A data science team retrains the same algorithm many times, each time on a different random sample drawn from the same population. Across those runs, the model's predictions on a fixed holdout set barely change. In bias–variance terms, what does this stable behavior primarily indicate?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Here's the deal: variance asks how much your model's answers jump around when you swap the training data. Think of it like this — if three different crews each train on a slightly different slice of the same population and all three spit out almost the same scores on a fixed test set, you're looking at low variance. Consistency is the tell. The exam trap is mixing this up with bias. Bias is about being systematically wrong on average; variance is about sensitivity to which sample you drew. High variance is the wild swingy overfitter; low variance is the steady one. Your boss walks in with five new training extracts and wants to know if scores will flip — with low variance, you can say they should hold pretty still. Remember: stable across training sets = low variance. Got it? Sweet. Keep that distinction cold for the test.
Full explanation below image
Full Explanation
In the bias–variance framework, variance describes how sensitive a model's predictions are to the particular training sample used. When you retrain the same learning procedure on different draws from the same underlying population and observe nearly identical predictions on a fixed evaluation set, the estimator exhibits low variance: its outputs do not fluctuate much with sampling noise in the training data.
This notion is distinct from bias. Bias reflects systematic error — the gap between the model's expected prediction (averaged over many training sets) and the true target function. A model can be low-variance yet high-bias (for example, a simple linear model that always underfits in a similar way), or high-variance yet low-bias (a flexible model that averages near the truth but swings hard on each sample). Low variance alone therefore does not imply perfect training fit or low bias; it only asserts stability under retraining on different samples.
The incorrect options encode common confusions. Equating the scenario with high variance reverses the definition: high variance is exactly large sensitivity to the training set, often associated with overfitting and unstable generalization estimates. Interpreting stability as high bias confuses a possible co-occurrence with a necessary implication; stability does not diagnose underfitting. Reading stability as low bias or as nearly perfect training fit conflates average correctness (or training error) with cross-sample consistency.
Practically, low variance is desirable when you need reproducible predictions and reliable validation metrics across data splits. Techniques that often reduce variance include regularization, ensembling (bagging, random forests), more training data, and less aggressive model capacity relative to sample size. Trade-offs remain: overly aggressive variance reduction can increase bias. Exam questions typically probe whether you can separate "does the answer change when the training set changes?" (variance) from "is the model systematically off on average?" (bias). When predictions barely move across resamples, mark low variance and avoid smuggling bias language into the answer unless the stem also describes systematic underfitting or oversimplification.