This process involves setting up an analysis job for Signals. These configuration settings will vary based on the nature of the business, available data, and users’ preferences. The ultimate goal here is ensuring that the correct number of the most relevant insights are delivered to the most appropriate users at the earliest available time.
Signals analysis configuration may require adjustment over time in order to optimize the settings, or as the underlying data changes. |
Signal configuration is done at the View level. Yellowfin will perform its analysis based on the setting details that you provide here; these must include:
Follow the instructions below to set up an automated Signals analysis schedule.
You will require the ‘Configuration’ role function to perform these steps. If you want to learn more about the Signals role functionality, refer to this section. |
In the next step, provide the basic details to describe this Signals analysis.
Continue to the next step.
The Time & Schedule step involves highlighting the time periods that you want the system to analyze, along with configuring how frequently this analysis should be executed. Below is a description of each of the settings. (Note: The fields and their purpose differ according to the different analysis types.)
Field | Analysis type | Description | Example |
Date field | All | The Signal analysis must include a date field that gets analyzed by the system. This is required to form a time series. You can choose a date field of any granularity (unit). | |
Granularity | Outlier detection, Trend changes | Specify the granularity (i.e. date unit) of the time series to analyze. This value should not be lower than the unit of your selected date field. | For example, if you have a date field with day, then your granularity must be Day or a higher value (Month, Quarter, etc.) |
Period comparison | Not required for Period Comparison.he granularity is set to be equal to the window size by default. | ||
Period Configuration | All | This selection changes the time window selection fields. Options are:
Note: For Outlier analysis, only a single time period needs to be specified. | |
Schedule | All | Set a schedule for when and how often frequently this analysis job should run. Note: For a weekly frequency or higher, use the Advanced Setting link for further specification. | You can schedule your analysis to automatically run every Monday at noon, for example. |
Basic compare fields | |||
Analyze the last | All | Specify the main time period window to be analyzed. | Analyze the last ‘month’ will simply analyze the previous month (going back from the current date). |
Compare to the same time last | Only Period compare and Trend changes | Specify a secondary time period to be analysed. This time period is compared to the main target period specified above. | Based on the above example, if this field was set to ‘quarter’, then the previous month will be compared and analyzed with the same month from the last quarter. (So the third month will be compared to the third month of the last quarter.) |
Advanced compare fields | |||
Window size | All | This advanced configuration specifies the length of the time period to be analyzed. In case of period compare and trend changes where two time periods are analyzed, both the time periods will be of the same length. | A window size of 3 months will analyze the last 3 months for Outlier detection; or compare the last 3 months, with previous 3 months prior to that, for Period comparision or Trend Changes. |
Offset | Period comparison | Use this advanced field to specify a time distance between the two time periods. A distance of the size specified here is formed from the end of the previous period to the beginning of the current period. An offset of zero creates two sequential periods. | Say your window size is 3 months (from 1st Jan to 31st Mar 2018), an offset of 11 months, will compare the same 3 months (1st Jan to 31st Mar) of the previous year, i.e. 2017. Or if offset is set to 0, compare the last 3 months with previous 3 months. |
Trend changes | Use this advanced field to specify a time distance between the two time periods. A distance of the size specified here is formed from the end of the previous period to the end of the current period. An offset of 1 window size unit creates two sequential periods. | For a window size of 3 months, where the offset is set to 1 year, the last 3 months are compared to 3 months from a year ago. | |
Outliers | Specifies how long ago should the target window size be analyzed. | If window size is 3 weeks, and offset is 5 months, then analyze 3 weeks from 5 months ago. | |
Fixed range fields | |||
Target Period | All | Specify a fixed date range to analyze using the date picker | |
Compare period | Only period compare, trend analysis | If your selected analysis compares two time periods, then specify the other one here. Note that the target period must be after the compare period. Both the periods should be of similar length. |
In the Metrics & Dimensions step, you can select or unselect the fields in your data, that are important to you. Yellowfin will analyze these fields in order to detect Signals.
Fields that are also selected for Assisted Insights, will be selected here by default. |
Maximum alerts: You can set a limit to the number of alerts that are generated each time a Signals analysis is run.
