← Bank of America Interview Insights
The first part is easy enough, 30 out of 36 outcomes are unequal so 5/6.
Start by defining the sample space of 36 equally likely outcomes for two fair six-sided dice. For the first question, compute the probability of different values by subtracting the probability of matching values from 1. For the second question, recognize that by symmetry, the probability that one specific die is strictly higher than the other is half the probability of different values.
Pro tip: Explicitly state the assumption of independence and fairness, and mention that the symmetry argument relies on the dice being identical and fair. This shows rigor and avoids ambiguity.
State that there are 36 equally likely outcomes when rolling two fair six-sided dice, as each die has 6 independent outcomes.
Calculate the number of outcomes where the dice show different values: 6 * 5 = 30. Then divide by 36 to get 5/6.
Use symmetry: since the dice are fair and identical, the probability that die A > die B equals the probability that die B > die A. These two events are disjoint and together make up the 'different values' event. Thus, each has probability (5/6)/2 = 5/12.
Optionally, list or count the outcomes where one die is strictly higher (e.g., 15 outcomes for each direction) to confirm the result.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.
Felt like a warmup but they wanted proper definitions, not hand-waving.
Start with clear, concise definitions of mean and variance, using both mathematical formulas and intuitive explanations. Then, connect these concepts to practical applications in data science, especially in the context of banking and finance. Finally, briefly mention related concepts like standard deviation and expected value to demonstrate depth of understanding.
Pro tip: Emphasize that variance is in squared units, so standard deviation is often more interpretable; this shows attention to practical detail. Also, relate the concepts to real-world banking scenarios, such as risk assessment or customer lifetime value, to stand out.
State that the mean (expected value) is the weighted average of all possible values of a random variable, where weights are probabilities. Provide the formula: E[X] = Σ x * P(X=x) for discrete, or ∫ x f(x) dx for continuous.
Define variance as the expected squared deviation from the mean, measuring spread. Provide the formula: Var(X) = E[(X - μ)^2] = E[X^2] - (E[X])^2.
Describe mean as the 'center of mass' or long-run average, and variance as the average squared distance from that center, indicating how spread out the values are.
Discuss how mean and variance are used in data science, e.g., in feature scaling, anomaly detection, portfolio risk (variance as risk), and model evaluation.
Briefly note standard deviation (square root of variance) for interpretability, and possibly covariance for multivariate cases.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.
Yes, dividing by n gives a biased estimator because it underestimates the true variance.
Start by directly answering 'yes' and explaining that the standard sample variance with denominator n is biased because it underestimates the true population variance. Then introduce the unbiased estimator using denominator n-1 (Bessel's correction) and briefly explain why it works, perhaps with a simple example or intuition.
Pro tip: Mention that while n-1 is unbiased for variance, the sample standard deviation (square root of unbiased variance) is still biased for the population standard deviation due to Jensen's inequality. This shows depth and awareness of subtle statistical issues.
Clearly state the standard sample variance formula with denominator n and the corrected version with denominator n-1. Use proper notation: s_n^2 = (1/n)Σ(X_i - X̄)^2 and s_{n-1}^2 = (1/(n-1))Σ(X_i - X̄)^2.
Show that E[s_n^2] = σ^2 - σ^2/n = ((n-1)/n)σ^2, so it underestimates the true variance. This demonstrates that the bias is negative and depends on sample size.
State that using n-1 in the denominator (Bessel's correction) yields an unbiased estimator: E[s_{n-1}^2] = σ^2. Explain that this correction accounts for the fact that the sample mean is estimated from the same data, reducing the degrees of freedom by one.
Mention that while s_{n-1}^2 is unbiased for variance, its square root is not unbiased for standard deviation. Also note that for large n, the bias of s_n^2 becomes negligible, but for small samples it matters.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.
No, zero correlation does not generally imply independence.
Start by clarifying that zero correlation does not generally imply independence, then explain the special case of jointly normal variables where it does. Use a counterexample to illustrate the general case and state the condition precisely.
Pro tip: Mention that in finance, assuming independence from zero correlation can be dangerous, especially during market stress when correlations can spike, so always check for joint normality or use copulas.
Briefly define correlation and independence, emphasizing that correlation measures linear dependence while independence means no relationship whatsoever.
State that zero correlation does not imply independence in general, and provide a counterexample such as Y = X^2 with symmetric X.
Explain that if X and Y are jointly normally distributed, then zero correlation does imply independence.
Emphasize that the condition is joint normality (or more generally, that the joint distribution is determined by its mean and covariance, as in elliptical distributions).
Discuss why this matters in data science and finance, e.g., risk modeling, and caution against assuming independence without checking the joint distribution.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.
Linearity, independence of errors, homoscedasticity, normality of errors, no perfect multicollinearity.
Start by defining linear regression in simple terms, then explain the OLS method and its objective. Systematically walk through the assumptions, emphasizing how violations can be detected and addressed, and relate them to practical implications in a banking context.
Pro tip: Mention that in finance, assumptions like no autocorrelation and homoscedasticity are often violated due to time series data, and briefly discuss robust alternatives like Newey-West standard errors to show depth.
Explain that linear regression models the relationship between a dependent variable and one or more independent variables by fitting a linear equation.
Describe OLS as a method that estimates coefficients by minimizing the sum of squared residuals, and mention its closed-form solution.
Enumerate the key assumptions: linearity, independence, homoscedasticity, normality of errors, no multicollinearity, and no endogeneity.
For each assumption, briefly explain how to detect violations (e.g., residual plots, VIF) and potential fixes (e.g., transformations, robust standard errors).
