open-phase-ensemble

An open-source, heterogeneous multi-tool ensemble for streaming time-series anomaly detection and forecasting.

Preliminary Research Self-Reported Metrics

Time-series anomaly detection suffers from methodological fragmentation and evaluation pitfalls. Many academic benchmarks rely on flawed point-adjustment protocols that artificially inflate random guessers to >90% accuracy.

open-phase-ensemble is a heterogeneous ensemble that maps scalar temporal observations into phase-space manifolds via Takens' delay embedding, combined with a battery of six expert detectors — spanning non-parametric phase-space methods, covariance geometry, and parametric models including a transformer autoencoder and isolation forest — governed by an online multiplicative-weights Meta-Judge. On the TSB-AD-U industry benchmark, the system does not yet outperform simpler baselines on the primary metric (VUS-PR); see the TSB-AD-U results section for details.

0.0303 MIT-BIH VUS-PR (primary) — VUS-ROC 0.9354 (+0.4324 over null)

0.5223 CWRU VUS-PR (primary) — VUS-ROC 0.6434 (+0.3087 over null)

0.0000 Stream vs Batch Distance (Zero-Lookahead)

System Foundation & Architecture

The system decomposes time-series evaluation into six modular, causal compartments:

Heterogeneous Ensemble Foundation

Combines non-parametric phase-space methods (Simplex, Mahalanobis, Matrix Profile) with parametric models (Isolation Forest, AR ridge filter, Transformer autoencoder) under a unified online fusion framework.

Orthogonal 6-Detector Battery

Combines phase-space trajectory prediction, covariance geometry, subsequence motifs, subspace isolation, linear autoregression, and neural MSE reconstruction error. Pairwise Pearson correlation between detector score streams is low (all pairs $|r| < 0.3$, max $|r| = 0.14$); see the correlation matrix.

Online Meta-Judge

Employs the Hedge multiplicative-weights algorithm with fixed-share mixing floors to dynamically reweight experts in real time without ground-truth labels.

Zero-Lookahead Invariant

Enforces strict element-by-element streaming execution. No future data points are ever exposed during online scoring.

Methodological Rigor

Evaluated using un-buffered Volume Under Surface (VUS-ROC/PR) metrics with label-only buffering and IAAFT phase-randomized surrogate null baselines.

Research Background

open-phase-ensemble originated from an audit of a closed-source time-series verification engine that mapped raw observations into multidimensional phase-space manifolds via Takens' delay embeddings to predict future dynamics through geometric trajectory matching.

Auditing the mathematical bounds of pure phase-space trajectory matching exposed a structural limitation: trajectory projection captures non-linear dynamic shifts well, but struggles with spatial covariance shifts, variance drops, and isolated global point outliers. No single detection algorithm is universally optimal across all anomaly types; robustness requires an ensemble of mathematically orthogonal experts.

To address this, the system combines a battery of six complementary algorithms:

Simplex Projection (EDM) — Phase-space trajectory geometry & non-linear prediction
Robust Mahalanobis (Ledoit-Wolf) — Covariance geometry & variance shift detection
Matrix Profile (Single-Window) — Subsequence motif & discord discovery
Isolation Forest — Subspace partitioning & global point anomalies
AR Linear Ridge Filter — Linear stochastic prediction residuals
MSE Transformer Autoencoder — Neural sequence reconstruction error

Choosing to Build Openly

This research is developed and released entirely in the open for three primary reasons:

Auditable Reproducibility: Closed-source benchmarks cannot be independently audited. Open code guarantees that every score, calculation, and claim can be verified.
Metric Integrity: Volume Under Surface (VUS-ROC/PR) with label-only buffering eliminates point-adjustment inflation.
Community Collaboration: A clean, modular design allowing researchers to plug in new detectors, surrogate generators, or gating strategies.

Significance & Scientific Outlook

The observations documented in this repository are offered as directions for further investigation, not as established conclusions. The results suggest that an open, composable, online-learning ensemble can achieve meaningful predictive edges over phase-randomized null baselines while remaining fully inspectable and deterministic.

Key open research questions include:

Ensemble Scalability: How does the Hedge Meta-Judge scale as the detector battery expands to 20+ experts?
High-Noise Limits: What is the breakdown threshold of non-parametric delay embeddings when signal-to-noise ratio drops below 0 dB?
Cross-Domain Generalization: How well does the architecture maintain its predictive edge across non-stationary geophysical, financial, and industrial streams?