PredictionErrorDriftDetector

Summary

PredictionErrorDriftDetector is RTTA's streaming implementation of: EWMA absolute prediction-error drift detector.

Update API

result = rtta.PredictionErrorDriftDetector().update(prediction, actual)

The update(...) call consumes one observation using prediction, actual. advance(...) uses the same inputs when the caller wants to update state without materializing a Python return value.

Theory Of Operation

PredictionErrorDriftDetector standardizes the current error or move against an EWMA mean and variance estimated from prior samples. The detector uses the resulting z-score with hysteresis or reset logic so isolated noisy observations do not become persistent regimes by themselves.

Recurrence

Let \(z_t = (prediction_t, actual_t)\) denote the observation consumed by one update(...) call and let \(\theta\) denote constructor parameters such as window lengths, thresholds, and smoothing constants.

\[e_t=|actual_t-prediction_t|, \qquad q_t=\frac{e_t-\mu_{t-1}}{\sqrt{\max(\sigma^2_{t-1},\epsilon)}}\]

\[\mu_t=\mu_{t-1}+\alpha(e_t-\mu_{t-1}), \qquad \sigma^2_t=(1-\alpha)(\sigma^2_{t-1}+\alpha(e_t-\mu_{t-1})^2)\]

\[r_t = \begin{cases} 1, & r_{t-1} = 0 \text{ and } q_t \ge e \\ 0, & r_{t-1} = 1 \text{ and } q_t \le x \\ r_{t-1}, & \text{otherwise} \end{cases}, \qquad x < e\]

The return value is the current scalar indicator value.

Implementation Notes

The recurrence is implemented in src/rtta/indicator.cpp in class PredictionErrorDriftDetector.

Reference

Background reference