Below are answers to the most frequently asked questions about how rAI-powered Predictive Modeling works. If you have additional questions, contact your Customer Success Manager or reach out to resonatesupport@resonate.com.

General FAQ

• What is rAI-powered Predictive Modeling
• How does Resonate rAI-powered Predictive Modeling overcome common modeling challenges other solutions can't?
• What's the difference between rAI-powered Predictive Modeling next-best customer application and Resonate's Look-alike modeling? 

Model Development and Delivery

• How long does the Predictive Modeling process take once a sales agreement is signed?
• What deliverables will I receive?
• What are the input requirements for the model?
• What type of first-party CRM data should I include?
• What is the recommended refresh frequency?
• How can I interpret the model's results?

Technical FAQs

• What IDs are accepted for matching? 
• What is the minimum number of records required for matching? 
• What are the expected match rates? 
• How can I send and receive data?
• How do I know if my model is effective? 
• Why are my Estimated Targetable IDs so high compared to my matched records for Next Best Customer Models? 
• Why are my Estimated Targetable IDs so low compared to my matched records for Churn Models? 

General FAQs

What is rAI-powered Predictive Modeling?

rAI-powered Predictive Modeling tells you who your next-best customers are and who will churn—so you can act on who matters most and improve LTV more effectively than ever.

Resonate offers the following applications:

How does Resonate rAI-powered Predictive Modeling overcome common modeling challenges other solutions can't?

Ninety-five percent of marketers use predictive modeling—but few are getting the results they need. What's needed isn't just prediction—it's precision, freshness, and the ability to take immediate action.

Resonate's solution offers several advantages over alternative methods:

• High-quality data: Models reliant on first-party data alone offer a partial view of consumers limited to who they already know, on their own properties and channels. This is a narrow and partial view that leads to partial results. Resonate combines first-party data with recent behavioral data to show what consumers are doing today, with behavior beyond your brand—tapping into the earliest intent, lifestyle shifts and psychographic cues most other models miss. 
• Models built in days: Instead of custom models that take months to build—we deliver initial models within five days.
• Precise models that perform in the real-world: Instead of overfit or under fit models, we used advanced modeling techniques and tune to your businesses' data to ensure ROI-boosting predictions you can trust. 
• Move with the market: We use recent data and regularly update scores to keep models reflective of today's market dynamics. 
• Connect predictions to activation: Don't wait months for disconnected insights. Achieve improved performance this week with fast targets that easily deploy to your campaign workflows so you never miss the moment to convert. 

Now you can reach more of the right customers at scale with custom modeling without the cost or effort. 

What's the difference between rAI-powered Predictive Modeling next-best customer application and Resonate's Look-alike modeling? 

Next-best isn’t the same as lookalike—it’s smarter targeting.

Look-alike Modeling helps you find more people who resemble previous customers—based on shared traits or behaviors.

Next-Best Customer modeling, on the other hand, predicts which individual people are most likely to buy soon—based on actual outcomes, not just similarities. It runs your database—or a broader audience—through a model that scores each person individually, so you can prioritize based on purchase likelihood.

Think of it as resemblance vs. readiness.

• Use Look-alike Modeling when you want to grow your audience by finding new people who look like a successful group.
• Use Next-Best Modeling when you want to focus budget and attention on those most likely to take action now.

Some brands use both—first expanding reach with lookalike audiences, then applying predictive scores to focus on the individuals within that group most likely to convert.
 

Model Development and Delivery

How long does the Predictive Modeling process take once a sales agreement is signed? 

The modeling process delivers results within 5 days of receipt of a clean file. 

What deliverables will I receive? 

You will receive two files: 

• your scored CRM file — your original first-party file with a propensity score (0–1) and decile appended to each matched record
• a results document (Word) covering our modeling methodology, your match rates, and model performance details.

What are the input requirements for the model? 

The input file should include specific data types, fields, and formatting requirements as outlined in the Data Requirements article.

