How does the opportunity Predicted Win Probability get calculated?
Oracle Sales uses a built-in machine learning model to calculate the predicted win probability of a given opportunity using the historical data about your won and lost opportunities.
The model is automatically retrained on the prior 3 years' worth of data every month.
The model recalculates the Predicted Win Probability every 12 hours based on updates salespeople have made in the meantime.
For the Predicted Win Probability changes to be made visible in the UI, the opportunities must be first indexed for Adaptive Search. The indexing process, which runs often, indexes a record depending on the date and time of the last update (Last Updated Date). Any change to the record normally updates that date. Since many changes in the Predicted Win Probability are very incremental, however, Oracle prevents unnecessary reindexing of records by using the profile option AI Update Who Columns Threshold (ORA_ZCA_AI_UPDATE_WHO_THRESHOLD).
By default, if the Predicted Win Probability changes by less than 5 percent, there's no update to the Last Updated Date. You can set this profile to a different value by using the Manage Administrator Profile Values task in the Setup and Maintenance work area. The threshold you apply here also applies to changes in the AI Lead Score.
When you enable the Predicted Win Probability feature, the application checks that you've enough historical data for the model to be trained. It checks you've at least the following:
-
1,000 closed opportunities
- 100 lost opportunities
- 1,000 opportunity contacts
- 1,000 opportunity stages
- 100 won opportunities
These are just the bare minimum requirements. Chances are that you won't get meaningful predictions on small numbers.
The model trains on the shape of your data for the different opportunity stages. Here are the objects the data is drawn from:
-
Account
-
Activity
-
Contact
-
Opportunity
-
Opportunity Contact
-
Opportunity Revenue
-
Opportunity Stage Snapshot
The snapshot compares information in opportunities by stages. For example, it compares the activities in the initial Qualifying stage to the same stage in won and lost opportunities.
Attributes that don't include enough data are discarded. For example, if your opportunities typically include few contacts, the contact information will be excluded from the model and predictions. Attributes with lots of data are combined into features, which the model uses to make the predictions.
While you can't configure or copy the model itself, you can make some changes:
- You can exclude records from being used in the model training and predictions
For example, you can exclude opportunities that have completely different sales processes from those you're trying to predict, because including them would yield inaccurate predictions. If you're primarily selling equipment, for example, you may want to exclude subscription renewals. You can exclude opportunities using multiple attributes, including sales method and sales channels, for example.
Excluded records still display the Predicted Win Probability field, but the field value is blank.
-
Use custom attributes to train the model.
While you can't add attributes to the model, you can substitute custom attributes for a subset of the attributes included in the model. The following objects permit substitutions for a small subset of existing attributes:
-
Activity
-
Opportunity
-
Opportunity Contact
-
Opportunity Stage Snapshot
-
| Callout | Description |
|---|---|
| 1 | Select the object you want to configure. |
| 2 | Specify the attributes and values that you want to use to exclude records. You can enter multiple filters. |
| 3 | You can replace the attributes listed under the Model Feature Configuration heading with your own custom attributes. |
