Skip to content
Tech Interview Prep
Technical interview guide

Bias-Variance Tradeoff

Why model error splits into bias and variance, and why reducing one often increases the other.

Read
24 min
Practice MCQs
25
Interview Q&A
25
Edition
v5
Editorial status
Reviewed

Scope: scikit-learn stable model-selection and ensemble guidance and Google ML Crash Course overfitting guidance accessed 2026-08-31..

Curated: · Written: · Reviewed:

25 questions

MCQ-2What does estimator variance describe?
MCQ-1A team fits the same model class to 200 bootstrap resamples of one training set. At a fixed test point x, the mean prediction across the 200 fits is 0.40 above the true value, and the predictions have a standard deviation of 0.05. Which term does the 0.40 measure?
MCQ-19A team is asked to predict the number of heads in 10 fair coin flips from features that describe the room (temperature, time of day, who is flipping) but carry no information about the flips. What is the lowest mean squared error any model can achieve on unseen data?
MCQ-4A team trains a gradient-boosted model on 180k rows and holds out 20k. At max_depth=4 the logs read training log-loss 0.21, holdout 0.24; at max_depth=16 they read training 0.04, holdout 0.31. Which pattern in these logs most strongly indicates overfitting on the holdout split?
MCQ-9A team tunes a random forest on 40,000 labelled rows with 5-fold cross-validation, sweeping max_depth over 2, 4, 8, 12 and 20 and recording mean training and validation ROC AUC: | max_depth | train AUC | validation AUC | | --- | --- | --- | | 2 | 0.72 | 0.71 | | 4 | 0.81 | 0.79 | | 8 | 0.89 | 0.77 | | 12 | 0.95 | 0.74 | | 20 | 0.999 | 0.70 | They must ship one configuration. Which choice does this curve support?
MCQ-13A team fits ridge regression on 500 training rows with 200 strongly correlated features. Sweeping λ gives training error 0.18 → 0.24 → 0.35 and validation error 0.42 → 0.31 → 0.38 at λ = 0.01, 1, 10. Which reading of the bias–variance tradeoff does this trace support?
MCQ-17A team compares a single depth-3 decision tree with 400 rounds of gradient boosting on depth-3 trees. Same 50k training rows, same 10k held-out rows, same features. Single tree: 22% training error, 24% validation error. Boosted: 3% training error, 26% validation error. Which reading of the bias-variance profile do these numbers support?
MCQ-15A bagged ensemble averages 10 trees. Each tree's prediction variance is 0.25. The pairwise correlation between tree errors is 0.6. Using the standard bagging variance approximation, what is the variance of the ensemble's averaged prediction?
MCQ-16A team bags 200 fully grown decision trees, each trained on a bootstrap sample drawn from the same 500-row, 12-feature dataset, and averages their predictions. On held-out data the ensemble's variance is barely below a single tree's. Which of these best explains the weak variance reduction?
MCQ-21A team doubles a 5,000-row training set to 10,000 rows by adding near-duplicate copies of rows it already has. The run-to-run variance of the fitted model barely drops. Why do near-duplicate rows buy so little variance reduction compared with independent rows?
MCQ-6A fraud-detection model scores 0.99 on its training split and 0.71 on a held-out split. Before funding 200k more labelled transactions, the lead asks for a plot that shows whether more data would close that gap. What belongs on the horizontal axis of that plot?
MCQ-7A team training a gradient-boosted model on 200k rows sees train log-loss 0.11 and validation log-loss 0.19. After adding 800k more rows from the same pipeline (1M total), train log-loss rises to 0.14 and validation falls to 0.16. Validation is still dropping about 0.01 per doubling of data. Which action do the numbers best support?
MCQ-8A team trains a 2-layer MLP (64 units per layer) on 1.5M rows with 25 features for binary classification; 40% of rows are positive. Training log loss falls to 0.38 by epoch 8 and is unchanged at epoch 300, and validation log loss holds at 0.39. Predicting the 0.4 prior scores 0.673 log loss, and a gradient-boosted tree on the same features reaches 0.22. A pilot that added 4M more rows moved validation loss from 0.390 to 0.388. Which change is most likely to lower validation loss?
