← Optiver Interview Insights

Optiver·Software Engineer·Technical Phone Screen·Intermediate

Intermediate
May 2026

Summary

Optiver quant engineer interview with a pretty gnarly probability/statistics question built around regression diagnostics. One question, but it had a lot of moving parts and required you to actually reason through residual distributions rather than just recite definitions.

Questions Asked (1)

Q1

You have a scatter of (x, y) points with a fitted linear regression line. Rank these three events from most to least probable: (1) the regression overestimates y for a new x drawn uniformly from the training x-range; (2) a new (x, y) pair has y greater than 20; (3) the absolute residual for a new (x, y) pair exceeds 1.5 times the standard deviation of the training residuals. Justify your ranking using residual distribution properties like zero-mean, approximate normality, and homoscedasticity.

A/B Testing & ExperimentationTechnical Trade-offsProduct Analytics & Metrics
Author's notes

This tripped me up more than I expected.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify the assumptions of linear regression: residuals are zero-mean, approximately normal, and homoscedastic. Then, analyze each event's probability using these properties, noting that event (1) is about the sign of the error (50% chance), event (2) depends on the distribution of y (which may be centered around 20 or not), and event (3) is about the tail probability of residuals exceeding 1.5 standard deviations (about 13.4% if normal). Finally, rank them based on these probabilities, justifying with the properties.

Pro tip: Emphasize that the ranking depends on the distribution of y and the model assumptions; without knowing the mean of y, event (2) could be more or less probable. A strong candidate will state assumptions and consider edge cases.

1. State model assumptions

Recall that in linear regression, residuals have mean zero, are approximately normally distributed, and have constant variance (homoscedasticity).

2. Analyze event (1)

The regression overestimates y for a new x if the residual is negative. Since residuals are zero-mean and symmetric (under normality), the probability is approximately 0.5.

3. Analyze event (3)

The absolute residual exceeds 1.5 times the standard deviation of training residuals. Under normality, P(|Z| > 1.5) ≈ 0.1336, so about 13.4%.

4. Analyze event (2)

The probability that a new y > 20 depends on the distribution of y. Without additional information, we cannot determine it precisely, but we can compare it to the others by considering typical scenarios.

5. Rank and justify

Compare the probabilities: event (1) is ~50%, event (3) is ~13.4%, and event (2) is unknown but likely less than 50% unless the mean of y is well above 20. Typically, event (2) might be the least probable if y is centered near 20, but it could be more probable if the mean is high. Justify based on assumptions.

Key Points to Mention

  • Residuals have zero mean, so P(overestimate) = P(residual < 0) ≈ 0.5 under symmetry.
  • Under normality, P(|residual| > 1.5σ) ≈ 0.1336, which is less than 0.5.
  • The probability of y > 20 depends on the distribution of y; if y is centered around 20, it could be close to 0.5, but if centered lower, it could be small.
  • Homoscedasticity ensures the residual variance is constant, so the 1.5σ threshold applies uniformly.
  • The ranking is not absolute without knowing the mean of y; state assumptions clearly.
  • Mention that if the model is good, residuals should be small relative to the range of y, but event (2) is about the absolute value of y, not residuals.

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