← Capital One Interview Insights

Capital One·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Capital One data scientist interview that went deep into linear regression diagnostics. The whole thing felt like a grad school exam on applied stats, which I was not fully expecting from a DS role at a fintech company.

Questions Asked (6)

Q1

You have a standardized coefficient for a variable like tailwind at -0.8 with a p-value of 0.07. How do you interpret that, and how do you distinguish statistical significance from practical significance?

Product Analytics & MetricsTechnical Trade-offs
Author's notes

The p-value being above 0.05 tripped me up for a second.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by interpreting the coefficient and p-value in context, noting that the effect is marginally non-significant at the 0.05 level but may still be practically meaningful. Then clearly distinguish statistical significance (likelihood the effect is due to chance) from practical significance (real-world impact and business relevance). Use the specific example to illustrate how you would communicate this to stakeholders and make decisions.

Pro tip: Emphasize that p-values are not a measure of effect size or importance; a non-significant result can still be actionable if the effect size is large and the business context supports it. Also, mention that with a p-value of 0.07, you'd consider the study's power and whether a larger sample might yield significance.

1. Interpret the coefficient and p-value

Explain that the standardized coefficient of -0.8 indicates a strong negative relationship (in standard deviation units) between the predictor and outcome. The p-value of 0.07 means there is a 7% chance of observing such an effect if the null hypothesis were true, which is above the conventional 0.05 threshold for statistical significance.

2. Define statistical significance

Clarify that statistical significance is about whether an observed effect is likely due to chance, typically assessed via p-value < 0.05. It does not speak to the size or importance of the effect.

3. Define practical significance

Explain that practical significance concerns whether the effect size is large enough to matter in the real world, considering business context, costs, and benefits. A statistically non-significant result can still be practically significant if the effect is meaningful.

4. Apply to the example

For the tailwind variable, the coefficient of -0.8 suggests a strong negative effect, but the p-value of 0.07 means it's not statistically significant at the 5% level. However, given the large effect size, it could be practically significant; you'd assess the business impact and consider if further data collection is warranted.

5. Communicate and decide

Discuss how you would communicate this nuance to stakeholders: highlight the uncertainty but also the potential impact. Recommend actions such as running a larger study, checking for confounding variables, or conducting a cost-benefit analysis to determine if the effect is worth acting on.

Key Points to Mention

  • Standardized coefficient interpretation: -0.8 means a one standard deviation increase in the predictor is associated with a 0.8 standard deviation decrease in the outcome, all else equal.
  • P-value interpretation: 0.07 is not statistically significant at the 5% level, but it is close; consider the context and consequences of Type I vs Type II errors.
  • Statistical significance vs practical significance: statistical significance is about probability, practical significance is about magnitude and real-world impact.
  • Effect size matters: a large coefficient can be practically significant even if not statistically significant, especially in small samples.
  • Consider study power and sample size: a non-significant result might be due to low power; a larger sample could yield significance.
  • Business context: at a company like Capital One, decisions should weigh the cost of action against the potential benefit, not just p-values.

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

Q2

Walk through the differences between R-squared, adjusted R-squared, and out-of-fold RMSE. Can R-squared go up while adjusted R-squared goes down, and what would you do in that situation?

Technical Trade-offsData Modeling
Author's notes

Yes it can, when you add a predictor that contributes basically nothing.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clearly defining each metric and its purpose: R-squared measures in-sample fit, adjusted R-squared penalizes for model complexity, and out-of-fold RMSE estimates out-of-sample prediction error. Then explain that R-squared can increase while adjusted R-squared decreases when adding a predictor that improves fit slightly but not enough to justify the complexity penalty. Finally, discuss your diagnostic and decision-making process, emphasizing the importance of out-of-sample validation and business context.

Pro tip: Mention that adjusted R-squared is not a substitute for out-of-sample validation; always prioritize out-of-fold RMSE or cross-validated metrics for model selection, especially in high-stakes domains like finance.

1. Define the metrics

Briefly explain what each metric represents: R-squared as proportion of variance explained in-sample, adjusted R-squared as a penalized version for number of predictors, and out-of-fold RMSE as a cross-validated measure of prediction error on unseen data.

2. Explain the divergence

Describe how adding a predictor can increase R-squared but decrease adjusted R-squared if the improvement in fit is smaller than the penalty for the additional parameter. Use a simple example or formula to illustrate.

3. Diagnose the situation

Outline steps to investigate: check for multicollinearity, assess the predictor's significance, and compare out-of-fold RMSE to see if the new model actually improves out-of-sample performance.

4. Decide on action

Based on diagnostics, decide whether to keep the predictor. If adjusted R-squared decreases and out-of-fold RMSE worsens or stays the same, consider removing the predictor or using regularization.

5. Emphasize out-of-sample validation

Conclude by stressing that out-of-fold RMSE (or other cross-validated metrics) should guide model selection, as it directly measures generalization performance, which is often the ultimate goal.

