第 13 / 31 页
非代码面试题
显示 20 / 611 道匹配题目
答题状态：未尝试未正确已正确
ID题目领域难度题型进度权限
2463Reusing the Test Set After DebuggingA model is evaluated on test, a bug is found, the code is fixed, and the same test set is used again to verify the fix and choose among two corrected versions. Why is that second use no longer a clean test?机器学习中等essay未尝试面试订阅2464No Test Labels Touched Is Not EnoughSomeone argues there was no leakage because the code never accessed test labels. Give the core reason this defense can fail in real ML pipelines.机器学习困难essay未尝试面试订阅2465Why Nested Validation ExistsIf the same validation set is repeatedly used for model family choice, feature engineering, and threshold tuning, why is a second outer holdout or nested procedure conceptually necessary?机器学习困难essay未尝试面试订阅2466What to Audit in a Leakage ReviewYou are auditing a pipeline for leakage. Beyond checking the split line in the final dataframe, what is the highest-value thing to inspect in the code path?机器学习简单essay未尝试免费2467Unsupervised Preprocessing Can Still Distort EvaluationWhy can fitting an unsupervised step like PCA or quantile normalization on all rows still make the final reported test error too optimistic?机器学习简单essay未尝试免费2468Group Leakage Inflates Confidence TooWhy does entity overlap across train and test typically make confidence intervals and model-stability assessments look better than they really are?机器学习中等essay未尝试面试订阅2469Why Point-in-Time Feature Stores MatterA team says they can avoid leakage by using the latest vendor table everywhere because the values are more accurate. What core point about deployment reality are they missing?机器学习中等essay未尝试免费2470Rare Category Thresholding After Seeing Test CompositionSuppose you choose the minimum frequency for keeping a category only after inspecting how many rare categories appear in the test set. Why is that already a contaminated design choice?机器学习困难essay未尝试面试订阅2649How Many Expanding-Window Folds Fit in a Monthly Panel?You have 60 months of data. Each expanding-window fold uses 24 months for training, the next 6 months for validation, and then advances by 6 months. How many validation folds fit?机器学习中等数值题未尝试免费2650Usable Training Days After Purging and EmbargoA 100-day event-study sample uses a contiguous 20-day validation block in the middle. Labels look ahead 5 days, and you impose a 2-day embargo on each side of the validation block. How many days remain usable for training?机器学习困难数值题未尝试面试订阅2651Why Repeated CV Replicates Are Not Independent ExperimentsWhy should a practitioner be cautious about treating 20 repeated CV scores as if they were 20 independent experiments?机器学习简单essay未尝试免费2653Training Observations Per Fold in Grouped Cross-ValidationThere are 12 issuers and each issuer contributes 5 observations. In 3-fold grouped cross-validation, one fold holds out 4 issuers at a time. How many observations are used for training in each fold?机器学习中等数值题未尝试面试订阅2656Why Random Row CV Breaks With Overlapping Label HorizonsWhy can ordinary random row cross-validation severely overstate performance when each label depends on the next 5 trading days and adjacent rows overlap in those horizons?机器学习简单essay未尝试免费2657Why Fold-to-Fold Variation Is Not a Standard Error of DeploymentWhy is the standard deviation of fold scores not automatically the standard error of future production performance?机器学习简单essay未尝试免费2658Repeated Fold Inspection as a Tuning ChannelA researcher keeps examining which dates underperform in each fold and then adjusts features accordingly. Why is this still overfitting even if no formal hyperparameter optimizer is used?机器学习中等essay未尝试面试订阅2662Why Overlapping Validation Windows Complicate Score AggregationWhy should a practitioner be careful when averaging performance over overlapping validation windows?机器学习中等essay未尝试面试订阅2667Why Class Stratification Is Not Enough for Repeated EntitiesWhy can class-stratified cross-validation still fail badly when the same issuer appears many times and issuer identity carries predictive information?机器学习简单essay未尝试免费2668Why Embargo Matters Even With Backward-Looking FeaturesSuppose features use only past prices, yet labels depend on future returns over an event window. Why can an embargo still be necessary around the validation block?机器学习中等essay未尝试面试订阅2673Average Edge Across Opposing RegimesA signal earns +6 bps on 70% of days in calm markets and -10 bps on 30% of days in stressed markets. What is its unconditional average daily edge in bps?机器学习中等数值题未尝试面试订阅2674Precision of a Rare Alpha Event DetectorOnly 2% of days contain a true dislocation worth trading. A classifier catches 65% of those days but fires falsely on 4% of normal days. What is the precision of a positive alert?机器学习中等derivation未尝试面试订阅