How to Set Up Predictive Models in Demandbase

What you’ll learn

This comprehensive playbook will walk you through setting up two distinct predictive models within DemandbaseOne; Qualification Score and Pipeline Predict. Not only will you learn the technical steps, but you will also learn how different factors help drive each model, such as the data you input and the GTM strategies within your organization. By the end, you’ll have the tools to set up and customize these models for your unique needs while also avoiding potential pitfalls.

What are predictive models

Qualification Score

Qualification Score is designed to indicate how similar an Account is compared to a group of Accounts you have given the model for training. This concept of similarity is derived by looking at the Firmographic and Technographic makeup of an Account as well as their Intent signals over the past year. A simple example of this is to feed the model a group of your best Customers, the model will then give each Account a score of 0 to 100 to indicate how similar an Account is to that group of Customers. However you can also use this model to ultimately find look-a-like Accounts so utilizing it for your ideal customers as you launch a new product line is also a fantastic use case.

Pipeline Predict

Pipeline Predict on the other hand is designed to identify which Accounts are ready to start engaging with your Sales Team in hopes to purchase your product(s). Just like Qualification Score, you will need to provide the model with a group of Accounts for training except now instead of Customers we want to focus on Accounts that have opened an Opportunity with your organization as well as have engagement activities prior to the Opportunity being created. In this model we will once again look at the Firmographic and Technographic makeup of the Accounts, but now also include the activities within the Account to help identify how likely they are to open an Opportunity with your Sales team. It will then rank Accounts with a score of 0 to 100. An example of this would be to give the model a list of Accounts that have Opened Opportunities over the past 1 year for your primary product in order to identify Accounts for your Sales Team to follow up with both on the New Business and Customer side. 

What you need to get started

Prerequisites for Qualification Score

• Accounts added to Demandbase One for Marketing (DB1M)
• Demandbase Intent Keyword Sets created in alignment with your GTM strategies
• Collaboration across GTM Teams to align on the purpose of the model(s)

Prerequisites for Pipeline Predict

• Live CRM and MAP integrations with DB1M
• Engagement Minutes Model is updated for CRM and MAP activities
• Ensure your Account Journey has at least one stage designated for Pipeline and one for Customer. This will aid the model in understanding how you view new business and growth opportunities

Predictive models depend on the completeness and relevance of your dataset. Without properly structured accounts, working system integrations and well-defined activities, your models may not perform optimally.

GTM strategy review

Predictive models are highly dependent on input data, which must align with your specific objectives and strategies for the best results. Please go through this brief three question GTM Strategy Review process in order to understand the best methodology for building your models. 

1. Do you currently use one or more Account Journeys in DB1M?
2. Do you sell a single or multiple products?
   1. If you sell multiple products are they sold to the same Accounts or would they never be sold to the same Account? (An example of the latter would be if Product A was sold only to Healthcare companies and Product B was only sold to Aviation companies) 
3. Do you sell into different countries?

• Determining how many Account Journeys you use in DB1M should directly align to the amount of models you create. The goal is to have one paradigm; One Account Journey which uses one of each Predictive Model.  
• If you answered yes to selling a single product, you should be able to utilize one Account Journey and one of each Predictive Model
• If you answered yes to selling multiple products, but they are sold to the same Accounts, you should be able to utilize one Account Journey and one of each Predictive Model. If you sell multiple products to different types of Accounts you would benefit from having multiple Account Journeys and Predictive Models. 
• If you answered yes to selling into different Countries, you would benefit from using multiple Account Journeys and multiple Predictive Models for each Country grouping; for example North America vs Europe. This is because the more granular we can be in a model’s training set the higher the precision of the model will be and the more than likely the engagement needed to move Accounts down the funnel will vary. Additionally this ensures a non-bias within the model by ensuring each country based model has enough data to train with precision.  

