Anomaly Detection Settings

Preview

When setting up an Anomaly it's useful to understand the sensitivity of the algorithm, in order to know how many alerts you are likely to receive. So in the preview you can see the last 3 months for Daily Time Period and the last 7 days for Hourly Time Period. 

The Time Period field determines the data points that are used to determine Anomalies, therefore when selecting Hourly the algorithm will input each hour of the day to calculate the presence of an Anomaly. This is not to be confused with Frequency, which is how often the algorithm is run to check for the presence of Anomalies (Only available on higher plans).

NOTE: Not all Time Period, Metric and Segment combinations are possible, and you will receive an error if the combination is invalid.

Advanced Settings

You have the option of customizing the parameters used in the Anomaly Detection algorithm. See below for explanations on each setting and when to adjust them.

Sensitivity

This sets the significance level for anomaly detection. Decrease the sensitivity when your data is in stable conditions like regular sales periods to minimize false positives. Increase during volatile periods like holiday seasons to capture more potential anomalies.

Threshold of Positive Anomalies

Options:

• 99th Percentile: This sets the threshold at a very high level, only flagging data points that exceed 99% of all other values. It's useful for extremely conservative anomaly detection, where only the most extreme cases are considered anomalies.
• 95th Percentile: This is a less stringent threshold than the 99th percentile but still focuses on significantly high values. It's suitable for cases where you want to detect anomalies that are not just the most extreme.
• Median of the Data Max Values: This option sets the threshold at the median of the highest values in your dataset. It's a more adaptive approach and can be useful in datasets where the range of maximum values varies significantly.
• Default Setting for Digital Marketing Data: A recommended default setting could be the 95th percentile. This setting is a good balance for digital marketing data, where you might want to be alerted to significant spikes in metrics (like website traffic or ad clicks) that are unusual but not extremely rare. The 95th percentile will capture notable deviations without being overly sensitive to every fluctuation.

The choice of threshold depends on the nature of the digital marketing data and the specific goals of the analysis:

• For highly variable marketing campaigns (e.g., during product launches or major promotions), you might prefer the 99th percentile to avoid too many false alarms.
• For more stable, ongoing campaigns, the 95th percentile could be a better choice to ensure you're alerted to any significant changes.
• If the dataset is particularly noisy or uneven, using the median of the data max values might provide a more tailored threshold.

Max Anomaly Ratio

This number balances sensitivity and practicality, assuming digital marketing data would not typically have a large number of anomalies.

When to Increase:

1. During High-Impact Periods: Increase during major marketing campaigns, product launches, special events (e.g., Black Friday), or when testing new marketing strategies. These periods typically see surges in data points like web traffic or social media engagement.
2. Post-Major Changes: If there have been significant changes to your website or marketing platforms, an increased threshold helps in identifying the impacts of these changes on user behavior.
3. Unstable Market Conditions: In times of crisis or unexpected market shifts, a higher setting helps in quickly identifying a range of anomalies in consumer behavior or campaign performance.

When to Decrease:

1. For Regular Monitoring: In stable conditions, like monitoring ongoing campaigns, a lower value helps focus on the most significant anomalies, avoiding noise from minor, day-to-day fluctuations.
2. In Small Businesses or Niche Markets: For smaller businesses or niche markets with fewer and less variable data points, a lower setting is more suitable, as large anomalies are less common.

The key to adjusting the ratio is to consider the context of your data and the specific objectives. Increase it when expecting a high volume of significant data fluctuations or during periods of uncertainty. Decrease it for more routine monitoring or in more stable, predictable environments.

Piecewise median time window

This parameter dictates how the data is segmented for analysis. Set this to 7 for weekly patterns like retail foot traffic, or 12 for monthly patterns like subscription renewals. Here’s a simplified guideline on when to adjust this hyperparameter:

When to Increase:

1. For Long-Term Trends Analysis: Increase the window size when you’re interested in understanding long-term trends in your data. This is particularly relevant for analyzing seasonal effects or long-duration marketing campaigns where changes occur over weeks or months.
2. For More Stable, Consistent Data: In scenarios where the data is generally stable and consistent, a larger window can help smooth out minor fluctuations, allowing a clearer view of significant, sustained changes.
3. For High Volume, Less Granular Data: If your dataset is large and less granular (e.g., aggregated monthly data), a larger time window is appropriate for capturing the broader patterns and trends.

When to Decrease:

1. For Short-Term Fluctuations Analysis: Decrease the window size when focusing on short-term fluctuations and anomalies. This is useful for tracking the immediate impact of specific marketing activities like flash sales, short-term promotions, or social media campaigns.
2. For Highly Variable or Granular Data: In cases where the data is highly variable or granular (e.g., daily or hourly data), a smaller window allows for a more sensitive detection of sudden changes and anomalies.
3. For Niche or Targeted Campaigns: When analyzing the effects of targeted marketing efforts or niche campaigns, a shorter time window can provide more relevant insights into their immediate impact.

Long term

Disable this option when your primary focus is on short-term effects, like the immediate impact of a specific marketing campaign, promotional event, or sudden market changes.

Overriding the fetching time

Some data sources can take a while after the day has ended for post-processing and the data considered stable. Therefore use this setting to delay the default fetching time of 9am in your data sources timezone.