Propensity modeling: predict your donors' next move – Avid

This article explains propensity modeling in Avid, predicting which current or lapsed donors are likely to take specific actions. It details creating models from seed segments with transaction history above 1,000, managing audience depth, refreshing models, and interpreting model metrics like Ranking and Model Scores. Propensity models prioritize existing donors, unlike lookalike models for new prospects. Use View Model Details to assess model reliability and adjust outreach lists accordingly.

Use this article when

You want to prioritize which current or lapsed donors to ask next, rather than find brand-new prospects.
You are setting up a propensity model and need to know what your seed segment must include. A seed segment is the audience that defines the behavior you are modeling.
You want to know who a propensity audience includes, and why the Contacts count can exceed account depth.
You need to open View Model Details to read Ranking, This list, Model scores, and What drove the score.
You need to change list size with Update depth, or rebuild scores with Refresh model.
You need to compare two models that share a requested size. That size is not always the same depth.

What is propensity modeling?

Propensity modeling predicts which donors are most likely to take a specific action in the near future.

Avid looks at what past donors looked like right before they took an action, such as starting recurring giving. It then finds donors in your file who currently look similar.

For agency partners: technical definition

A propensity model is a supervised classification model. It is trained on labeled outcomes: donors who did vs. did not take the action. It produces a probability score for each donor.

Avid uses point-in-time feature engineering to avoid leakage. Training features are built from donor behavior before the qualifying action. Scoring features are built from each donor's current profile.

When holdout metrics exist, Ranking and Model scores use the holdout split. Holdout metrics come from data the model did not train on.

If they do not exist, Avid shows in-sample numbers and a warning to refresh. In-sample metrics include training data.

How your seed segment and scoring audience work

Two things matter: what trains the model, and who ends up in the resulting audience.

Your seed segment's filters directly define who trains the model. Whatever your segment's filters currently return is the seed. Avid does not add or remove anyone from it.
Avid then scores everyone else in the same constituent data blend your seed segment is built on. That is every other donor in that data source. Avid excludes anyone already in your seed segment and anyone marked as deceased. The resulting audience never includes people who are already in your seed. Avid does not limit scoring to non-donors or any other subset.
The result is a ranked list of accounts. Avid ranks that scoring pool by predicted likelihood. It includes the highest-scoring accounts up to the current account depth. At create, that depth is the Target Audience Size you entered. Later, you change depth with Audience depth (accounts) and Update depth.
Avid only learns from information that would have been known before the behavior happened. It is not trained on hindsight. The giving signals used in training come from before the seed action.

The propensity audience never contains members of your original seed segment. Deceased donors are excluded from scoring.

Account depth follows Target Audience Size at create, or Audience depth (accounts) after Update depth. The Contacts column can be higher than that account count. One account can have more than one contact.

Propensity vs. lookalike: quick difference

Propensity models and lookalike models both use statistical matching, but they solve different problems.

	Lookalike model	Propensity model
Primary goal	Find new people	Prioritize existing donors
Best for	Acquisition	Retention, upgrades, cultivation
Population	Prospects / non-donors	Current and lapsed donors in your CRM (customer relationship management system)
Output	New audience expansion	Ranked list of who to ask now

Examples of when to use propensity (and why)

Recurring donor conversion. Identify one-time donors who resemble past donors right before they became recurring donors.
Appeal theme affinity. Find donors likely to respond to a specific appeal theme based on who responded to similar themes before. An appeal is how you ask for support.
Next best ask. Rank donors by likelihood of responding to your next planned ask, so you can prioritize outreach.
Lapsed donor re-engagement. Find lapsed donors who look like donors who successfully came back in the past.
Channel optimization. Predict which donors are most likely to respond through a specific channel, based on past channel-response patterns.

When not to use propensity

You want to acquire new donors who have never given before. Use a lookalike model instead.
Your seed segment does not have transaction history to learn from.
You need a one-time list based only on current attributes, not predicted future behavior.
Your audience is too small for the model to find a reliable pattern.

