Adaptive Thresholding for Non-Stationary Anomaly Detection

This direction is merged into another direction: adaptive anomaly thresholding. Its last version stays readable and citable here.

Version 1 of 1437 papers508 conceptsfrozen 2026-10-04

In traditional anomaly detection, the boundary between "normal" and "anomalous" is often a fixed line. A sensor reading above a certain value triggers an alarm; a network packet larger than a specific size is flagged. However, real-world systems—from spacecraft and power grids to connected cars—are rarely static. They age, they respond to changing weather, and they shift through different operational phases. In these non-stationary environments, a static threshold is a liability: set it too low, and the system is overwhelmed by false positives; set it too high, and it misses stealthy, critical failures.

The research direction emerging across these 257 papers focuses on Adaptive Thresholding. Instead of a hard-coded limit, these systems treat the detection boundary as a dynamic variable that evolves alongside the data it monitors.

The Shift from Raw Values to Residuals

The core intuition behind modern adaptive detection is that we should not threshold the raw data itself, but rather the prediction error (or residual) of a model.

If we have a model that understands the "physics" or typical patterns of a system, it can predict what the next data point should be. An anomaly is then defined as a significant deviation from this prediction. As Dongyu Liu and colleagues demonstrate in the AQEyes system, calculating thresholds based on the distribution of these errors allows a model to function across regions with vastly different baseline pollution levels without manual recalibration.

Similarly, researchers at NASA’s Jet Propulsion Laboratory, including Kyle Hundman, utilized LSTMs with nonparametric dynamic thresholding to monitor spacecraft. By using a sliding window to calculate thresholds from recent prediction errors, the system naturally adapts as the spacecraft ages or enters new mission phases, ensuring that "normal" signal drift isn't mistaken for a catastrophic failure.

Statistical Self-Calibration

How does a system decide a threshold is "correct" without human intervention? The papers suggest several nonparametric and statistical mechanisms:

  1. Distribution-Free Metrics: Yingbin Liang and H. Vincent Poor propose using Maximum Mean Discrepancy (MMD) to compare data streams. This allows for "order-level optimal" detection without knowing the underlying probability distributions in advance.
  2. Relative Outliers: In medical contexts, such as identifying salient brain regions in EEG data, Christian Rummel and others use the inter-quartile range (IQR) at each time step. This labels data points as outliers relative to the current network state, rather than an arbitrary global constant.
  3. Bayesian Bootstrapping: For complex "bags of data" like email networks, Hideitsu Hino and Noboru Murata use non-overlapping confidence intervals to trigger alerts only when the statistical difference between a reference and a test window becomes significant.

Context-Aware Sensitivity

A major breakthrough in this field is the move toward contextual adaptation. A car driving at 70 mph on a highway has a different "normal" sensor profile than one idling in a parking lot.

Matan Levi and Aryeh Kontorovich addressed this by combining Hidden Markov Models with regression-based thresholds. The regression model predicts the expected likelihood of a behavior sequence given the current driving context, effectively "tuning" the sensitivity of the security system in real-time.

In aerospace, Brian M. de Silva and Jared Callaham took this further by using supervised classifiers to automate thresholding. By training a decision tree on prediction residuals, the system learns to distinguish between high-variance maneuvers (which are normal) and actual sensor degradation.

Control Loops for Threshold Stability

Finally, some researchers treat threshold adjustment as a control problem. Jakub Marecek and Tigran Tchrakian introduced the ADMI (Additive-Decrease Multiplicative-Increase) logic. When an observation exceeds the threshold, the system flags an anomaly and aggressively increases the threshold (multiplicative increase) to account for a potential shift in environmental noise. When errors are low, the threshold gradually tightens (additive decrease).

This "elastic" boundary ensures the system remains robust against transient noise while maintaining high sensitivity to persistent structural changes, a theme echoed in Ruimin Chu's work on online fuel leakage detection, where historical data diversity dictates the current alarm level.

By moving from static limits to these self-correcting, context-aware mechanisms, anomaly detection becomes a living process capable of navigating the entropy of the real world.

Papers in this version

Showing 60 of 437 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adaptive_thresholding_for_non_stationary_57e23673,
  title = {Adaptive Thresholding for Non-Stationary Anomaly Detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-thresholding-for-non-stationary-anomaly-detection/rdr_57e2367362b84ca6805bed121f9986f5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_adaptive_thresholding_for_non_stationary_57e23673_v1,
  title = {Adaptive Thresholding for Non-Stationary Anomaly Detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-thresholding-for-non-stationary-anomaly-detection/rdr_57e2367362b84ca6805bed121f9986f5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}