Compression-Based Anomaly Detection for Business Intelligence: A Study of LZ78 and Neural Models in Time Series

Abstract

Anomaly detection in time series is an important problem in domains where abnormal behavior may indicate critical system failures and labeled data is rarely available, making unsupervised approaches particularly relevant. This paper relies on two compression-based methods for time series anomaly detection. The first approach implements universal probability assignment derived from the Lempel-Ziv 78 compression algorithm, while the second method employs neural lossless compression using an autoregressive recurrent model. Both techniques interpret anomalies as sequences that are difficult to compress under models learned from normal behavior. The methods are evaluated on the Electricity Transformer dataset, focusing on the temporal behavior and statistical properties of the anomaly scores, in order to highlight the complementary characteristics of universal and learned compression strategies. By eliminating the need for costly manual labeling, these unsupervised models provide a scalable and cost-efficient solution for real-time monitoring in resource-constrained industrial environments. This research demonstrates how information-theoretic metrics can be effectively integrated into Business Intelligence systems to facilitate informed decisions regarding predictive maintenance and risk mitigation.