The process of transforming raw data into informative input features that improve machine learning model performance, including aggregating events into behavioral metrics, encoding categorical variables, constructing interaction terms, and deriving time-based signals. For agencies, feature engineering is where domain expertise about what actually drives client outcomes gets built into the model, and it is typically the step that determines whether a predictive system performs well or barely at all.
Also known as feature construction, feature extraction, feature creation
Feature engineering encompasses the full process of converting raw data into the numerical representations that machine learning models can learn from effectively. Raw data is almost never in the right form: a timestamp needs to become “days since last purchase” and “day of week”; a categorical label like “product category” needs to become a numerical encoding; a sequence of purchase events needs to become aggregate signals like “average inter-purchase interval” and “trend in purchase frequency over the last 30 days.” Every transformation requires a decision about what information to extract, how to extract it, and whether the extracted signal will generalize beyond the training set.
Interaction features capture relationships between variables that neither captures alone. A customer who views a product page and then visits the pricing page within five minutes is exhibiting different intent than one who views the same product page and then immediately leaves. Neither the product page view nor the pricing page view alone encodes this intent as well as a feature that captures the transition between them. Constructing interaction features requires understanding the behavioral process being modeled well enough to anticipate which combinations of raw signals are likely to be informative.
Deep learning partially automates feature engineering for image, text, and audio data by learning useful representations directly from raw inputs during training. This is why deep learning replaced much of the hand-crafted feature engineering that earlier computer vision and natural language processing systems required. For structured tabular data, however, automated feature learning is less effective, and thoughtful feature engineering by practitioners who understand the domain remains the primary path to high model performance. This is why gradient boosted models trained on well-engineered tabular features frequently outperform neural networks on structured marketing and customer data.
Agency AI projects involve structured client data: CRM records, email engagement logs, web behavioral events, purchase histories, and campaign performance tables. This is exactly the data type where feature engineering has the most leverage. A working ad agency building predictive models on structured client data will get substantially more performance improvement from better feature engineering than from trying different model architectures, and that leverage compounds: a well-engineered feature set supports better models, which produce better predictions, which drive better client outcomes.
Recency, frequency, and monetary features are the foundation of most customer prediction models. The RFM framework, recency of last purchase, frequency of purchases, and monetary value of purchases, encodes the behavioral signals most predictive of future customer value across industries. These are not raw data fields; they are derived features computed from transaction histories. Agencies that know how to build and tune RFM-style features, and the dozens of variations on them, have a repeatable foundation for customer prediction models across clients regardless of industry.
Campaign performance features encode what works. A creative performance prediction model trained on raw campaign settings will underperform one trained on features that encode domain knowledge: ad format type, relative headline length, emotional tone category, whether the creative includes a price, and whether the call to action is action-oriented or benefit-oriented. These features require someone who understands advertising to construct them from raw data. The model cannot construct them on its own. The agency’s domain expertise is what gets encoded in the feature engineering step, and it is what separates a generic model from one that actually reflects what drives performance in a specific client’s category.
Temporal features require special care to avoid leakage. Features derived from data that occurs after the event being predicted, such as using future purchase behavior as a feature for predicting current intent, introduce data leakage that produces artificially high training accuracy and real-world model failure. Feature engineering for time-series data requires careful attention to temporal boundaries: every feature must be computable from information available at the time the prediction would be made in production, not from information available only in hindsight.
An agency is building a reactivation propensity model for a subscription software client to identify churned users most likely to re-subscribe if contacted. The initial model uses four raw features from the CRM: days since cancellation, plan type at cancellation, tenure in months, and cancellation reason code. It achieves 61% accuracy on a held-out test set. The agency conducts a feature engineering sprint that adds 14 new features: engagement trend in the 60 days before cancellation, number of logins in the final active month, ratio of feature usage in last month to prior six-month average, number of support tickets in the final quarter, whether the user exported their data before cancelling, and interaction terms between tenure and cancellation reason. The retrained model achieves 79% accuracy on the same test set. The most informative new feature is the data export indicator: users who exported their data before cancelling are 3.4 times more likely to re-subscribe within 90 days of a targeted outreach.
The generative AI foundations module of the workshop covers the feature engineering practices that convert raw client data into the signals that make predictive models perform at production quality.