In order to perform honest assessment on a predictive model, what is an acceptable division between training, validation, and testing data?

Question

Options

• ATraining: 50% Validation: 0% Testing: 50%
• BTraining: 70% Validation: 0% Testing: 0%
• CTraining: 0% Validation: 100% Testing: 0%
• DTraining: 50% Validation: 50% Testing: 0%

• B
  3% (1)
• C
  3% (1)
• D
  95% (36)

Explanation

Option D is correct because it separates data into a training set (used to build the model) and a validation set (used to evaluate it on unseen data), which is the minimum structure needed for honest assessment - the model never trains on the validation data, so performance measurements are unbiased.

Option A fails because having zero validation data means there is no independent set to evaluate model selection or tuning decisions, even though it has a test split. Option B is entirely unusable for assessment since 100% of data goes to training with nothing held back. Option C is the inverse problem - with no training data, no model can be built in the first place.

Memory tip: Think of it as "you can't grade your own homework" - the model must be evaluated on data it has never seen during training, which requires at least a training/validation split (D). Any option with 0% held-out data (B, C) or 0% training data (C) immediately fails this rule.

No community discussion yet for this question.