Propensity predictions enable you to predict the likelihood of virtually any action, such as lead conversion or customer churn.

In Outcomes, you'll describe the business goals you'd like to predict. Let's say a major business goal of yours is that you want to acquire more high-quality leads–you'll plug the outcome's cohorts in, and Faraday will create a likely-to-buy predictive model for that goal by applying dozens of strategies to build a range of candidate models, then selecting the one that most accurately predicts your outcome.

Getting started

Inside Outcomes, you'll find a list of your current outcomes if you have any, with columns for:

  • Performance: the performance score of the predictive outcome.
  • Eligiblity: the cohort chosen for who you want to be able to achieve this outcome.
  • Attainment: the cohort selected as attainment represents the "finish line" for the prediction. What cohort do you want the eligibility cohort to look like?
  • Attrition (optional): the cohort people will enter if they fail to enter the attainment cohort by achieving the outcome.
  • Status: whether the outcome is ready, queued, or errored.

Creating an outcome

  1. Select new outcome in the upper right of the Outcomes list view.

Screenshot of the Outcomes list view

  1. Next, select an eligibility cohort. The eligibility cohort is who you want to be able to achieve this outcome. An important note on eligibility cohorts, however, is that it should not be a subset of your attainment cohort: if your attainment cohort is set to Customers, your eligibility cohort should not be something like Customers with a basement.

Screenshot of a new outcome

  1. Next, select your attainment cohort. This cohort represents the "finish line" for your prediction. If you're predicting which leads are most likely to become customers, your attainment cohort

  2. Optionally, select an attrition cohort for users that fail to attain this outcome.

📘Example: creating a lead conversion outcome

  1. Optionally, select certain traits to block in this outcome. For example, you may want to ensure protected classes aren't used.

  2. Once your cohorts are selected, give your outcome a unique name.

  3. With your desired fields are filled out, click save outcome to save the outcome, after which you'll receive a popup telling you that your outcome is building. You'll receive an email when the outcome is ready for use, and its status in the Outcomes list view will display as Ready.

Analyzing an outcome

Once your outcome is complete and its status is ready, various features to analyze will populate in the outcome. This includes the performance of the model (through the score, lift table, lift curve, and ROC), or what kind of results you can expect when using this outcome, as well as the data features that were most important during the predictive model's build. Dynamic prediction indicates that each section will include breakdowns based on how long the individuals in the outcome were in the cohort in question.

📘What is lift?

Outcome performance

Each outcome's default tab is performance. This tab provides a chart that includes a score based on its ability to predict the outcome versus predicted lift. The closer the performance score to the upper right, the better the performance you can expect in a real-world use case. The available scores can be seen below.

  • Misconfigured: This can happen when your cohorts have too few people in them to make meaningful predictions.
  • Low: Some lift might be generated, but since signal strength is low, performance may not be consistent. This could mean that we can't easily distinguish between positive and negative examples due to them being too similar (e.g. very similar age, gender, etc), or being too dissimilar (within a given cohort).
  • Fair: Lift should be generated, and this is definitely better than random. For example, if your average lead conversion rate is 3%, you can likely expect a pipeline using this outcome to produce a higher conversion rate.
  • Good: Lift should be generated with some consistency.
  • Great: Lift should be generated reliably.
  • Excellent: Strong lift should be generated reliably.
  • Warning: Your outcome is predicted better than we would typically expect. You should check that only predictors known prior to the outcome are included. In other words, the model's performance is too good to be true. This can happen when the model calls on first-party data that's directly related to the outcome. If your model reports this, ask support for assistance.

📘Getting the most out of an outcome score

Additionally, each outcome includes other performance metrics through which you can assess how an outcome is performing. These metrics are displayed in the Faraday app as well as the full technical report, which is available via the three dots in the upper right, then the full technical report button.

Expand the tabs below to see info specific to each section of outcome performance.

Outcome performance: metrics

Screenshot of an outcome's metrics tab

The metrics tab includes various metrics to help assess model performance, including ROC AUC value, average conversion rate, and average lift over random.

