Transparent comparisons, from input to residual
A baseline that matters
Start with a mean predictor. Compare linear regression, a quadratic regression candidate, and a random forest where the data and sample size support them.
Keep the split honest
Choose random, group-aware, or ordered validation to suit the study. Preprocessing is fitted within training partitions, and cross-validation selects models without selecting on the held-out test.
Read more than a score
Review RMSE, MAE, R², cross-validation error, and observed-versus-predicted values. A favorable metric on one dataset is evidence for that evaluation, not a general validation claim.
Keep the recipe
Store selected features, target, seed, split indices, parameters, versions, and input hash with each run. Repeat the calculation from the same recorded inputs.
Predictive capability depends on sample size, data-generating conditions, and representativeness. Out-of-domain estimates and small test sets require caution. The workspace does not call a model accurate without evaluation evidence.