Highlight why these assumptions matter in banking, such as for risk modeling, and mention common challenges like non-normality in financial data.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.
Beta equals (X transpose X) inverse times X transpose Y.
Start by stating the closed-form solution clearly: β = (XᵀX)⁻¹Xᵀy. Then explain the intuition behind it—minimizing the sum of squared residuals—and discuss the assumptions and practical considerations, especially in a banking context where data may be large or collinear.
Pro tip: Mention that in practice, you rarely compute the inverse directly; instead, you use QR decomposition or gradient descent for numerical stability and scalability. This shows you understand both theory and real-world implementation.
Clearly write the closed-form OLS solution: β = (XᵀX)⁻¹Xᵀy, defining X as the design matrix and y as the target vector.
Describe that OLS minimizes the residual sum of squares (RSS), and setting the gradient to zero yields the normal equations, leading to the closed-form solution.
Mention key assumptions: linearity, independence, homoscedasticity, normality of errors, and no perfect multicollinearity (XᵀX invertible).
Note that computing the inverse can be unstable or slow for large datasets; alternatives like QR decomposition, SVD, or regularized regression (ridge) are often used.
Connect to Bank of America by mentioning applications like credit scoring, risk modeling, or fraud detection, where interpretability and computational efficiency matter.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.
Multicollinearity inflates the variance of coefficient estimates, making them unstable and hard to interpret even if predictions are okay.
Start by defining multicollinearity and explaining its consequences on coefficient estimates and model interpretability. Then outline a systematic approach to detect it using VIF, correlation matrices, and condition indices, and finally discuss remedies such as removing variables, combining them, or using regularization. Emphasize the importance of context and business impact, especially in banking where interpretability and stability are critical.
Pro tip: In banking, multicollinearity can make it hard to explain which factors drive risk, so always tie your solution back to model interpretability and regulatory requirements. Mention that sometimes keeping correlated variables is acceptable if prediction is the only goal, but for inference, you must address it.
Define multicollinearity as high correlation among predictors, leading to unstable coefficient estimates, inflated standard errors, and reduced interpretability. Explain why this matters for inference and business decisions.
Use correlation matrices to spot pairwise correlations, calculate Variance Inflation Factor (VIF) for each predictor (VIF > 5 or 10 indicates concern), and consider condition indices for broader diagnostics.
Evaluate whether multicollinearity affects your specific goal: if prediction is key, it may be less critical; if inference or coefficient interpretation is needed, it must be addressed. Consider business implications and regulatory needs.
Apply remedies such as removing one of the correlated variables, combining them via PCA or factor analysis, using regularization (ridge, lasso), or collecting more data. Choose based on model purpose and domain knowledge.
Re-check VIF and model stability after changes, validate with cross-validation, and clearly communicate the rationale and impact to stakeholders, ensuring transparency.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.
Ridge adds an L2 penalty and shrinks coefficients toward zero without zeroing them out, which helps with collinearity by spreading the weight.
Start by defining Ridge and Lasso as regularized linear regression techniques, then explain how their penalties (L2 and L1) shrink coefficients differently. Connect this to multicollinearity by discussing how Ridge stabilizes estimates when predictors are correlated, while Lasso performs feature selection by zeroing out coefficients, which can arbitrarily drop correlated features. Emphasize the trade-offs and when to use each in a banking context.
Pro tip: Mention that in banking, interpretability and regulatory compliance often favor Lasso for its feature selection, but if all correlated features are important, Ridge is preferred. Also, note that Elastic Net combines both and can be a safe default.
Explain that both are linear regression variants with regularization: Ridge adds an L2 penalty (sum of squared coefficients), Lasso adds an L1 penalty (sum of absolute coefficients).
Describe how Ridge shrinks coefficients towards zero but never exactly zero, while Lasso can shrink some coefficients exactly to zero, performing feature selection.
Define multicollinearity as high correlation among predictors, which inflates coefficient variance and makes estimates unstable and hard to interpret.
Explain that Ridge handles multicollinearity by distributing coefficient weights among correlated features, reducing variance. Lasso, however, tends to pick one correlated feature and discard others, which can be problematic if all are relevant.
Highlight that Ridge is better when all features contribute, Lasso when sparse solutions are desired. Mention Elastic Net as a compromise and the importance of cross-validation for tuning lambda.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.
PCA finds directions of maximum variance in the data.
Start with a clear, intuitive definition of PCA as a dimensionality reduction technique that finds new orthogonal axes (principal components) maximizing variance. Then explain eigenvalues and eigenvectors as the variance magnitude and direction of these components, and finish by discussing limitations such as interpretability, linearity, and sensitivity to scaling.
Pro tip: Relate PCA to a real-world banking scenario, like reducing the dimensionality of customer transaction data for credit risk modeling, to show practical understanding and business relevance.
Explain PCA as a method to reduce dimensionality by projecting data onto a lower-dimensional subspace that captures the most variance.
Describe eigenvectors as the directions of the new axes (principal components) and eigenvalues as the amount of variance explained by each component.
Outline the steps: standardize data, compute covariance matrix, find eigenvectors/eigenvalues, sort by eigenvalues, and project data onto top components.
Mention key limitations: assumes linearity, may lose interpretability, sensitive to scaling, and can discard useful information if variance is not aligned with predictive power.
Summarize when PCA is useful (e.g., high-dimensional data, noise reduction) and when to consider alternatives (e.g., nonlinear methods like t-SNE or autoencoders).
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.