Advanced parameters for Outliers analysis:
Parameter | Description | Example |
Spikes, drops, and steps | This algorithm looks for sudden increases (steps) or decreases (drops) in metric value. If successive Steps or Drops are detected, they are reported as a Step Up or Step Down respectively, signifying that there has been a breakout. Use this toggle to enable this algorithm and bring up additional configuration options related to it. | Example of a spike, a lot of extra website visits occurred on a specific day. |
Lag | Specify a numeric value as lag to calculate the moving average. It affects how slowly the moving average will reflect the changes in the time series trend. The higher the lag, the smoother will be the moving average, which will be more tolerant to erratic data. This value should be lower than the number of data points in each analysis period to generate a signal. Note: The minimum value of 3 is required for the lag, no lower than that. | A typical value of 10% of the data might generate signals. For a month (with roughly 30 days) worth of daily data, use a lag of 3 (or higher). |
Influence | Define a numeric value between 0 and 1 which controls how strongly an outlier will impact the moving average. A high influence value will cause the moving average to more closely follow the time series. | A typical value for influence is 0.3. |
Threshold | Specifies the moving average corridor size in standard deviations. A higher value causes the corridor to be larger, requiring a larger magnitude of change to detect an outlier. This value defines the width of the border used to detect outliers. A high value will mean that outliers detected above it will be identified, resulting in big anomaly detection. | A typical value of 3.5 to 4.5 should detect significant outliers. |
Advanced configuration for Period Comparison:
Parameter | Description | Example |
Aggregate value changes | This algorithm detects changes in sum and average value of a metric from one period to another. Use this toggle to enable this algorithm and bring up additional configuration options related to it. See below. | Eg. Total sales amount has dropped significantly this month compared to last month. |
Sum aggregates | This algorithm detects changes in the total value of the metric field. Use this toggle to enable this algorithm. | |
Threshold % | A change in total must have an absolute percentage greater than this threshold to constitute a signal. | For example, a threshold of 10 will result in any sum changes greater than 10% being created as Signals. |
Threshold Absolute | A change in sum must have an absolute value greater than this threshold to constitute a signal. | For example, a value of 20 will consider any sum changes greater than 20 as a signal. |
Average aggregates | This algorithm detects changes in the average of a metric field. Use this toggle to enable this algorithm. | |
Threshold % | A change in average must have an absolute percentage greater than this threshold to constitute a signal. | For example, a threshold of 10 will result in any average changes greater than 10% being created as Signals. |
Threshold Absolute | A change in average must have an absolute value greater than this threshold to constitute a signal. | For example, a value of 20 will classify any average changes greater than 20 as a signal. |
New and lost attributes | This algorithm detects key dimensions that entered or left the data this period. Use this toggle to enable this algorithm and bring up additional configuration options related to it. See below | Eg. A customer that bought a lot last month stopped buying this month. |
Minimum significance | A new or lost attribute's total series percentage must be greater than this threshold to constitute a signal. | A minimum significance value of 20% would mean an attribute must make up more than 20% of total data to be detected as a signal. |
Advanced parameters for Trend Changes:
Parameter | Description | Example |
Linear Trend Changes | This algorithm detects changes in trend line slopes from one period to another, such as trend going from up to down, or growing significantly faster or slower. Use this toggle to enable the algorithm, and configure settings related to it. | PC sales were growing but started to decline last month. |
Threshold % | A change in trend slope must have an absolute percentage greater than this threshold to constitute a signal. | For eg, a value of 10 will result in any trend slope change greater than 10% being created as Signals. |
Threshold absolute | A change in trend slope must have an absolute value greater than this threshold to constitute a signal. | For example, a value of 20 will consider any trend slope change value greater than 20 as a signal. |
Flat tolerance | This numerical value will form a range about zero (that is, the same value in -/+ range) and any trends which have a slope falling inside this range will be considered to be flat. Values falling out of this range will be considered as having a positive slope (if a positive value), or a negative slope (that is the trend is going down because of a negative value). | For example, a flat tolerance of 0.01 means that if a trend line slope falls within -0.01 to 0.01 then it will be considered to be flat. |
Volatility changes | Detect changes in volatility between the compare and target periods. Measures changes in consistency of metric values. Use this toggle to enable the algorithm, and configure settings related to it. | Eg. Daily sales of Blue Shoes was sporadic last month, but has become more regular. |
Threshold % | A change in volatility must have an absolute percentage greater than this threshold to constitute a signal. | For example, a threshold % of 10 will result in any volatility changes greater than 10% to be classified as a Signal. |
Threshold absolute | A change in volatility must have an absolute value greater than this threshold to constitute a signal. | For example, a value of 20 will classify any volatility changes greater than 20 as a signal. |
Continue to the next step.