The quantity a supervised machine learning model is trained to predict: the output variable whose values in the training data serve as labels that guide the model’s learning. Choosing the target variable is the most consequential modeling decision in any supervised learning project, because it determines what the model optimizes and whether the resulting model is actually useful for the business decision it is intended to support.
Also known as dependent variable, output variable, label, response variable
In supervised learning, the target variable is the output the model learns to predict from input features. It is called the dependent variable because its values depend on (are to be predicted from) the independent input features. Training data consists of (features, label) pairs where the label is the observed value of the target variable for each example. The model learns the statistical relationship between features and labels during training, and uses that relationship to predict the target value for new examples where only the features are known.
The definition of the target variable encodes the modeling objective. A customer churn prediction model might define the target as binary: churned within 90 days (1) or not (0). A lifetime value prediction model might define the target as continuous: total revenue generated in the next 12 months. A lead quality scoring model might define the target as ordinal: high, medium, or low qualification. Each definition makes different assumptions about what matters, what information is available for labeling, and what downstream decisions the model will support. The target variable definition must be aligned with the business decision context, not just technically convenient to label.
Proxy targets, where the model predicts an observable metric that is correlated with but not identical to the true business outcome of interest, introduce risk that the model optimizes the proxy rather than the true objective. A model trained to predict email click-through rate as a proxy for purchase intent will learn to identify content that generates clicks, which may be maximally sensational or curiosity-driven rather than genuinely purchase-intent-inducing. A model trained to predict attributed conversions as a proxy for incremental sales impact will optimize toward channels and users that receive conversion credit, not toward the actions that cause conversion. Proxy target selection should explicitly acknowledge the gap between the observable proxy and the true business objective, and model predictions should be qualified accordingly when used to drive decisions.
A working ad agency that defines a target variable incorrectly, labels it noisily, or selects a proxy that diverges from the true business objective will build a model that optimizes the wrong thing. The model will pass technical validation metrics (good cross-validation AUC, acceptable test accuracy) while failing to produce the intended business outcome. This failure is particularly insidious because it is invisible in the model evaluation: a model that correctly predicts the proxy target looks good on all standard metrics even if the proxy does not actually correspond to the business goal. Target variable definition deserves as much attention as feature engineering and model selection, and arguably more.
The prediction horizon in the target variable definition determines what behavior the model captures and how actionable its predictions are in the deployment context. A churn model that defines the target as “churned within 30 days” captures only customers who are already very close to churn and may be too late to intervene effectively for most of them. A model that defines the target as “churned within 90 days” captures earlier-stage churn risk and gives the retention team more lead time. A model that defines the target as “churned within 12 months” captures very early risk signals but produces a prediction so distant from the current decision window that it may not reflect the customer’s current trajectory. The right prediction horizon depends on the lead time required for an effective intervention and the client’s capacity to act on a large predicted-at-risk population.
Label quality directly bounds model accuracy: a target variable labeled with 85% accuracy cannot produce a model that exceeds 85% accuracy on the same definition, regardless of model complexity. A lead qualification model where the “qualified” labels are assigned by a sales team that uses inconsistent criteria will have a ceiling on achievable accuracy equal to the inter-rater agreement among sales team members on qualification. Improving the model architecture or adding more training data will not overcome this ceiling. Diagnosing label quality through inter-rater agreement analysis, label audit sampling, and consistency checks before investing in modeling effort is the prerequisite for determining whether improved labeling procedures or a simplified qualification definition would more cost-effectively improve model quality than additional modeling work.
Continuous target variables typically produce more actionable models than binary targets when the business decision involves ranking or tiering rather than a hard yes/no decision. A lead quality model that predicts a continuous score (probability of converting to a sales-qualified opportunity, or predicted revenue from the lead) enables flexible threshold adjustment for different capacity constraints without retraining the model. A binary qualified/not-qualified model fixes the decision threshold at training time and must be retrained when the sales team capacity changes. For use cases where the downstream decision involves ranking, prioritization, or tiering, predicting a continuous score preserves more information and produces more flexible tooling than collapsing the prediction to a binary label.
An agency is building a subscriber retention model for a digital media client with 840,000 active subscribers. The client’s goal is to identify subscribers at risk of cancellation 60 to 90 days in advance to enable targeted retention interventions including personalized content recommendations, pause-subscription offers, and loyalty pricing. The agency initially proposes a binary churn model with the target variable defined as “subscription cancelled within 30 days.” After reviewing the client’s intervention capacity and timeline, the agency identifies two problems with this definition. First, subscribers identified as at-risk with only a 30-day horizon have often already decided to cancel; the intervention success rate on 30-day-horizon predictions is historically only 8%. Second, the retention team can proactively contact only 2,500 subscribers per week; with 840,000 subscribers, they can reach each subscriber approximately once per 67 weeks, meaning a 30-day prediction window is too narrow to systematically act on all identified at-risk subscribers before the predicted churn date. The agency redefines the target as “subscription cancelled between 60 and 150 days from prediction date,” a 90-day window starting 60 days out. This definition identifies at-risk subscribers with enough lead time for the intervention team to contact them (60 days of lead time) but not so early that the prediction is too uncertain to act on (the signal would be too noisy at 6-month-plus horizons). A second change: the target is defined as continuous (probability of cancellation in the 60 to 150 day window) rather than binary, enabling the team to rank subscribers by predicted risk and contact the highest-risk subscribers first when capacity is constrained. Model validation on a held-out 8-week period shows AUC of 0.78 and a 4.2x lift ratio (subscribers in the top 20% of predicted risk cancel at 4.2 times the base rate during the prediction window). Intervention success rate on model-identified at-risk subscribers increases to 22% with the longer lead time, compared to 8% on the original 30-day target definition.
The generative AI foundations module covers target variable definition, proxy target risks, prediction horizon selection, and label quality diagnostics, with examples from churn, lead quality, and conversion prediction models built for agency clients.