Having multiple Predictive Models in a singular Journey can cause confusion across your entire GTM organization, it’s a best practice to make your Account Journey definitions as simple as possible in terms of why an Account would qualify for a stage. For example, if you have 2 Pipeline Predict Models used in the MQA Stage, all of your Marketing efforts based on that stage need additional criteria to specify the model score and your Sales team would always need to identify which model caused the Account to move into that stage

Step-by-step playbook

Part 1: Setting Up Qualification Score

1. Access the Qualification Score Section
   1. Log into DB1M, click on Settings, and choose Qualification Score within Predictive Models.
2. Identify Accounts for Training
   1. Select the accounts you want to supply as the model’s training dataset using one of three approaches:Use hard-coded CRM Account IDs.Reference CRM Account fields.Utilize the Accounts with Opportunities relationship to reference specific opportunity characteristics within Accounts
   2. When picking a methodology for identifying these Accounts you should be aware that this model will automatically retrain itself every 6 months and the methodology you selected can impact the model during the re-training
   3. If you choose to use hard-coded CRM Account IDs and do not update those IDs when the model retrains, you will be missing out on any Account that could have qualified for the group since the last training. While this is not going to break anything you are missing out on crucial new data that could impact your model’s output
   4. If you choose to reference either the Account and/or Opportunity fields, make sure you do so in a way that when the model retrains those fields still represent the right Accounts. One issue we typically see is referencing Opportunity fields that at one point indicated the Account as a Customer, however since the last model training that Account has churned and is no longer a Customer. Again this will not break your model but will be feeding it data that is potentially no longer accurate if you want to train the model with only Customer Accounts
3. Add Intent Data
   
   1. Under the Keyword Sets section, choose Intent Keyword Sets relevant to the overall strategy of the Qualification Score Model. For example, if you built a model to identify customers for Product A, you should select Keyword Sets relevant to Product A. 
4. Review Data Availability
   
   1. Use the Refresh button to check that your criteria capture a sufficient volume of accounts. A dataset of at least 50 accounts is required as a baseline; however, using 100+ accounts provides greater consistency and accuracy.

Part 2: Setting up pipeline predict

Identifying Criteria

The first thing you want to do with any Pipeline Predict Model setup is to update the out-of-the-box setting to ensure you are capturing all accounts that have opened an opportunity in the last year. Now given the strategy of your model, this could have additional criteria, such as product, region, or industry. One thing to consider is that if your Operations team creates a Renewal opportunity soon after closing a new business opportunity, you will want to restrict those renewal opportunities from entering your model. Pipeline Predict is based on how to land new and expand existing customers, feeding it “system-created” renewal opportunities can incorrectly bias the model’s calculations. We highly recommend you work with your Operations team to ensure you are using the right criteria to find these accounts and opportunities.

The Pipeline Predict Model utilizes similar data as the Qualification Score Model, except now it will include the vast activities associated with an account to determine how likely it is the account will open a new business or customer opportunity with your organization. Due to this, you must ensure that you only include the activities that will contribute to the accuracy of your model. 

Regardless of your GTM Strategy, these best practices for identifying the activities to include and exclude from your model should always be followed.

Activity Inclusions

• When selecting Activities to include in your model, leave the out-of-the-box setting in place within the model to ensure you only include activities with an Engagement Minute score of above 0
• Ensure you are capturing all relevant marketing activities such as High Strength Demandbase Intent, Demandbase Trending Intent, Web Visits, and Marketing Programs/CRM Campaigns.

If you are using a combination of CRM and MAP technologies that do not have the concept of Marketing Programs or CRM Campaigns, use the Form Submission activity to identify activities with your Marketing efforts.

Activity Exclusions

• Exclude Low and Medium Strength Intent from Demandbase. Accounts can exhibit dramatic amounts of intent at any given time. Due to this, you should train your model on the most valuable intent signals; High Strength and Trending Intent.
• Exclude all Sales Activities. This model is designed to identify accounts ready to speak to Sales, so there is no need to include activities from your Sales Team.
• Exclude Email Opens and Email Clicks. Most organizations have extra layers of security around their email system, which leads to Marketing Automation Platforms registering false positive activities for email engagement. 
• If you have connected Marketo to DB1M, exclude Interesting Moments from the activities ingested by the model. Interesting Moments are defined as clusters of activities that the model will automatically identify.

