Machine Learning
Software that learns patterns from historical examples instead of being programmed with fixed rules.
Applied: Useful when a company has past campaign outcomes and wants to predict future customer response.
Wiki / customer prediction
A practical ML dictionary.
Using AI, Machine Learning and Deep Learning, requires getting familiar with quite a few industry specific terms which can be a daunting task.
On this site we have compiled a wiki of selected terms for people who want to apply prediction models to customers, donors, subscribers, and campaign audiences without becoming specialists first.
01
Core Ideas
The minimum vocabulary for understanding what a prediction model is doing.
Supervised Learning
A model learns from examples where the right answer is already known.
Applied: For customer prediction, the known answer might be donated, purchased, renewed, churned, or ignored.
Classification
A prediction task where the outcome belongs to a category.
Applied: Examples: will respond versus will not respond, churn risk versus no churn risk.
Regression
A prediction task where the output is a number.
Applied: Examples: expected order value, expected donation amount, or predicted lifetime value.
Probability Score
A number between 0 and 1 that estimates how likely an outcome is.
Applied: A donor with a score of 0.78 is estimated to be more likely to respond than one scored 0.21.
Label
The answer the model is trained to predict.
Applied: In a direct-mail model, the label might be whether the person donated within 30 days of receiving the letter.
02
Customer Prediction Terms
Terms that appear when models are used for targeting, retention, and campaign prioritisation.
Conversion
The action you want a person to take.
Applied: A conversion might be a donation, purchase, renewal, signup, or booked meeting.
Churn
When a customer stops buying, donating, subscribing, or engaging.
Applied: Churn models help teams decide who may need retention outreach before they leave.
Propensity Model
A model that estimates how likely a person is to do something.
Applied: Common examples include propensity to buy, donate, renew, churn, or click.
Customer Lifetime Value
The expected total value of a customer relationship over time.
Applied: CLV helps decide when a high acquisition or retention cost is still economically sensible.
Segment
A group of customers with shared traits or behaviour.
Applied: Segments are useful for messaging, but prediction scores are often better for deciding exactly who to target.
Cohort
A group defined by a shared starting point or event.
Applied: A cohort could be customers acquired in March, donors from a holiday appeal, or subscribers in their first 90 days.
RFM
A simple customer scoring method based on recency, frequency, and monetary value.
Applied: RFM is often a strong baseline before more advanced machine learning is introduced.
Uplift
The extra effect caused by sending a campaign compared with doing nothing.
Applied: A person may be likely to buy anyway; uplift asks whether the campaign actually changes the outcome.
03
Data And Features
The raw material of prediction work: what the model sees and what can go wrong.
Feature
An input variable used by a model.
Applied: Features can include days since last purchase, number of past donations, email engagement, geography, or channel history.
Feature Engineering
Turning raw data into useful model inputs.
Applied: For example, converting transaction history into recency, frequency, total value, and trend features.
Training Set
The historical examples used to fit the model.
Applied: A campaign response model might train on previous sends and their outcomes.
Holdout Set
Data kept separate so you can test the model on examples it has not learned from.
Applied: A holdout set gives a more honest estimate of how the model may perform on the next campaign.
Data Leakage
When the model is given information that would not be available at prediction time.
Applied: Including a post-campaign donation date while training a pre-campaign model would make performance look falsely strong.
Missing Data
Values that are absent, incomplete, or unknown.
Applied: Missing values can be meaningful, such as a customer with no recent email opens or no recorded purchase history.
Class Imbalance
When one outcome is much rarer than the other.
Applied: If only 2 percent of recipients donate, accuracy alone can be misleading because almost everyone is a non-donor.
04
Models
Common model families used in practical customer prediction systems.
Logistic Regression
A simple, interpretable model for estimating the probability of a yes-or-no outcome.
Applied: Often a good baseline for churn, conversion, and response prediction.
Decision Tree
A model that splits data through a sequence of yes-or-no questions.
Applied: Easy to explain, but single trees can be unstable and less accurate than ensembles.
Random Forest
A model that combines many decision trees to produce a more stable prediction.
Applied: Useful for tabular customer data, especially when relationships are not purely linear.
Gradient Boosting
A method that builds many small trees, each correcting errors from the previous ones.
Applied: Often strong for campaign response, churn, and propensity models.
XGBoost
A popular gradient boosting library known for strong performance on structured data.
Applied: A practical choice when customer records are stored as rows with many behavioural features.
Calibration
How well predicted probabilities match reality.
Applied: If 1,000 people are scored near 0.30, roughly 300 should convert if the model is well calibrated.
Overfitting
When a model learns noise in the training data instead of patterns that generalise.
Applied: An overfit model can look excellent on old campaigns and disappoint on the next one.
Regularisation
Techniques that discourage a model from becoming too complex.
Applied: Regularisation helps keep predictions stable when customer data has many weak or noisy signals.
05
Evaluation
How to judge whether a model is good enough to guide spend.
Confusion Matrix
A table showing correct and incorrect predictions across yes and no outcomes.
Applied: It helps separate true responders, missed responders, wasted sends, and correctly skipped contacts.
Precision
Of the people the model selected, how many actually converted.
Applied: High precision matters when every mailed piece, ad impression, or call is expensive.
Recall
Of all the people who converted, how many the model found.
Applied: High recall matters when missing good prospects is more costly than contacting extra people.
ROC-AUC
A score for how well a model ranks positives above negatives across thresholds.
Applied: Useful for comparing models, but it does not directly tell you what campaign volume to send.
Lift
How much better a selected group performs compared with the average audience.
Applied: If the top 20 percent has 3x lift, it converts three times better than the campaign average.
Threshold
The score cutoff used to decide who receives an action.
Applied: A higher threshold reduces volume and waste; a lower threshold reaches more possible responders.
Cross-Validation
Testing a model across multiple train-test splits.
Applied: It gives a more robust view of performance when there are not many historical campaigns.
06
Operations And Trust
Terms that matter once a model influences real customer decisions.
Explainability
The ability to understand why a model made a prediction.
Applied: Important when teams need confidence before changing targeting rules or campaign budgets.
SHAP
A technique for estimating how much each feature contributed to a prediction.
Applied: SHAP can show whether recency, donation frequency, channel, or value history drove a score.
Model Drift
When patterns change after the model is trained.
Applied: A donor model trained before a major event may need review if behaviour changes afterwards.
Monitoring
Tracking model inputs, scores, and outcomes after deployment.
Applied: Monitoring catches broken data feeds, changing audience quality, and falling campaign performance.
Retraining
Updating a model with newer data.
Applied: Retraining keeps scores aligned with current customers, offers, seasons, and channels.
Pseudonymisation
Replacing direct identifiers with artificial IDs while keeping records linkable for analysis.
Applied: Useful for privacy-conscious modelling where names and raw contact details are not needed for training.
Consent
Permission or legal basis for using data for a purpose.
Applied: Prediction projects should confirm that the data use fits the organisation, region, and customer relationship.
Use the dictionary as a project checklist.
A useful prediction project should be clear about the label, available features, leakage risks, evaluation metric, threshold, monitoring plan, and privacy basis before the model is treated as a business tool.