Pega Sales Automation uses decisioning capabilities to provide predictive insights for opportunities.
The application provides the following predictions:
Opportunity insights are displayed in the opportunity record for each opportunity. The widget is dynamic and responds to changes, such as stage changes and changes in customer activity or contact growth.
The following figure displays sample data for an Opportunity insight:
Probability to move to next stage
Adaptive models predict the likelihood that an opportunity will move from the current stage to the next stage. The application uses the average number of all opportunities that have moved from the current stage to the next stage as the base propensity. The difference between the likelihood that an opportunity will move and the base propensity is an indicator of how you are progressing with the opportunity.
Adaptive models predict the likelihood of winning an opportunity. The application uses the average number of all won opportunities for a given stage as the base propensity. The difference between the likelihood that you will win an opportunity and the base propensity is an indicator of how you are progressing with the opportunity.
Adaptive models predict the quarter when an opportunity is most likely to close. The application compares the target date range that is set by the sales representative and the predicted closing quarter to indicate whether the opportunity close date is earlier than expected, on time, or delayed.
Historical data and adaptive learning
The application uses self-learning adaptive models to generate opportunity predictions. This approach is based on core decision management capabilities and provides the flexibility to add and remove predictors as your needs change.
- Historical data – The adaptive model provides a set of key predictors that are drawn from an analysis of data from actual production environments. Production data was used to create weekly snapshots for each closed opportunity over a two-year period. Over 2,000 opportunities were used to create the weekly snapshots and approximately 50,000 records were used to define the key predictors and train the adaptive models.
The following predictors are provided in the adaptive model:
|Pega Sales Automation||Pega Sales Automation for Financial Services|
|B2B and B2C opportunities||B2B opportunity||B2C opportunity|
|.AllTotal30Vs90||.AllTotal30Vs90||.Age (months or years in business)|
| ||.StageDuration|| |
| ||.StageSequence|| |
| ||.TotalOutstandingDebt|| |
| ||.ValueOfCollateral|| |
- Adaptive learning – An agent runs daily to execute a data flow that makes a call to a decision strategy. The decision strategy contains all of the adaptive models and runs on all of the open opportunities to capture the data that is required by the models. The decision strategy uses the standard Delayed Learning cache. When an opportunity is closed, the application triggers a response strategy from a data flow to fetch the data for relevant decisions and train the applicable models. You can use the same decision strategy at any time to evaluate an opportunity and return propensities.
The following figure displays an overview of the solution:
The application uses a single strategy to drive the insights that are displayed on the dashboard widget and to take periodic predictor snapshots. The strategy runs all three model types (move to next stage, opportunity win, and win date), distinguishing them by using labels that are mapped to the issue and group hierarchy. For the opportunity win and move to next stage models, the strategy also evaluates a model without predictors (opportunity win base propensities) to establish a base propensity for benchmarking.
The following figure displays the strategy flow:
The solution includes a set of models that provide predictors based on historical data.
Prediction: Will this opportunity ever be won?
The input predictor range should be as complete as possible. All models share the same predictors, although they might treat them differently.
- Opportunity win base propensity model
This model does not use predictors. Instead, the model uses an opportunity stage as the context to provide the base propensity of winning an opportunity in the current stage.
Prediction: Will this opportunity ever move to a higher stage?
When an opportunity moves from the current stage to the next stage, the application sends a positive, one-response to all opportunities in the Adaptive Decision Manager (ADM) data cache. The application also takes a snapshot of the previous stage. The response is filtered by using an additional Decision Data Store (DDS) dataset. The response strategy for the move to next stage model filters out the decision results for stages that have already been marked as positive. It then determines whether the current stage is higher than the stage in the snapshot, and returns only the decision results for which this is true.
- Move to next stage base propensity model
This model does not use predictors. Instead, the model uses an opportunity stage as the context to provide the base propensity of moving from the current stage to the next stage.
The following figure displays the move to next stage model:
ADM does not support exact date win predictions. Instead, ADM supports separate models for the following outcomes:
- Will this opportunity be won in 0-90 days from now?