Applies to and prerequisites

Item	What you need
Plan	Avid Intelligence or the Avid Fundraising Operating System (FOS). Your plan must support propensity models.
Permission	Pathways edit permission to create a model or select Refresh model. You can open View Model Details by selecting the audience name on any propensity row. The Action menu needs Pathways edit permission, Direct Mail edit permission, or dashboard pin permission. Exporting requires a File Download connection and permission to export an audience to a file.
Seed segment	The segment must be one you own (not a modeled or suggested segment). It must be a constituent data blend audience, include a transaction filter, and have a size above `1,000`. See Owned vs. modeled audiences in Avid.
Time	Create and Refresh model take several minutes. You get a notification when the job finishes.

Eligibility requirements for your seed segment

Before you can create a propensity model from a segment, that segment must meet all three of the following requirements. If it does not, the Create a Propensity Model option is disabled and shows a tooltip explaining why.

Your segment's source must be a constituent data blend audience. If it is not, you see: Propensity Models require a constituent data blend audience.
Your segment's filter must include at least one column from transactions. If it does not, you see: Propensity Models require transactions to be a part of the filter.
Your segment's current audience size must be over 1,000. Avid requires an audience size above 1,000 to use this option. If your segment is smaller, you see: Propensity Models require an audience size above 1,000. Broaden your segment's filters to grow the audience before you create a model from it.

Steps: create a new propensity model

Go to Pathways > Manage Audiences.
Find or create the seed segment that defines the target action. Example: donors who started recurring in the last 24 months. Confirm it meets all three eligibility requirements above.
In that segment's row, open the Action menu and select Create a Propensity Model.

You see the Create a Propensity Model modal. The buttons are Cancel and Create. This modal does not refresh an existing model.
Enter a value in Target Audience Size. The placeholder is Enter a target account count...

You see: This is a count of accounts, ranked by likelihood. This can be adjusted later.

The field is required. The minimum is 100. Avid pre-fills the seed segment's current size, or 1,000 if that size is missing. It never goes below 100.
Select Create.

You see: We're building your audience. This will take several minutes. You'll be notified on completion!
When processing finishes, Avid creates a new audience named Propensity Model of {Seed Segment Name}. For example, a seed named Active Donors produces Propensity Model of Active Donors.

The new row shows a Propensity type badge. The audience never includes anyone from your original seed segment. Account depth matches the Target Audience Size you entered. The Contacts column can be higher than that account count.
Selecting the output name opens View Model Details, not the regular audience editor.

Manage Audiences columns are Audience, Type, Contacts, Data Source, Usage, Updated By, Updated Date or Archived Date, and Action. Individual scores are not a column on this table.

The Contacts info tooltip is: The actual number of records synced to the destination will vary based what that system uses to determine unique records.

On a propensity row, the Contacts cell tooltip is: Model depth is {N} accounts. Contact count can be higher when an account has multiple contacts. If the model has no stored depth yet, it is: Model depth is a count of accounts. Contact count can be higher when an account has multiple contacts.

Steps: open View Model Details

Go to Pathways > Manage Audiences.
Use the Audience Type filter to find propensity rows. The placeholder is Select audience types...
Find the propensity audience. The Type badge is Propensity.
Select the audience name. Or open the Action menu and select View Model Details.

You can also select the propensity icon in the Usage column. The tooltip is Propensity model.

While the modal loads, you see: Loading model details…

The modal title is the output audience name. The subtitle is Seeded from plus the seed name. Status can be Ready, Scoring, Training, Failed, or Unknown.

You see: This ranking predicts who is most likely to take the same action as your seed audience. Model scores show how reliable that ranking is, the factors show which giving behaviors mattered most, and you can change how many people to include without rebuilding the model.

A Learn more in Avid's Help Center. link points to this article.

How to read View Model Details

Use this modal to judge whether the list is worth mailing or calling. Read which giving behaviors mattered, then decide how deep to go. The modal does not show a score for each person. Person-level scores are available later as the export column Propensity Score.

The modal includes Ranking, This list, Model scores, What drove the score, Build facts, and How deep to go. The footer has Refresh model and, when a File Download connection exists, Export the audience.

Ranking

Ranking is the first number to read. It is the model's fit number (also called a c-stat), shown to two decimals.

Use it to decide whether the list is reliable enough for outreach. The product shows one number and a Strong or Needs caution badge. It does not show a four-row meaning table on the screen.