Key Points to Mention

  • R-squared always increases or stays the same when adding predictors, while adjusted R-squared can decrease.
  • Adjusted R-squared penalizes model complexity based on the number of predictors and sample size.
  • Out-of-fold RMSE provides an unbiased estimate of out-of-sample error, unlike in-sample metrics.
  • Multicollinearity can cause unstable coefficient estimates and affect adjusted R-squared.
  • Regularization techniques (e.g., LASSO, ridge) can help manage complexity and improve out-of-sample performance.
  • Always consider business context and the cost of prediction errors when choosing a model.

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

Q3

How would you detect and handle multicollinearity in a regression model? When would you remove a variable versus apply regularization, and what happens to coefficients and standard errors when multicollinearity is present?

Data ModelingTechnical Trade-offs
Author's notes

VIF is the go-to.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining multicollinearity and its consequences, then walk through detection methods (VIF, correlation matrix, condition index). Finally, discuss decision criteria for removal vs. regularization, emphasizing the trade-offs in interpretability and predictive performance.

Pro tip: Mention that in regulated industries like banking, interpretability often trumps pure predictive performance, so removing variables or using domain knowledge to combine them may be preferred over black-box regularization.

1. Define and Diagnose

Explain what multicollinearity is and its impact on coefficient estimates and standard errors. Then list common detection techniques: correlation matrix, VIF, tolerance, condition index.

2. Assess Severity and Context

Interpret VIF thresholds (e.g., >5 or >10) and consider the goal: inference vs. prediction. High VIF may be acceptable if prediction is the sole focus.

3. Choose a Remedy

Decide between removing a variable (based on domain knowledge, VIF, or stepwise selection) or applying regularization (ridge, lasso, elastic net). Discuss pros and cons of each.

4. Explain Coefficient and SE Effects

Detail how multicollinearity inflates standard errors, making coefficients unstable and potentially flipping signs, while overall model fit may remain good.

5. Validate and Communicate

Emphasize the importance of validating the chosen approach with cross-validation and communicating the trade-offs to stakeholders, especially in a business context.

Key Points to Mention

  • Variance Inflation Factor (VIF) and its interpretation
  • Impact on standard errors: inflation leads to wider confidence intervals and less reliable p-values
  • Coefficient instability: signs may reverse and magnitudes change dramatically
  • Ridge regression shrinks coefficients but keeps all variables; lasso performs variable selection
  • Removal criteria: domain knowledge, VIF threshold, and impact on model performance
  • Trade-off between interpretability and predictive accuracy, especially in regulated industries

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

Q4

How do you diagnose heteroskedasticity from residual plots, and what tests or fixes would you apply?

Data ModelingRoot Cause Analysis
Author's notes

Fan-shaped residuals are the visual tell.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by explaining how to visually detect heteroskedasticity from residual plots, then discuss formal statistical tests to confirm it, and finally outline potential fixes. Emphasize the importance of addressing heteroskedasticity for valid inference and model performance.

Pro tip: Mention that heteroskedasticity doesn't bias coefficient estimates but affects standard errors, so while predictions might still be okay, inference and uncertainty quantification can be misleading. Also, consider the context: in finance, volatility clustering is common, so techniques like GARCH might be relevant.

1. Visual Diagnosis

Plot residuals against fitted values or predictors. Look for patterns like a funnel shape (increasing variance) or other non-random structures indicating heteroskedasticity.

2. Formal Tests

Apply statistical tests such as Breusch-Pagan, White, or Goldfeld-Quandt to formally test for heteroskedasticity. These tests help confirm the visual diagnosis.

3. Assess Impact

Determine if heteroskedasticity affects the goals. For inference, it can invalidate standard errors; for prediction, it might not be critical unless uncertainty matters.

4. Remedial Measures

Apply fixes: use robust standard errors (e.g., White's), transform variables (e.g., log), use weighted least squares, or adopt models that account for heteroskedasticity (e.g., GARCH for time series).

5. Validate and Iterate

After applying fixes, re-check residual plots and tests to ensure the issue is resolved. Iterate if necessary.

Key Points to Mention

  • Residual plots: residuals vs. fitted values, residuals vs. predictors, scale-location plot
  • Common patterns: funnel shape, increasing spread, non-constant variance
  • Tests: Breusch-Pagan, White, Goldfeld-Quandt, Park test
  • Consequences: inefficient OLS estimates, biased standard errors, invalid hypothesis tests
  • Fixes: robust standard errors, weighted least squares, variable transformations, heteroskedasticity-consistent models
  • Context: in finance, consider GARCH models for volatility clustering

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

Q5

What happens if you omit the intercept from a linear regression, or if you standardize only some of the features and not others?

Data ModelingTechnical Trade-offs
Author's notes

Omitting the intercept forces the regression line through the origin, which biases everything unless the data actually supports that constraint.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying that both issues relate to model specification and feature scaling, then explain the consequences of each separately. For omitted intercept, discuss how it forces the regression line through the origin, potentially biasing coefficients and inflating R-squared. For partial standardization, explain how it leads to incomparable coefficients and can affect regularization and interpretation. Conclude with practical recommendations for when each might be acceptable and how to avoid pitfalls.

