Prediction quality depends on the target, observation time, data-generating process, features and validation design established before modelling.
Pre-processing begins with the prediction moment
Before transforming data, define what is predicted, for whom, at what time and how far ahead. Every feature must have been available at that moment. A variable recorded after the outcome, or created by a response to it, leaks future information into the model. Leakage can make validation look excellent and deployment fail.
Represent the data-generating process
Missing values, repeated observations, delayed events and selection into the dataset have business causes. A zero may mean no demand, an unavailable product or a missing report. Customer transactions are nested within people and time. Machine readings follow operating states. Pre-processing should preserve these relationships rather than treating the table as an arbitrary matrix.
Feature engineering encodes a defensible hypothesis
Transformations can represent seasonality, exposure, rates, thresholds, interactions, lags or diminishing returns. Domain knowledge helps identify which representation could remain useful when conditions change. This does not guarantee causality. A variable may be predictively useful without being an intervention lever, and a causal claim requires a suitable design and assumptions beyond feature selection.
Fit every transformation inside validation
Imputation, scaling, encoding, feature selection and balancing must be learned from the training portion and then applied to later data. Time-based decisions require time-respecting splits. Grouped data may require customer, site or asset separation. The benchmark should reproduce the way new cases will arrive, including the missingness and category changes expected in operation.
Post-processing connects predictions to action
Calibration, thresholds, capacity and error consequences determine how a score is used. Marketways documents the target, features, transformations, validation and decision rule, then monitors drift and operating outcomes. Classification and regression provide model techniques; business understanding determines what the model is allowed to represent and how its result enters a decision.
