What AKF is (and what it is not)

An Adaptive Kalman Filter estimates an unobserved state (e.g., fair price / drift / micro-volatility) from noisy observations (ticks, OHLC, bid/ask). It becomes adaptive when it updates its noise assumptions (process noise Q, measurement noise R) as conditions change.

Reality check: AKF does not “forecast the market”. It reduces noise, provides an uncertainty-aware baseline, and makes downstream policies less reactive.

Core model (minimal)

State:        xₖ = F xₖ₋₁ + wₖ      (w ~ N(0, Q))
Measurement:  yₖ = H xₖ   + vₖ      (v ~ N(0, R))
Innovation:   eₖ = yₖ - H x̂ₖ|ₖ₋₁
  

Diagram: where AKF sits in a risk pipeline

The filter is a stabilizer. The gate is the enforcer. The dashboard is the observer.

Table: common failure modes and mitigations

Failure mode	Symptom	Mitigation (practical)
Wrong noise assumptions (Q/R)	lagging or overreacting estimate	adaptive update rules + regime bins + bounded updates
Bad input quality	unstable state, false confidence	schema validation, timestamp sanity, spread/latency monitors
Non-stationary shocks	innovation spikes, divergence	shock detector + temporary R inflation + gate cooldown

Requirements (non-negotiable)

Log raw observations and the filter state (x̂, P) per step.
Version your configuration (hash + effective_from) and keep rollback.
Bind the filter output to a gate policy — never to “prediction selling”.

Modules

Catalog