Pro tip: Mention that omitting the intercept can be valid if domain knowledge strongly suggests zero outcome at zero predictors, but always check residual plots and compare with the intercept model. For standardization, emphasize that it's crucial for regularized models and when comparing feature importance, but not strictly necessary for plain OLS if interpretation is in original units.

1. Clarify the two scenarios

Separate the question into two parts: omitting the intercept and partial standardization. State that each has distinct implications for model fit, interpretation, and performance.

2. Explain omitted intercept consequences

Discuss how forcing the line through the origin can lead to biased coefficient estimates, especially if the true relationship has a non-zero intercept. Mention that R-squared can be misleadingly high and residuals may show patterns.

3. Explain partial standardization consequences

Describe how standardizing only some features makes coefficients non-comparable and can distort regularization penalties (e.g., Lasso/Ridge) and distance-based methods. Also note that it complicates interpretation of feature importance.

4. Provide practical guidance

Recommend generally including an intercept unless theoretically justified, and standardizing all features (or none) for consistency, especially when using regularization or comparing coefficients.

5. Connect to business context

Relate to Capital One's focus on interpretability and model risk management: omitted intercepts can hide bias, and inconsistent scaling can lead to unstable models, affecting credit decisions.

Key Points to Mention

  • Omitted intercept forces regression through origin, violating the assumption that errors have mean zero if true intercept is non-zero.
  • R-squared can be artificially inflated when intercept is omitted, and its usual interpretation changes.
  • Standardization ensures features are on the same scale, which is critical for regularized models and comparing coefficients.
  • Partial standardization leads to coefficients that are not directly comparable and can mislead feature importance.
  • Regularization penalties (L1/L2) are scale-sensitive, so inconsistent scaling can unfairly penalize some features.
  • Always validate model assumptions via residual plots and consider domain knowledge before omitting intercept or scaling selectively.

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

Q6

Your residuals show curvature and heavy non-normal tails. What modeling changes would you consider, and how would those changes affect inference and prediction intervals?

Data ModelingTechnical Trade-offs
Author's notes

Curvature means the linear assumption is probably wrong, so polynomial terms or a nonlinear model.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by diagnosing the specific nature of the curvature and non-normal tails (e.g., heteroscedasticity, outliers, skewness) and connect them to potential violations of model assumptions. Then propose a hierarchy of modeling changes—from data transformations and robust methods to more flexible models—and explicitly discuss how each affects coefficient estimates, standard errors, and prediction intervals. Emphasize the trade-off between interpretability and flexibility, and validate changes with residual diagnostics and out-of-sample performance.

Pro tip: In regulated industries like banking, always tie model changes to business impact and regulatory constraints (e.g., interpretability for adverse action notices). Mention that while flexible models may improve fit, they can complicate explainability and require additional validation for fair lending compliance.

1. Diagnose the residual issues

Identify whether the curvature indicates non-linearity (e.g., missing polynomial terms, interactions) and whether heavy tails are due to outliers, skewness, or heteroscedasticity. Use plots (residuals vs. fitted, QQ-plots) and tests (Breusch-Pagan, Shapiro-Wilk) to confirm.

2. Consider data transformations

Apply transformations to the target (e.g., log, Box-Cox) or predictors (e.g., splines, polynomials) to address curvature and stabilize variance. Discuss how transformations affect interpretability and back-transformation for prediction intervals.

3. Explore robust and flexible models

If transformations are insufficient, consider robust regression (e.g., Huber, quantile) for heavy tails, or non-linear models (e.g., GAMs, tree-based ensembles) for curvature. Note that robust methods downweight outliers, while flexible models may capture complex patterns but risk overfitting.

4. Assess impact on inference

Explain how each change affects coefficient estimates, standard errors, and hypothesis tests. For example, robust regression yields consistent estimates under non-normality but requires robust standard errors; GAMs provide flexible inference but with smoothing parameter uncertainty.

5. Evaluate prediction intervals

Discuss how prediction intervals change: transformations require back-transformation (which may bias intervals), robust methods may widen intervals for outliers, and flexible models can produce narrower but potentially overconfident intervals. Validate with coverage probability on hold-out data.

Key Points to Mention

  • Heteroscedasticity and its impact on standard errors (use robust standard errors or weighted least squares).
  • Non-linearity: polynomial terms, splines, or generalized additive models (GAMs) to capture curvature.
  • Heavy tails: robust regression (Huber, quantile) or heavy-tailed distributions (t-distribution errors).
  • Transformations: log or Box-Cox for skewness and variance stabilization, and back-transformation for prediction intervals.
  • Trade-offs: interpretability vs. flexibility, and regulatory constraints in finance (e.g., explainability for adverse action).
  • Validation: residual diagnostics, cross-validation, and coverage of prediction intervals.

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