Abstract
Outlier detection in non-stationary time series is challenging due to regime changes and strong contextual dependence of system behavior. Observations numerically typical in a global sense may violate structural relationships that hold only within specific operating regimes.
This paper proposes a context-aware rule-based framework for outlier detection in complex time series. System behavior is partitioned into deterministic contexts representing volatility and trend regimes; for each context, a set of interpretable IF--THEN rules describes regular structural relationships. Outliers are identified as observations that violate rules of the active context or exhibit inter-context inconsistency---strong disagreement between active and alternative context rule sets.
The method is fully deterministic, requires no iterative training or probabilistic density estimation, and preserves interpretability at both rule and context levels. Rule ranges are calibrated via non-parametric percentile estimation with linear O(N) complexity. Evaluation on three daily currency pairs (EUR/USD, GBP/USD, USD/JPY, 2010--2022) under leave-one-year-out cross-validation and on a synthetic nonlinear system confirms that context conditioning increases sensitivity to structural regime transitions while maintaining stability compared to global or fully local approaches.
Paper Type
Short Paper
DOI
10.62036/ISD.2026.53
Context-Aware Rule-Based Outlier Detection in Non-Stationary Time Series
Outlier detection in non-stationary time series is challenging due to regime changes and strong contextual dependence of system behavior. Observations numerically typical in a global sense may violate structural relationships that hold only within specific operating regimes.
This paper proposes a context-aware rule-based framework for outlier detection in complex time series. System behavior is partitioned into deterministic contexts representing volatility and trend regimes; for each context, a set of interpretable IF--THEN rules describes regular structural relationships. Outliers are identified as observations that violate rules of the active context or exhibit inter-context inconsistency---strong disagreement between active and alternative context rule sets.
The method is fully deterministic, requires no iterative training or probabilistic density estimation, and preserves interpretability at both rule and context levels. Rule ranges are calibrated via non-parametric percentile estimation with linear O(N) complexity. Evaluation on three daily currency pairs (EUR/USD, GBP/USD, USD/JPY, 2010--2022) under leave-one-year-out cross-validation and on a synthetic nonlinear system confirms that context conditioning increases sensitivity to structural regime transitions while maintaining stability compared to global or fully local approaches.
Recommended Citation
Kacprowicz, M. & Niewiadomski, A.(2026). Context-Aware Rule-Based Outlier Detection in Non-Stationary Time Series. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.53