The Ranking help text is: Use this to decide if the list is worth mailing or calling. 0.50 is a coin flip. 0.70 or higher means people at the top are much more likely to respond than the rest of the file.

Use this table to interpret the number. This table is help text, not a screenshot of the product.

Ranking	What it means
`0.50`	A coin flip. No better than chance.
Below `0.60`	Close to chance. This list may not outperform a simple audience. The badge is Needs caution.
`0.60` to `0.69`	Aim for `0.70` or higher before you lean on this list. The badge is Needs caution.
`0.70` or higher	People at the top are much more likely to respond. The badge is Strong. The list is worth using for outreach.

If ranking is below 0.60, or the seed is thinner than Avid typically wants, a banner can open with Use this ranking with caution. The follow-on line is one of:

Use with caution. The seed audience is thinner than we typically want for a stable model.
Use with caution. Ranking quality below 0.60 is close to chance, so this list may not outperform a simple audience.

Between 0.60 and 0.69, the badge is still Needs caution. That ranking banner does not appear.

This list

This list shows how deep the current cut goes. It shows Top {x}% when a depth quality percent exists. Otherwise it shows {N} accounts.

When depth quality exists, the badge is one of: Highest-likelihood group, Past the best group, or Too deep. Older locked models may show only the account count.

A Strong ranking with a Too deep cut is still a weak campaign list. Tighten depth before you mail or call.

Model scores

Model scores shows Precision, Recall, Accuracy, and F1 score as percents. Use these as campaign-language checks, not as the reason to pick a list size.

Score	Help text
Precision	`Of the names at the top of this list, how many already did the action you care about. Higher means fewer wasted names if you only work the top of the file.`
Recall	`How many of the people who already did that action this ranking still finds. Higher means you can go deeper in the list without leaving as many good prospects out.`
Accuracy	`How often the ranking correctly sorts likely people from everyone else. Use it as a quick quality check, not as the reason to pick a list size.`
F1 score	`A balance of finding the right people and not wasting names. High means you can work the list without leaving too many good prospects out or including too many unlikely ones.`

If the numbers include training data, you see: These numbers include training data and may look better than live performance. Refresh the model to get holdout metrics.

An Analysis Audience compares giving traits or campaign results in Insights. It does not replace Ranking or Model scores. To build one, see How to Create an Analysis Audience.

What drove the score

What drove the score shows each giving signal's share of the ranking. These shares are for the model as a whole. They are not a score for each person.

The intro is: Each percentage is this signal's share of the ranking. Higher means that giving behavior influenced who scored well more than the others.

Avid shows the top signals with a share above zero. Shares under 1% display as <1%.

Examples of signal labels include:

Days since first gift
First gift channel
Trailing 12-month lifecycle
Contactability
Account type
Days since last gift
Lifetime gift count
Lifetime giving
Gifts in last 3/6/12 months

This is not a complete on-screen list.

If signals are not ready, you see: Feature importance is not available yet.

Build facts

Build facts shows:

Last trained: date as Mon D, YYYY
Requested audience size: N accounts
Current membership: N accounts / M contacts
Seed size: N

You see: People already in the seed are excluded from this list. Contact count can be higher when an account has multiple contacts.

How deep to go

Use How deep to go to change how many already-scored accounts are in the audience. This does not retrain the model.

If depth and scores are locked, you see: Refresh model to unlock depth and scores. Older models stay locked until you select Refresh model.

When unlocked, the control is Audience depth (accounts). The minimum is 100. You cannot set depth higher than the number of already-scored accounts.

You see: Top {N} of {M} scored accounts ({P}%). Avid then adds one of these hints:

Still in the highest-likelihood group.
Past the highest-likelihood group.
Too deep — this list will be less effective.

You may also see a warning banner:

The ranking is strong, but this audience is too deep.
The ranking is usable, but this audience is too deep.
Ranking is usable, but this list is past the highest-likelihood group.

Those depth banners also include Ranking quality is high enough to use. The linked action is Tighten depth below.

LockStep insight: Use the list when Ranking is Strong and This list is Highest-likelihood group. If ranking needs caution, do not lean on the list for outreach. If ranking is strong but the cut is Too deep, select Tighten depth below and then Update depth.

