← Boston Consulting Group Interview Insights

Boston Consulting Group·Data Scientist·Take-home Assignment·Senior

Senior
Jun 2026

Summary

BCG data scientist take-home, basically one giant multi-part problem covering the full ML pipeline from temporal splitting to production deployment. Pretty dense for a single assignment and felt more like a work sample than an interview.

Questions Asked (4)

Q1

Build a reproducible training pipeline for a severely imbalanced binary classification dataset (1% positive rate) that uses temporal train/val/test splits, applies standardization and one-hot encoding, handles class imbalance inside CV folds without leaking validation data, trains a strong baseline, and outputs calibrated probabilities.

System DesignTechnical Trade-offsData Modeling
Author's notes

This was the bulk of the work.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around a modular pipeline that enforces temporal order, prevents data leakage, and handles imbalance within CV folds. Emphasize that all preprocessing (standardization, one-hot encoding) and resampling must be fit only on training folds, and that calibration is applied after model training to output reliable probabilities.

Pro tip: Mention that you would use a time-based split and within each CV fold, apply SMOTE or class weights only to the training portion, never to validation. Also, highlight that calibration should be done on a separate calibration set or via cross-validation to avoid overfitting.

1. Data splitting and temporal ordering

Split the data into train, validation, and test sets based on time, ensuring no future data leaks into training. Use the earliest data for training, next for validation, and latest for testing.

2. Preprocessing pipeline with leakage prevention

Build a pipeline that applies standardization and one-hot encoding, fitting only on the training fold within each CV iteration. Use scikit-learn's Pipeline and ColumnTransformer to encapsulate this.

3. Handling class imbalance within CV folds

Within each training fold, apply techniques like SMOTE, class weights, or undersampling only to the training data, never to validation. Use imblearn's pipeline to integrate this seamlessly.

4. Model training and hyperparameter tuning

Train a strong baseline (e.g., XGBoost, LightGBM) with hyperparameter tuning using time-series cross-validation. Evaluate using metrics like AUC-PR, recall, or F1 due to imbalance.

5. Probability calibration and final evaluation

Calibrate predicted probabilities using Platt scaling or isotonic regression on a held-out calibration set. Evaluate on the test set with proper metrics and ensure reproducibility by setting random seeds.

Key Points to Mention

  • Temporal splits to respect time order and prevent lookahead bias
  • Leakage prevention: fit preprocessing and resampling only on training folds
  • Class imbalance handling: SMOTE, class weights, or undersampling within CV
  • Use of pipelines (scikit-learn, imblearn) for reproducibility and modularity
  • Calibration methods (Platt scaling, isotonic regression) and evaluation metrics (AUC-PR, Brier score)
  • Reproducibility: random seeds, versioning, and logging

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

Report ROC-AUC and PR-AUC on the test split, recall at 5% FPR, and the decision threshold that maximizes F1 subject to recall being at least 0.90.

Product Analytics & MetricsTechnical Trade-offs
Author's notes

PR-AUC is the one that actually matters at 1% positive rate and I made sure to say that upfront.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business context and evaluation goals, then walk through the computation of each metric on the test set, emphasizing the trade-offs between ROC-AUC and PR-AUC for imbalanced data. Finally, explain how to find the decision threshold that maximizes F1 while satisfying the recall constraint, and discuss the implications for model deployment.

Pro tip: Always relate the metrics to business impact—e.g., how a 5% FPR translates to false alarms per 1,000 customers—and mention that threshold selection should be validated on a separate validation set to avoid overfitting the test set.

1. Clarify context and metric definitions

Confirm the problem type, class balance, and business costs of false positives/negatives. Briefly define ROC-AUC, PR-AUC, recall at 5% FPR, and the F1-constrained threshold.

2. Compute ROC-AUC and PR-AUC

Calculate both metrics on the test set using predicted probabilities. Explain that ROC-AUC measures overall ranking ability, while PR-AUC is more informative for imbalanced data.

3. Calculate recall at 5% FPR

Determine the threshold where the false positive rate equals 5%, then compute the corresponding recall. Discuss how this metric balances detection rate with acceptable false alarms.

4. Find threshold maximizing F1 with recall ≥ 0.90

Search over thresholds to find the one that maximizes F1 score while ensuring recall is at least 0.90. If no threshold satisfies the constraint, report that and suggest revisiting the model or constraint.

5. Interpret and communicate results

Summarize the metrics, highlight trade-offs, and provide recommendations for threshold selection based on business priorities. Mention potential next steps like calibration or cost-sensitive learning.

