​
Historical Anomaly Detection

Historical anomaly detection identifies data points that deviate from the expected behavior in a given time series. It helps you pinpoint potential fraudulent activities, security breaches, or other significant outliers.

In this section, you will learn how to perform historical anomaly detection, including:

  • Generating predictions
  • Constructing a 99% confidence interval
  • Identifying data points that fall outside this interval as anomalies

Below are links to more specific use cases and tutorials:

Historical anomaly detection

Quickstart guide to performing basic anomaly detection.

With exogenous features

Learn how to incorporate external data (like weather or events) in your anomaly detection.

With date features

Enhance accuracy by including seasonal or holiday features tied to specific calendar dates.

Modifying confidence intervals

Adjust the confidence interval to suit your detection threshold requirements.

​
How It Works

By flagging anomalies quickly, you can prioritize investigation and reduce the risk of unnoticed irregularities.

​
Step-by-Step Guide

1

Collect and Prepare Data

Ensure your historical time series is clean and well-formatted. Handle any missing values or outliers before sending data to TimeGPT.

2

Choose Confidence Level

Decide on a confidence interval (e.g., 99%) suitable for your risk tolerance. A higher interval flags fewer anomalies but reduces false positives.

3

Configure and Run Detection

Use TimeGPT’s APIs or libraries with your preferred confidence level to generate predictions and intervals.

Historical Anomaly Detection with TimeGPT
import timegpt

# Initialize with your API key
t = timegpt.TimeGPT(api_key="YOUR_API_KEY")

# Prepare your time series data
data = [...]

# Run historical anomaly detection
anomalies = t.detect_anomalies(
    data=data,
    confidence_interval=0.99
)

print(anomalies)
4

Review Anomalies

Inspect the anomalies flagged by the model. These points are potential indicators of significant deviations in your data.

5

Adjust and Iterate

If you find that the model is overly sensitive or missing critical outliers, adjust the confidence interval or include additional features (e.g., exogenous data, date features) to improve detection accuracy.

Always validate anomalies. Some flagged data points may be normal operational fluctuations rather than true anomalies.

Use charts and graphs to visualize how your data behaves around the confidence interval. This helps you quickly spot anomalies.

​
Summary

Historical anomaly detection helps you proactively manage risks by pinpointing unusual data points in your past records. Whether you want to detect fraudulent transactions or system malfunctions, TimeGPT equips you with flexible confidence interval settings and the option to incorporate additional features for more refined results.

Feel free to explore our specialized guides for exogenous or date-based features, or learn how to tweak TimeGPT for different confidence levels.

Ready to get started? Check out our quickstart guide from the links above to start detecting anomalies in your historical time series today.