Best Practices for Setting Up and Using Predictive Scores

Demandbase predictive scores help you prioritize accounts for marketing campaigns and sales interactions. To calculate these scores, Demandbase applies machine learning with data from all available sources, including your CRM and Marketing Automation System (MAS).

Best Practices

Use the following best practices to get the most out of your predictive scores.

1. Retrain predictive scores regularly to capture model improvements and to include additional activities. We recommend enabling the Auto Retrain option. We also recommend retraining scores everytime you update your Engagement Minutes or intent keyword configurations.
2. Avoid constricting predictive score models to your training accounts by using too many firmographic or technographic Selectors.
3. For Pipeline Predict Scores, avoid using Selectors to only include a limited set of engagement activities. Instead, use Selectors to exclude the engagement activities you don’t want used to train the model, for example outbound sales emails.
4. For Pipeline Predict Scores, use a minimum of 10 distinct customers with 100 opportunities. 
5. For Qualification Scores, use a minimum of 50 accounts. We recommend 100 or more accounts. 

Causes for Reduced Score Accuracy

The following setup may reduce the accuracy of your Pipeline Predict Scores:

• Training scores without enough opportunities:
  • Demandbase only trains the model using accounts that have preceding engagement prior to opportunity creation.
  • Demandbase does not use an opportunity to train the model if the account has a preceding opportunity that is Closed/Won.
• Training scores using upsell and cross-sell opportunities. Demandbase does not use upsell and cross-sell opportunities to train the model.
• Training scores using limited activities. Excluding too many activities when setting up a predictive score may cause reduced score accuracy.
• Using firmographic Selectors to filter down a list of accounts. Using firmographic Selectors to only include specific types of accounts can affect model accuracy.
• Not retraining scores after making updates to your engagement minutes or intent keywords. Retraining Pipeline Predict Scores after updating engagement minutes and intent keywords ensures relevant activities are incorporated into the model.

The following setup may reduce the accuracy of your Qualification Scores:

• Training scores without enough accounts. Demandbase requires a minimum of 50 accounts, but 100 or more accounts create a more accurate score.
• Using firmographic or technographic Selectors to filter down a list of accounts. Using firmographic and technographic Selectors to only include specific types of accounts can isolate inputs for those account types and affect model accuracy.
• Not retraining scores after adding technographics. Retraining Qualification Scores after adding technographic Selectors ensures that they are incorporated into the model.
• Not retraining scores after adding technographics or intent keywords. Retraining Qualification Scores after adding technographic and intent keyword ensures that they are incorporated into the model.