- Will this opportunity be won in 90-180 days from now?
- Will this opportunity be won in 180-270 days from now?
- Will this opportunity be won in 270-360 days from now?
The application uses propositions in a Decision Data shape to model the date ranges. The propositions have both a minimum and maximum days attribute. If you require more granularity, you can edit the propositions to meet your needs. You can also draw a spline in the flow to create a smoother display for the dashboard widget.
- Win date base propensity model
This model does not use predictors. Instead, the model uses an opportunity stage as the context to provide the base propensity of winning an opportunity in a specified time frame.
The following figure displays the win date model:
The solution uses the D_PredictSAOpportunity data page.
For the Pega Sales Automation application, review the following D_PredictSAOpportunity data page description:
The Analytics widgets in the opportunity run in the context of the D_PredictSAOpportunity data page. This data page has the SAPredictOpportunity activity as a data source, which runs the GetContactGrowthRatio, GetEmailActivityRatio, GetTrendsDataRatio, and GetCIRatio activities before calling the SAPredictOpportunity data flow. The data flow uses the EvaluateOpportunity strategy to predict the propensity to move to the next stage, to win an opportunity, and to predict the close date quarter.
For the Pega Sales Automation for Financial Services application, review the following D_PredictSAOpportunity data page description:
This data page populates the propensity for an Opportunity case. The Analytics widgets in the opportunity run in the context of the D_PredictSAOpportunity data page. This data page has the SAPredictOpportunity activity as a data source, which runs the PopulatePredictors activity. The PopulatePredictors activity replaces the following activities: GetContactGrothRatio, GetEmailActivityRatio, GetTrendsDataRatio, and GetCIRatio, which are available in the Pega Sales Automation application. When there are multiple products specified, the PopulatePredictors activity iterates all of the products propensity, but displays the least propensity model. This workflow is applicable only for the credit and debit product types.
Review the following data flows list:
The solution uses the StoreOpportunitySnapshots data flow during the initial training of the models by using the data that is fetched from the internal production data. This data flow converts each record from the PegaCRM-Data-SFA-Predictors class into opportunity objects and then routes them to the EvaluateOpportunity strategy. In the data flow, the Mode for the EvaluateOpportunity strategy is set to Make decision and store data for later response capture.
The input for TrainFromHistory comes from the RD OpportunityStages, which fetches one record per opportunity from the predictors’ data table, sorted by maximum stage. This data flow calls the OpportunityClosed and MovedToNextStage data flows.
The OpportunityClosed data flow calls the CloseOpportunity strategy, which captures responses for decisions in the past period.
The MoveToNextStage data flow calls the HandleNextStageResponses strategy, which is the core strategy for training the PredictNextStageModels adaptive model. The available responses depend on whether an opportunity moves up or down from the current stage.
The SnapshotOneOpportunity data flow reuses the EvaluateOpportunity strategy, which is used to train the model during bulk processing. This data flow runs on a daily basis and captures all predictors for the opportunity each time it runs.
The SAPredictOpportunity data flow reuses the EvaluateOpportunity strategy to get the analytic results for the opportunity.
The StoreOpportunitySnapshots and the TrainFromHistory data flows are not applicable for the Pega Sales Automation for Financial Services application.
You can view details for each data flow in Dev Studio.
- In the panel, click .
- Click Data Model > Data Flow.
- To open the data flow record, click a data flow name.
The solution uses the following strategies:
You can view details for each strategy in Dev Studio.
- In the panel, click .
- Click Decision > Strategy.
- To open the strategy record, click a strategy name.
Each adaptive model has predictors, context, and outcomes that you must configure before you can train the models to predict outcome propensities based on your use cases. Base propensity models do not have any predictors, but they have the OpportunityStage property in their context so that you can train different models for each stage.
Outcome propensities are configured for the following models:
You can view details for each model in Dev Studio.
- In the panel, click .
- Click Decision > Adaptive Model.
- To open the model record, click a model name.