  • ROC AUC: ROC AUC is a score, on a scale of 0.0 to 1.0, that helps determine how well a model is performing. An ROC AUC of less than 0.5 indicates a configuration error. An ROC AUC of 0.5 indicates that the results are no better than a random guess, which means that the model is performing very poorly and is very unlikely to properly identify future conversions. On the other end, an ROC AUC of 1.0 indicates that the model is likely too good to be true, thus an ideal ROC AUC should be in somewhere in the middle.
  • Average conversion rate: Average conversion rate is the number of individuals that were eligible and became attainers, divided by the number of individuals in your eligibility cohort. This value, when used with values from the lift table can help you calculate the conversion rate you might expect per decile when using these predictions.
  • Average lift over random: Similar to ROC AUC, the value in average lift over random aims to quantify the benefit of using this outcome as opposed to not using predictions at all within a single number, therefore higher numbers often indicate better performance.

📘How to calculate predicted conversion rate

The average conversion rate (ACR) is equivalent to not using predictions at all, and lift * average conversion rate is equivalent to Faraday's added value. As an example, using the above formula, if you have an average conversion rate of 10%, and your top 20% of scorers have a lift of 5.0 per the lift table in an outcome, the formula would look like 10 + 5 * 10, resulting in a 60% predicted conversion rate for your top 20% of scorers in this outcome.

Outcome performance: lift curve

Screenshot of an outcome's lift curve tab

The lift curve measures the estimated lift within each target percentile. The X-axis, target percentile, indicates the fraction of the people that have moved from the eligiblity to attainment cohort that is being captured. A well-performing lift curve will generally have a strong "bump" up the Y-axis that represents the highest-scoring individuals in the outcome (the left-hand side of the X-axis).

Outcome performance: lift table

Screenshot of an outcome's lift table tab

The lift table provides an alternate view of the lift curve, in which each decile of scorers is given its own row. Each row includes:

  • The specific decile (e.g. top 10% of scorers).
  • The probability threshold, which can be used to customize deployment filtering in a pipeline.
  • The number of positive examples, or the individuals who have successfully completed the journey from the eligiblity to attainment cohorts.
  • The number of negative examples, or those still in the eligibility cohort who have yet to complete the journey to attainment.
  • The precision, which is a way to quantify how good the model is at putting the positive examples in the given decile at the highest ranks of scoring.
  • The lift, which is an indicator of how well the model is predicting positive examples over a random choice from the eligility cohort.
Outcome performance: ROC

Screenshot of an outcome's ROC tab

The ROC curve measures how well a model is performing. A well-performing will curve toward the Y-axis of true positive rate (TPR) before leveling out at 100%, as the TPR tells you how well the model is predicting true positives, or individuals from the eligibility moving to the attainment cohort. The dotted line represents a model that is as good as a random guess, or a coin-flip.

Outcome data

The data section of an outcome breaks down the top features or traits used in creating the outcome, ordered by their level of impact. These can include features from both your first-party data (indicated by a sparkle) and the Faraday Identity Graph. Each feature that had an impact on the outcome will be assigned a percentage based on its level of impact.

Screenshot of an outcome's features of importance

By expanding a feature, you'll see its directionality, or the impact the feature has on the score of individuals in the eligiblity cohort. Each feature's directionality is broken down into different bins based on how the feature in question is broken down. For example, the age feature will include bins for people age 18-31, 31-37, etc.

Screenshot of an outcome's features directionality

In the above example, if an individual is 45 years old, the directionality of their corresponding age bin is -17, so–for a population of 100 people–this individual would rank on average 17 ranks lower than the average scored individual. This is a rather negative impact on their predictive score for the outcome.


The data section also includes any exceptions that are selected when creating or editing an outcome. Exceptions allow you to prevent the outcome from using certain traits while building the outcome, and can be made for both Faraday traits, and first-party traits defined from your data. You can add exceptions for all protected class traits via the checkbox, or add individual exceptions via the add trait button.

