Streaming wavelet transforms for online systems.

FerroWave is built for cases where feature extraction has to happen as data arrives, not after an offline batch job finishes.

Why this needs production signal processing

!

Batch transforms are not enough when inference or monitoring has to update on each tick or event.

!

Rolling windows are easy to code but weak at scale-localized signal behavior.

!

Online systems need predictable allocation and latency behavior.

How FerroWave handles it

01

Maintain state

Streaming transform state updates as observations arrive instead of recomputing the full window.

02

Emit features

Online features can represent scale energy, detail coefficients, denoised values, or transform summaries.

03

Feed inference

The emitted features can drive monitoring, classification, regime detection, or trading-system context.

What makes this FerroWave-shaped

The public site calls out nanosecond-class streaming MODWT/tick paths.

Streaming features are named as one of the ML workflow shapes.

Rust implementation supports deployment-sensitive feature extraction.

Common concerns

Is streaming support only for market data?

No. Market data is a natural use case, but streaming wavelet features also apply to other time-series systems.

Does streaming replace batch transforms?

No. It complements batch workflows when latency or online updates matter.