[Paper Review] Anomaly Detection in Univariate Time-series: A Survey on the State-of-the-Art

Mohammad Braei, Sebastian Wagner|arXiv (Cornell University)|Apr 1, 2020

Anomaly Detection Techniques and Applications90 references154 citations

TL;DR

This survey compares statistical, classical ML, and deep learning approaches for anomaly detection in univariate time-series, evaluating 20 methods across public datasets in terms of accuracy and runtime.

ABSTRACT

Anomaly detection for time-series data has been an important research field for a long time. Seminal work on anomaly detection methods has been focussing on statistical approaches. In recent years an increasing number of machine learning algorithms have been developed to detect anomalies on time-series. Subsequently, researchers tried to improve these techniques using (deep) neural networks. In the light of the increasing number of anomaly detection methods, the body of research lacks a broad comparative evaluation of statistical, machine learning and deep learning methods. This paper studies 20 univariate anomaly detection methods from the all three categories. The evaluation is conducted on publicly available datasets, which serve as benchmarks for time-series anomaly detection. By analyzing the accuracy of each method as well as the computation time of the algorithms, we provide a thorough insight about the performance of these anomaly detection approaches, alongside some general notion of which method is suited for a certain type of data.

Motivation & Objective

• Motivate the need for broad, cross-category evaluation of time-series anomaly detection methods.
• Compare statistical, classical machine learning, and neural network approaches on univariate time-series.
• Provide insights on method suitability for different data characteristics and performance trade-offs.

Proposed method

• Discuss foundational concepts: anomalies, time-series, types of anomalies, and evaluation framing.
• Introduce and categorize anomaly detection methods into statistical, classical ML, and neural-network-based approaches.
• Describe evaluation setup, including dataset usage and hyperparameters, to ensure fair comparisons.
• Summarize representative methods within each category (e.g., AR/ARMA/ARIMA, SES/TES/DES, clustering and One-Class SVM, autoencoders, GANs).
• Analyze how prediction-based versus shape-based detection applies to univariate time-series.

Experimental results

Research questions

• RQ1 What are the relative accuracies of statistical, classical ML, and deep learning methods for univariate time-series anomaly detection?
• RQ2 How do computation times compare across method families on benchmark datasets?
• RQ3 Which data characteristics (stationarity, seasonality, windowing) favor particular anomaly detection approaches?
• RQ4 What are the practical considerations (supervised vs. semi-supervised vs. unsupervised) in applying these methods to univariate time-series?

Key findings

• The paper provides a quantitative comparison across three method categories on multiple datasets.
• It analyzes both accuracy and computation time to map performance trade-offs.
• It discusses the suitability of methods for different data properties and anomaly types (point, collective, contextual).
• It highlights the dependency of performance on model choices (e.g., AR vs. ARMA vs. ARIMA, SES/DES/TES, clustering vs. density-based methods).
• It emphasizes that there is no single best approach; effectiveness is data-dependent and linked to forecasting versus shape-based detection strategies.