Skip to content
Interview prep / ML/Data

ML/Data interview questions: page 2 of 3

Questions 51–100

  1. Early stopping consumes validation dataEasy
  2. A gap based only on autocorrelationHard
  3. Duplicated alternative-data recordsMedium
  4. Confidence after a short regimeMedium
  5. Validation does not match executionMedium
  6. Tree depth and sample fragmentationEasy
  7. Random forests and correlated treesMedium
  8. Boosting learning rate and tree countMedium
  9. Permutation importance with correlated factorsMedium
  10. Impurity importance favours flexible splitsMedium
  11. Trees beyond the training rangeMedium
  12. Leakage-safe lagged featuresEasy
  13. Missingness as a predictorMedium
  14. Monotonic constraintsMedium
  15. Out-of-bag error for time seriesMedium
  16. Cross-sectional ranks and tiesEasy
  17. Feature importance is not a trading ruleEasy
  18. Interactions and partial dependenceMedium
  19. Category drift in a live modelMedium
  20. A complex model loses its advantageMedium
  21. Implement ridge with an unpenalised interceptMedium
  22. Numerically stable binary log lossMedium
  23. Standardise using training statisticsEasy
  24. Stable softmax by rowEasy
  25. A mean-squared-error gradientMedium
  26. PCA projection without covariance inversionMedium
  27. Pairwise squared distancesMedium
  28. One k-means assignment and updateMedium
  29. Purging labels with interval overlapHard
  30. Generate expanding chronological splitsMedium
  31. Brier score for event probabilitiesEasy
  32. Chronological lagged rolling featuresMedium
  33. Finite-difference gradient checkMedium
  34. Sequential transaction-cost accountingMedium
  35. Reproducible synthetic linear dataEasy
  36. Better prediction error but worse net returnsHard
  37. Direction accuracy versus expected payoffMedium
  38. An alternative-data feature appears too earlyMedium
  39. Portfolio concentration behind an ML gainMedium
  40. Distribution shift after a market eventHard
  41. A tiny edge and uncertain costsMedium
  42. Learning a ranking rather than a levelMedium
  43. Retrain frequently or preserve stabilityMedium
  44. Calibration at the traded thresholdMedium
  45. A backtest cannot be reproducedMedium
  46. A neural network wins only in trainingMedium
  47. Transfer learning across marketsHard
  48. An anomaly detector flags every crisisMedium
  49. A feature derived from news textHard
  50. Choosing a research acceptance ruleHard