Anomaly Detection: Understanding Our Anomaly Detection Model

Our anomaly detection model is designed to help identify unusual patterns in your data, which could indicate potential issues or significant changes. It works by looking at your data over the past few months to understand the normal patterns and trends that typically occur.

How it Works:

1. Learning from Past Data:
The model starts by analysing your historical data, focusing on the general trend over several months. This helps it understand what "normal" looks like for your metrics.

2. Focusing on Recent Trends:
While it takes a long-term view, the model also emphasizes recent data, especially the past few weeks. This is important because some patterns, like seasonal changes (for example, weekly or monthly fluctuations), can have a significant impact on your metrics. By focusing on recent data, the model is better at capturing these short-term seasonal trends.

3. Detecting Anomalies:
Once the model understands the typical trends and seasonal patterns, it forecasts expected values for your metrics going forward. Each forecasted value comes with an upper threshold (the highest value we expect to see) and a lower threshold (the lowest value we expect to see).

If a new data point is higher than the upper threshold or lower than the lower threshold, it gets flagged as an anomaly. This means that the value is unusual compared to what we would normally expect, based on the trends the model has learned.

Why This Matters:
By detecting these anomalies early, you can investigate what might be causing the unexpected changes. It could help uncover anything from system issues to unexpected market shifts, allowing you to take action quickly.

Important to Note:
Because the anomalies are based on predictions, they may not always be 100% accurate. There’s always a slight margin of error in forecasting, so it’s important to account for that when reviewing anomalies. While the model is designed to be as accurate as possible, occasional deviations are possible.

How We Expand Anomaly Detection Over Time:
When you first onboard with us, the model will provide anomaly detection for the past 31 days of data. As you continue to work with us and accumulate more data, we’ll be able to extend the anomaly detection period, offering insights for up to 90 days in the past.