Screenshot of an outcome's exceptions

This can be useful for excluding protected classes, such as age and gender, from your predictions so that you can remain compliant to relevant laws & regulations.

Outcome bias

Each outcome you create includes a section that breaks down any bias that Faraday detects in your predictions and the data that was used for them, including a summary tab for an at-a-glance overview.

Screenshot of an outcome's bias summary

Bias reporting is broken down into four categories: data, power, predictions, and fairness.

  • Data: The underlying data used to build an outcome can introduce bias by unevenly representing subpopulations. This bias is measured by comparing distributions of sensitive dimensions across labels. Categorical distributions (e.g. for gender) are compared using proportions. Numeric distributions (e.g. for age) are compared using a normalized Wasserstein distance on the space of empirical distributions.

  • Power: A subpopulation is a subset of the eligible population defined by a set of sensitive dimensions (e.g. age and gender) and values (e.g. adult and female).

    Outcome performance can be measured for a subpopulation and compared to the overall performance on the entire population.

  • Predictions: Outcome predictions can target subpopulations with bias. Measuring that targeting discrepancy fall under this heading.

  • Fairness: Fairness metrics aim to give an overall picture of how a subpopulation is treated. There are many metrics in the literature and the appropriate metrics depend on the specific situation.

By clicking details in the summary tab next to a category–or the tab itself–you can see the level of bias that was detected for that category.

Screenshot of an outcome's bias reporting on predictions

In the above example, Faraday found that this lead conversion outcome had a strong favorable bias (indicated by the green arrow) toward people in the sensitive dimensions breakdown of senior/male. Expand the table below to see how each sensitive dimension breaks down into its subgroups.

Sensitive dimensions breakdown
Sensitive dimensionSubgroupValue
Ageyears old
Young Adult21-30
Middle Age41-60

Mitigating bias

When you find that an outcome of yours has bias being reported that you'd like to address, you can apply bias mitigation strategies to mitigate or reverse the bias.

Currently available bias mitigation strategies:

  • None: Ignore bias
  • Equality: Neutralize bias.
    • As an example, if you have 48% men in the outcome's eligiblity cohort, and you mitigate gender using equality, then any output of a pipeline using the mitigated outcome will have 48% men and 52% women.
      • For equality, the basis is to preserve the distribution from the eligible population, which means that you don't want your outcome to create bias by ranking people of a certain subpopulation higher than another.
  • Equity: Invert bias.
    • For equity, in a perfectly fair world, each subpopulation in the outcome's eligibility cohort is of the same size (e.g. 33% Senior Male, 33% Young Female, 33% Teens with Unknown gender). If one of these subpopulations–Senior Male, for example–is 52% of the overall population in the eligiblity cohort, it's 19% too large. Because it's 19% too large, Faraday shrinks it by 19% from the ideal 33%, which is roughly 14%. This process repeats for each subpoplation.
    • At the end of this process, Faraday over-promotes rarer classes so that the target population of your predictions ends up being made of more people from the originally under-represented populations. From a business standpoint, this serves two purposes:
      • Prevents your marketing campaign from being trapped in a vicious circle (i.e. I market more to middle-aged men, therefore I sell more to middle-aged men, therefore I should market more to middle-aged men.)
      • Allows you to identify which under-marketed population has the best potential to convert.

🚧️With great power...

In general, using equality to neutralize any bias discovered is the most common use case for bias mitigation in cases where you do want to actively mitigate (see warning above). It's often helpful, if you have outcomes available, to create a new outcome when applying a bias mitigation strategy for easy comparison.

📘Bias mitigation: further reading

Deleting an outcome

To delete an outcome, click the options menu (three dots) on the far right of the outcome you'd like to delete, then click delete. If the outcome is in use by other objects in Faraday, such as a pipeline, the delete outcome popup will indicate that you need to modify those in order to delete the outcome. Once there are no other objects using this outcome, you can safely delete it.

📘Deleting resources

Screenshot of deleting an outcome