The model learns from the first-party attributes attached to each record, so more variety and better recency produce stronger predictions. Include the outcome label (the column marking the behavior to predict — customers who have already churned for a churn model, or your best or converted customers for a next-best model), which is required. Add transaction history (recency of last purchase, frequency, and spend or value — the strongest predictors for both use cases), tenure and lifecycle fields (account age, first-purchase date, contract or renewal dates), product and category mix (plan or tier, what they have bought), and engagement and channel data (activity or login recency, email or site engagement, acquisition channel). For churn especially, include service and billing signals such as support tickets, escalations, payment failures, and downgrades — a drop in engagement is often the earliest churn signal. Add any demographic or firmographic fields you hold.

Two rules: keep feature columns as complete as possible (minimize nulls), and favor recent data, since fast-moving behavioral fields lose value quickly while stable fields like tenure hold. Weight a churn file toward engagement and service history; weight a next-best file toward transaction value and what defines your high-value customers.

Three operations run on different cadences. New-record scoring: records you send are scored incrementally, up to monthly. Rescoring: existing records are re-scored against refreshed behavioral data and any updated first-party data, which updates their probability scores and decile assignments — available bi-monthly, quarterly, or annually. Rescoring re-runs the existing model against new inputs; it does not retrain the model. Model rebuild: the model is retrained from scratch, annually.

We recommend rescoring quarterly and rebuilding annually. More frequent rescoring is available.

How can I interpret the model's results? 

The methodology document will guide interpreting probability scores and other metrics, and how to use the information for activation.

Technical FAQs

What IDs are accepted for matching? 

Resonate accepts hashed emails (HEMs) in three formats: HEMMD5 (MD5, 32 hexadecimal characters), HEMSHA1 (SHA-1, 40 hexadecimal characters), and HEMSHA2 (SHA-256, 64 hexadecimal characters).

Hashed values are hexadecimal only and case-insensitive. See Data Requirements article for a format example of each.

What is the minimum number of records required for matching? 

The minimum is 20,000 matched records — records that match to Resonate's identity graph — with 10–15% of the file representing the desired outcome (customers who have already churned for a churn model, or your best or converted customers for a next-best model). There is no maximum. Because match rates run 40–60% on hashed-email files, size your raw submission above the minimum to land 20,000 matched records — roughly 35,000–50,000 raw records depending on the recency and quality of your emails.

What are the expected match rates? 

Expected match rates vary depending on the quality of data provided. When providing HEMs, we generally see a 40-60% match rate.

How can I send and receive data? 

You can send and receive data in a variety of ways, our preferred is via AWS S3, but we can also facilitate SFTP, BOX, and Snowflake Direct share.

We prefer to have our customers write data to an AWS Bucket hosted by Resonate within the US-East-1 region. Resonate will deliver the final results to this bucket which customers can then ingest into their systems. (Read more about our data requirements). 

How do I know if my model is effective? 

There are a few metrics to consider when evaluating a Predictive Model:

• Precision: Precision typically ranges from 60% to 80%, depending on the type of business and the data quality. A higher precision rate is valuable because it means fewer false positives—i.e., customers who are incorrectly predicted to churn.
• Compare the Average Precision score to the Distribution Metric. Distribution is the baseline according to the file provided. The Average Precision score should be higher than the Distribution score. If these metrics are too similar the model is not providing lift over random predictions.

Why are my Estimated Targetable IDs so high compared to my matched records for Next Best Customer Models? 

• The records found in each decile have been gradient boosted or undergone a look-alike step that enables clients to find more customers in the Resonate ecosystem for Activation purposes.
• For the Activation step it may not be feasible to retarget all of the IDs that are available. Customers can request a specific subset or percentage of IDs from a decile or combination of deciles for retargeting.

Why are my Estimated Targetable IDs so low compared to my matched records for Churn Models? 

• For churn the baseline for estimated targetable IDs is limited to the matched records. There is no gradient boosting or look-alike function for the deciles in churn models.
• The estimated targetable IDs found in each decile for churn exclude the records where the client file indicates the customer has already churned. Therefore, if you desire to push a decile or combination of deciles to an Activation endpoint the audience will be created for IDs that show the highest probability for churn.