MCQ-12A fraud model scores 0.62 AUC on training and 0.61 on a same-distribution holdout, and both learning curves flatten by 200k rows. The team can spend the quarter either labelling 1M more rows or rebuilding features. Which measurement would most justify rebuilding features?
MCQ-3A team trains a gradient-boosted tree on 500k rows with 40 features, using a random 80/20 split. Training AUC is 0.63 and held-out validation AUC is 0.61. A logistic regression on the same features and split reaches 0.79 validation AUC. Which diagnosis do these numbers support?
MCQ-14A team trains an image classifier for 8 epochs, logging mean loss on the training split and on a validation split after each epoch; a held-out test set is kept aside for the final report. The training loop uses early stopping with patience = 2, counting epochs since the best validation loss, and keeps a copy of the best weights. | epoch | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | train loss | 0.62 | 0.48 | 0.39 | 0.33 | 0.28 | 0.24 | 0.21 | 0.19 | | val loss | 0.58 | 0.47 | 0.41 | 0.40 | 0.44 | 0.51 | 0.60 | 0.71 | What does the loop do, and which weights does the pipeline ship?
MCQ-5A team ships a fraud classifier because its training accuracy (93.5%) and validation accuracy (93.2%) differ by only 0.3 points. Fraud is 6% of the 500,000 transactions. What do the two accuracy numbers establish?
MCQ-20A subscription churn model scores AUC 0.52 on a held-out 6-month window. Its features (account age, plan tier, monthly spend and support-ticket count) are all computed at the scoring date, and the holdout window has not been used for any tuning. Which change is most likely to raise the holdout AUC?
MCQ-24A fraud model was trained on Jan–Jun transactions where fraud prevalence was 3.0%. Training ROC AUC was 0.99. On a random 20% holdout it reached ROC AUC 0.94 and, at threshold 0.5, precision 0.68 and recall 0.62. The model's top feature is transaction amount. On July 1 a new upstream rule began blocking 60% of the obvious fraud attempts that the model previously scored. For the first week of July, labels are complete; the scored population's fraud prevalence is 1.2%. At the same 0.5 threshold, production precision is 0.41 and recall is 0.62. Aggregate production ROC AUC on scored transactions is 0.93. What best explains the precision drop?
MCQ-22A fraud model is trained on 12M transactions from three markets. Overall validation AUC is 0.91. A fourth market holds 18% of validation traffic but only 4% of training rows; its slice AUC is 0.72. Ten retrains with different seeds give overall AUC 0.905–0.913 (sd 0.003) and fourth-market AUC 0.71–0.73 (sd 0.006). Within the fourth market, rows whose value of the model's top-gain feature falls inside the training range score AUC 0.89; 31% of rows fall outside it. What does this evidence establish?
MCQ-18A team stacks five base models with a logistic-regression meta-model. The meta-model is trained on predictions each base model made on the rows it was fit on, and the stack scores worse on held-out data than the best single base model. Why must the meta-model be trained on out-of-fold predictions instead?
MCQ-10A team tunes 300 gradient-boosting configurations on one fixed 5-fold split, fitting the scaler and encoder inside each fold. It picks the configuration with the highest mean validation AUC (0.912) and reports that number. On a random holdout from the same pool, the same configuration scores 0.874. What accounts for the gap?
MCQ-11A team tunes a gradient-boosted model over 20 hyperparameter settings with 5-fold cross-validation on all 1,000 labeled rows, picks the setting with the best mean fold AUC (0.87), and reports 0.87 as the model's expected AUC on unseen data. No rows were held back. Which reading of that 0.87 is right?
MCQ-23A team compares a depth-3 tree with a depth-12 tree on the same 5-fold split of 10,000 rows. The depth-12 tree's min_samples_leaf was tuned on those same folds. Per-fold accuracy, depth-3: 0.80, 0.82, 0.78, 0.81, 0.79; depth-12: 0.78, 0.94, 0.77, 0.95, 0.81. The team plans to ship the depth-12 tree. What does the comparison actually support?
MCQ-25A team has finished hyperparameter search and threshold selection on cross-validation folds. They need one number for the model's expected accuracy on next quarter's traffic, and they have a single 10k-example held-out set that no model has seen. What should that set be used for?