Define the task and baseline
Classification predicts a category; regression predicts a numerical value. Specify the target, scoring time, label window, metric, and operating constraints before selecting a model. A rule, historical average, or majority-class prediction can provide a useful benchmark. Then consider a simple linear or logistic model where appropriate. Its coefficients can help inspection, but correlated features, transformations, and confounding mean coefficients do not automatically establish causation.
A team predicts delivery duration in minutes. Identify the task type, propose a simple baseline, and choose a metric that reflects whether large late deliveries matter more than small errors.
Compare capacity and operational cost
Tree ensembles can model nonlinear relationships and interactions in tabular data, while neural networks and managed AutoML approaches offer other tradeoffs in capacity and development effort. No family is universally best. Compare candidates with the same data splits and metrics, and measure inference latency, training cost, maintenance, and explanation requirements. BigQuery ML supports multiple supervised model types, though feature types and regional support differ by model.
A logistic baseline meets the latency limit but misses the recall target. Plan a controlled comparison with a boosted tree using the same validation cohort, then state how you will check the serving constraint.
Protect the comparison from leakage
Use features available at the decision time, fit preprocessing on training data, and choose splits that match deployment. Preserve the final test set while selecting a model, features, and threshold on validation data. A more complex model may exploit leakage more effectively than a baseline, producing an impressive but misleading score. Track data snapshots, code, parameters, and evaluation results for each run.
A model predicts account cancellation at the beginning of a month, but a feature records the cancellation reason entered at the end of the month. Explain the leakage and redesign the feature set.
Make a release decision, not just a leaderboard
Evaluate the selected candidate on a representative holdout and important slices, then test the complete serving path. Check calibration or operating thresholds if probabilities drive decisions, as well as latency, access controls, and rollback. A model with a slightly higher aggregate score may fail a mandatory safety or operational gate. Document the tradeoff and continue monitoring once outcomes arrive.
Candidate A has lower overall error and meets all latency requirements. Candidate B has slightly lower error still, but fails a critical slice threshold. Which candidate is eligible under the stated gates, and what additional evidence would you request?
A more complex model greatly outperforms a baseline, but one feature is created after the prediction time. What is the first issue to investigate?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.