Key Points to Mention

  • ROC-AUC vs PR-AUC: PR-AUC is preferred for imbalanced datasets as it focuses on the positive class.
  • Recall at 5% FPR: a fixed operating point that quantifies detection ability at an acceptable false alarm rate.
  • Threshold optimization: F1 maximization with a recall floor ensures a balance between precision and recall while meeting a minimum detection requirement.
  • Business context: translating FPR and recall into concrete costs (e.g., false alerts per 1,000 customers) to guide threshold selection.
  • Validation strategy: using a separate validation set for threshold tuning to avoid overfitting the test set.
  • Model calibration: ensuring predicted probabilities are well-calibrated for threshold-based decisions.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

How would you choose the operating point for a production system with a hard constraint of at most 2 false positives per 1,000 predictions?

Product Analytics & MetricsTechnical Trade-offs
Author's notes

Took me a minute to translate that into a precision floor.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business context and the cost asymmetry between false positives and false negatives, then translate the constraint into a threshold on the model's predicted probabilities. Use the precision-recall curve to find the threshold that satisfies the false positive rate, and validate it on a holdout set while considering operational factors like model calibration and monitoring.

Pro tip: Emphasize that the constraint is on the false positive rate per 1,000 predictions, not on precision, and that you would monitor the actual false positive rate in production to detect drift and adjust the threshold if needed.

1. Clarify business context and costs

Understand the implications of false positives and false negatives, and confirm that the hard constraint is indeed on false positives per 1,000 predictions. Discuss with stakeholders whether the constraint is absolute or if there is flexibility.

2. Evaluate model performance

Use a validation set to compute the precision-recall curve and the false positive rate at various thresholds. Identify the threshold that yields at most 2 false positives per 1,000 predictions.

3. Select operating point

Choose the threshold that satisfies the constraint while maximizing the true positive rate (or minimizing false negatives). Consider if the model is well-calibrated; if not, calibrate it first to ensure the threshold corresponds to the desired false positive rate.

4. Validate and simulate

Test the chosen threshold on a holdout set or through cross-validation to ensure the false positive rate holds. Simulate production scenarios to estimate the impact on business metrics.

5. Monitor and iterate

Implement monitoring to track the actual false positive rate in production. Set up alerts for drift and define a process to re-evaluate the threshold periodically or when data distribution changes.

Key Points to Mention

  • Cost asymmetry between false positives and false negatives
  • Precision-recall curve and threshold selection
  • Model calibration (e.g., Platt scaling, isotonic regression)
  • False positive rate per 1,000 predictions vs. precision
  • Validation on holdout set and cross-validation
  • Production monitoring and drift detection

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

How might model calibration drift over time, and what technique would you use to monitor and re-calibrate without leaking future labels?

Root Cause AnalysisTechnical Trade-offsA/B Testing & Experimentation
Author's notes

Went with expected calibration error tracked on a rolling window of incoming predictions, then re-calibrate on a held-out recent slice once enough labels come in.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by explaining why calibration drift occurs (e.g., data drift, concept drift, model staleness) and its impact on decision-making. Then describe a monitoring framework that uses only past data and delayed labels, such as calibration curves on a rolling window with statistical process control. Finally, propose a re-calibration technique like Platt scaling or isotonic regression that can be updated without future labels, emphasizing the importance of avoiding leakage.

Pro tip: Emphasize that in practice, labels are often delayed, so you need to design a monitoring system that works with proxy metrics or delayed labels, and use techniques like 'delayed feedback' models to adjust. Also, mention that re-calibration should be done on a separate validation set that is time-based to avoid leakage.

1. Define calibration and drift

Clarify what model calibration means (e.g., predicted probabilities matching observed frequencies) and how drift can cause it to degrade over time due to changes in data distribution or concept.

2. Monitor calibration without future labels

Describe monitoring using only past data: track calibration on a rolling window with delayed labels, use proxy metrics (e.g., prediction distribution, confidence histograms), and apply statistical tests like Kolmogorov-Smirnov or calibration error metrics.

3. Detect drift and trigger re-calibration

Set up alerts when calibration error exceeds a threshold or when drift is detected via statistical process control. Use a time-based validation set to avoid leakage.

4. Re-calibrate without future labels

Apply post-hoc calibration methods like Platt scaling, isotonic regression, or beta calibration on recent labeled data (which may be delayed) or using techniques like online learning to update calibration parameters incrementally.

5. Validate and iterate

Validate the re-calibrated model on a holdout set from the same time period, ensuring no future data is used. Monitor continuously and iterate as needed.

Key Points to Mention

  • Types of drift: covariate shift, label shift, concept drift
  • Calibration metrics: reliability diagrams, expected calibration error (ECE), maximum calibration error (MCE)
  • Monitoring techniques: rolling window analysis, statistical process control (e.g., CUSUM), drift detection methods (e.g., ADWIN, Page-Hinkley)
  • Re-calibration methods: Platt scaling, isotonic regression, temperature scaling, beta calibration
  • Avoiding leakage: use time-based splits, delayed labels, and ensure no future data is used in training or calibration
  • Business impact: how miscalibration affects decision-making and costs, especially in consulting contexts

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.