Update depth versus Refresh model

After you create a model, you manage list size in View Model Details. Create still uses Target Audience Size. Later depth changes use Audience depth (accounts) and Update depth.

Action	Where it lives	What it does
Update depth	How deep to go, then Audience depth (accounts)	Does not retrain. Changes how many of the already-scored accounts are included. Minimum is `100`.
Refresh model	Modal footer only	Rebuilds scores from the latest giving. Uses the current requested audience size. The footer has no separate size field. Use this for new predictions. Also use it to unlock holdout metrics and full-universe scores on an older model.
Export the audience	Modal footer, when a File Download connection exists	Downloads the current list. Map the Propensity Score column if you need person-level scores.

Update depth is available only after you change the value, and only when depth is unlocked. Refresh model is in the modal footer, not in the row Action menu.

Steps: change audience depth

Open View Model Details for the propensity audience.
Go to How deep to go.
Change Audience depth (accounts). The minimum is 100. You cannot set depth higher than the number of already-scored accounts.
Select Update depth.

You see: Audience depth updated. Contact counts will refresh shortly. Membership changes without a full retrain.

If the update fails, you may see: Unable to update audience depth.

Steps: refresh an existing model

Refreshing fully retrains the model on your current data. It is not the same as Update depth.

Go to Pathways > Manage Audiences.
Find the propensity audience and open View Model Details.
Select Refresh model in the modal footer.

You see: We're refreshing your audience. This will take several minutes. You'll be notified on completion!
Wait for processing. After an older model refreshes, you get holdout metrics, an unlocked depth control, and scores for the full universe.

Compare two models

Open View Model Details on each model. Compare Ranking, Model scores, and What drove the score. Also compare how deep the current cut is on This list (Top {x}%).

The same requested audience size is not the same depth if the scored universes differ. Read This list and the How deep to go helper, not only the requested size.

Do not treat two models' Propensity Score values as one shared scale. Each score is the chance of resembling that model's seed, not a shared likelihood of giving.

You can also select 2 or 3 rows and select Compare audiences. The button appears when you select at least one row.

If Compare audiences is disabled, the tooltip explains why:

Select 2 or 3 audiences to compare.
Compare is available for Owned and Propensity audiences only.
Compare is available for constituent data blends only.
Select audiences from the same constituent data blend.
Compare is not available for archived audiences.
Archived audiences cannot be compared.

Export and use the audience

To export, use Export the Audience in the row Action menu, or Export the audience in the modal footer. Exporting requires a File Download connection and permission to export an audience to a file.
When you map columns, the source column is Propensity Score in the Scores category. Use that column if you need person-level scores. View Model Details does not preview those scores.
Use a propensity audience as the target for a Playbook or Pathways push, the same way you use any other audience.
Use Refresh model in View Model Details when you want the list to reflect current donor activity.

If something is blocked

If Create a Propensity Model is grayed out, read the tooltip. Fix the matching eligibility requirement.
If depth controls are missing, you see: Refresh model to unlock depth and scores. Select Refresh model in the footer.
If metrics look perfect on an older model, refresh so Avid can show holdout metrics.
If features are empty, you see: Feature importance is not available yet.
If the modal cannot load, you may see: This audience is not a propensity model. or Unable to load propensity model details. Close it and open View Model Details again from a Propensity row.

What you will see

After create, the new row is named Propensity Model of {Seed Segment Name} with a Propensity badge.

After Update depth, membership changes and contact counts refresh shortly.

After Refresh model, wait several minutes. Then open View Model Details again to read holdout scores and an unlocked depth control.

Next steps

Review your Manage Audiences list to find a segment that meets the eligibility requirements above.
Create a propensity model. Then open View Model Details to read Ranking and This list before you mail or call.
If two models share a requested size, compare Ranking, drivers, and This list Top %. Do not compare requested size alone.
See How to Create an Analysis Audience if you want to compare giving traits or campaign results. Those audiences do not replace View Model Details metrics.
See Owned vs. modeled audiences in Avid if you need to confirm the seed is an owned audience.

Notes and warnings

Refresh model is only in the View Model Details footer. It is not on the row Action menu.

The kebab in the Action column has tooltip Actions. The column header is Action.

Related to