Microsoft Azure Data Scientist Associate (DP-100) practice exam

When data is not missing at random and missingness correlates with the target, the fact that a value is missing carries signal -- adding a missingness indicator and using a group- or model-aware imputation preserves that information while filling the gaps. Plain mean imputation collapses all missing entries to one constant, erasing the informative missingness pattern and artificially shrinking the column's variance, which biases the model. The claim that 18% missing makes bias negligible ignores the mechanism -- bias from NMAR data depends on why values are missing, not just how many. Complete-case deletion is not 'always unbiased': dropping rows whose missingness depends on the diagnosis systematically removes a non-random slice of patients and biases the sample. Imputing zero is the worst option here -- zero is a real, meaningful magnitude for a lab result, so it silently injects extreme low values and corrupts the distribution.

Forecasting must respect causality: you split chronologically and use rolling/expanding-window time-series CV so the model is always trained on the past and evaluated on the future, mirroring real deployment. A random shuffled split lets future transactions land in the training set while their own past lands in the test set, leaking future information and producing wildly optimistic metrics that collapse in production. Stratifying a random split by week does not fix the core problem -- it still mixes future and past rows within folds. The belief that shuffling 'always produces the most representative folds' is a general-purpose heuristic that is exactly wrong for temporal data. Simply increasing k to 10 folds changes nothing about temporal leakage; more random folds still interleave past and future.

Grid sampling enumerates every combination of discrete values and therefore only supports hyperparameters expressed as choice; a continuous distribution such as uniform has no finite enumerable set, so the search space is invalid for grid sampling and the correct fix is to either discretize the learning rate with choice or switch to random/Bayesian sampling. The 'infinite count but otherwise valid' option is wrong because the continuous parameter makes the configuration invalid, not merely large. Grid sampling does not consume Bayesian priors; priors are a Bayesian-sampling concept, not a grid prerequisite. An early-termination policy bounds runtime of trials, not the legality of the search space, so attaching one would not make a continuous parameter valid for grid sampling.