Description
Mean velocity	D	Average velocity per dimension
Drift magnitude	1	L2 drift from first to last
Volatility	D	Per-dimension realized volatility
Soft change point count	1	Smoothed number of regime changes
Multi-scale drift	S	Drift at S temporal scales

Total: 2*D + 2 + S features.

Structs

AnalyticBackend

Pure-Rust analytic feature extractor (non-differentiable).

AnalyticConfig

Configuration for the analytic backend.

Traits

TemporalOps

Temporal feature extraction operations.

Functions

compute_scale_drift 🔒

Average drift magnitude at a given temporal scale.

soft_change_point_count 🔒

Soft change point count: number of points where the local deviation exceeds threshold standard deviations, smoothed by a running window.