Part 3: Types of models

Now that we have reviewed the basic criteria setup of a model, let’s dive into our two types of model setups: Simple and Advanced.

Simple Model

Best suited for organizations with a GTM Strategy of selling one primary product or service and/or are focused on specific geographical regions or segments of accounts. This setup will mirror what we have previously seen with some added focus.

Holistic Model

In this example of a simple model, the organization is simply using the recommended inclusion and exclusion criteria mentioned above, with no geographical or account segmentation focus.

Geographical Model

However, in this next example, the organization has decided this model will be used for just accounts located in their North American territory. As you can see, the setup is almost the same however we are only allowing accounts into the model which meet the territory criteria. If you were to utilize this methodology you would most likely be building a model for each territory.

Account Segment Model

However, in this next example, the organization has decided this model will be used for just accounts located in their North American territory. As you can see, the setup is almost the same however we are only allowing accounts into the model which meet the territory criteria. If you were to utilize this methodology you would most likely be building a model for each territory.

Advanced Model

Best suited for organizations with a GTM Strategy of selling multiple products to different buying groups within an account or to completely separate accounts. For the former example, imagine you sell a product to a buying group to land within an account, but over time you expand into other buying groups within that account. Since the buying groups are unique within that account, we recommend using a different model since your strategies would be different. A great example of the latter would be a company like General Electric and the fact they sell into the Aerospace and Healthcare industries. Those two strategies require selling to vastly different accounts and utilizing very different strategies, which is why we would recommend creating multiple models, one for each overarching strategy.

Below we will cover one example of this by building a Product Focused Pipeline Predict Model.

Product Focused Model

Building a product-focused model requires two main components: finding accounts that have opened an opportunity with the product of focus and identifying activities related to that product. It is important to continue to use the activity inclusion and exclusion recommendations mentioned above, however now you will also need to decide if you want to hyper-focus your model on activities only related to the product or be more generalized with the activities you include. 

If you choose to be more generalized in your activity inclusion, your Pipeline Predict Model may look similar to the image below. However, you can also implement some of the other techniques mentioned in the Simple Model, such as geographical and account segmentation, to further enhance your model.

Now if you want to get more specific in the activities inclusion, you could also simply reference the activities that relate to your focused product. The easiest way to do this would be to build an Activity Segment, which classifies activities by product. While we want to include those product-specific activities, we would also want to include generalized activities, since not all activities will be product-focused. In the image below, you will see how we implement this activity segmentation in addition to our more generalized activity inclusions.

Where can this be used?

Once trained, these models enable a litany of various use cases across Demandbase products and your existing tech stack. While the list below is not exhaustive, these are areas where customers start to implement these model scores into their existing GTM Strategies. 

DemandbaseOne Marketing

• Building Lists and Filters to find similar accounts or accounts likely to engage with sales
• Create Audiences for Programs and Campaigns
• Use in Automation criteria for launching programs in your Marketing Automation or activating Sales Engagement Platforms
• Embed in existing and/or new reports
• Embed in Slack Alerts to Sales
• Account Journey Stage criteria

DemandbaseOne Sales

• Used as part of the prioritization mechanisms in the Predictive Sales Dashboard

CRM

• Add Model Scores to Account Layouts
• Embed into new and existing reports for deeper prioritization
• Review Qualification Score as opportunities move down the funnel to aid forecasting
• Identify new customer growth opportunities
• Prioritize how you do account enrichment based on Model Scores

Business Intelligence Tools

• Snapshot Pipeline Predict Score to find trends in different account segments
• Identify trends with both models as accounts move down the funnel to gain deeper insight
• Embed model scores into existing internally built models as new characteristics