What This Chart Shows
This chart provides a deep-dive diagnostic view into a specific data series, showing the raw values over time alongside all detected anomalies and the thresholds used to detect them. It helps you understand why certain data points were flagged as anomalies and what type of anomaly they represent. This is a diagnostic tool for data quality investigation, model debugging, and understanding the anomaly detection process.
Key Questions This Chart Helps Answer
- What does this series' raw data look like over time?
- When and where did anomalies occur in this series?
- What type of anomaly was detected (structural, flow, or statistical)?
- Why was a specific data point flagged as an anomaly (which threshold was exceeded)?
- How do the different detection thresholds compare over time?
Axes, Metrics, and Units
Element | Description |
|---|---|
| X-axis (both panels) | Calendar time: Date labels at month boundaries, shared between panels. |
| Y-axis (Top Panel) | Raw Value: The actual observed value for this series over time. |
| Black line (Top Panel) | Raw Data: The time series of observed values. |
| Y-axis (Bottom Panel) | Error / Threshold: Reconstruction error and detection threshold values. |
| Red solid line (Bottom Panel) | Reconstruction Error: How far the observed value is from the model's prediction. |
| Blue triangle-down | Structural Anomaly: Detected change in data structure or pattern. |
| Purple triangle-up | Flow Anomaly: Detected change in data flow or distribution. |
| Red circle | Statistical Anomaly: Detected statistical outlier. |
| Green dashed line | Short-window adaptive threshold. |
| Blue dashed line | TDigest Threshold: Percentile-based threshold from streaming quantile estimation. |
| Black solid line | Final Threshold: The combined threshold used for anomaly detection. |
| Red dotted line | Short Fallback: Fallback threshold when short-window data is insufficient. |
| Orange solid line | Structural Distance: Distance metric used for structural anomaly detection. |
| Yellow solid line | Flow Distance: Distance metric used for flow anomaly detection. |
All values are raw (not normalized). Anomaly markers appear on both panels at the same time points.
Control Options Reference
Control | Meaning |
|---|---|
| Event Classification | Filter available series by event type (e.g., Statistical Anomaly, Narrative Mention). |
| Dimension Value | Filter by specific dimension values (e.g., search query, campaign name). |
| Search Series | Text search to filter series by name, channel, or source. |
| Select Series | Dropdown to choose the specific series to inspect. Series are ranked by anomaly count. |
You must select a series to view the chart. Without a selection, the chart displays instructions on how to select.
How to Interpret the Results
- Top Panel (Raw Data): The black line shows the actual data values. Anomaly markers (triangles, circles) appear at time points where the model detected something unusual.
- Bottom Panel (Error vs Thresholds): The red line (Recon Error) shows how far each observation was from the model's expectation. When the error exceeds the Final Threshold (black line), an anomaly is flagged.
- Anomaly types:
- Structural: The pattern or structure of the data changed (e.g., a new trend or regime shift). Blue triangle-down.
- Flow: The distribution or flow of data changed (e.g., a sudden spike or drop). Purple triangle-up.
- Statistical: A simple outlier based on statistical thresholds. Red circle.
- Threshold comparison: Multiple thresholds are shown to help understand how the final decision was made. The Final Threshold is the operative one; others are diagnostic.
- Hover: Shows the exact date and value for each point.
Practical Applications for Marketers
Application | How to use this chart |
|---|---|
| Data quality investigation | Inspect flagged anomalies to determine if they are real events or data errors. |
| Campaign impact assessment | Check if a campaign launch coincides with a detected anomaly in a KPI series. |
| Model debugging | Understand why certain points were flagged (or not) by comparing error to thresholds. |
| Event root cause analysis | Use structural vs flow vs statistical classification to understand the nature of the change. |
| Threshold tuning validation | Verify that threshold settings are appropriate for this series' behavior. |
Common Mistakes & Misinterpretations
Mistake | Why it is a problem | How to avoid |
|---|---|---|
| Assuming all anomalies are bad | Anomalies can be positive (e.g., a viral campaign) or negative (e.g., data outage). The chart only detects unusual, not good vs bad. | Investigate the context of each anomaly before drawing conclusions. |
| Ignoring anomaly type | Structural, flow, and statistical anomalies have different causes and implications. Treating them the same loses information. | Use the anomaly type to guide your investigation (structure = regime shift, flow = distribution change, statistical = outlier). |
| Over-focusing on threshold lines | Multiple thresholds are shown for diagnostic purposes. Only the Final Threshold determines anomaly flags. | Focus on the Final Threshold (black) for understanding what triggered detection. |
| Comparing across different series | Each series has its own scale and threshold dynamics. Comparing thresholds across series is not meaningful. | Use this chart for one series at a time; use other charts for cross-series comparison. |
| Treating reconstruction error as prediction error | Reconstruction error is how far the value is from the model's learned representation, not a forecast error. | Interpret error as "unusualness" relative to learned patterns, not as a forecast miss. |
Caveats & Considerations
- Series selection required: You must select a specific series to view this chart. Without a selection, the chart shows instructions.
- Threshold interpretation: Multiple thresholds are shown for transparency, but only the Final Threshold is used for detection. Others (Welford, TDigest, Fallback) are intermediate calculations.
- Anomaly classification: A data point may be flagged as multiple anomaly types simultaneously if it triggers multiple detectors.
- Time range: The chart shows the full available time range for the selected series. Use hover to see exact dates.
- Diagnostic tool: This chart is intended for deep-dive investigation, not for high-level portfolio views. Use other charts for cross-series or aggregate analysis.