Interview questions
1,322 questions in the style trading and research firms ask, from probability and brainteasers to options, statistics and coding. Every question is open to read, with its hints. The 50 in the free sample can be answered and marked on the spot, no account needed; open 3 of their worked solutions to try, then all of them with a free account.
1,322 questions · page 25 of 34
Random question- 961A softmax output layer receives logits (2,1,0).
FoundationMachine learningCitadelTwo Sigma
- 962Bias and variance of nearest neighbours, part 1 of 3
FoundationMachine learning
- 963One step of gradient descent, part 1 of 4
FoundationMachine learning
- 964Cross-validating a return forecaster, part 1 of 3
FoundationMachine learning
- 965Scoring a tree split, part 1 of 3
FoundationMachine learning
- 966Why is ordinary shuffled k-fold cross-validation wrong on a financial time series?
AppliedMachine learningCitadelTwo SigmaPoint72 and CubistFree solution
- 967Why do gradient-boosted trees usually beat deep networks on daily…
AppliedMachine learningCitadelTwo SigmaD. E. Shaw
- 968You cluster assets by return correlation with k-means and the result looks arbitrary.
AppliedMachine learningTwo SigmaAQR
- 969A classifier assigns scores at random. What is its expected AUC?
AppliedMachine learningTwo SigmaPoint72 and Cubist
- 970In a ten-dimensional unit cube, what fraction of the volume lies within 0.1 of some face?
AppliedMachine learningJane StreetCitadelTwo Sigma
- 971A support vector machine has |w|= 4. What is the width of its margin?
AppliedMachine learningJane StreetCitadelD. E. Shaw
- 972Gradient descent with momentum uses a learning rate of 0.01 and β = 0.9.
AppliedMachine learningJane StreetCitadelHudson River Trading
- 973A batch normalisation layer follows a fully connected layer of 128 units.
AppliedMachine learningJane StreetCitadelJump Trading
- 974A classifier flags 500 cases, of which 40 are true positives, and misses 60 positives.
AppliedMachine learningJane StreetCitadelD. E. Shaw
- 975Data are uniform in the unit hypercube in 10 dimensions.
AppliedMachine learningJane StreetCitadelTwo Sigma
- 976A node has 50 positives and 50 negatives.
AppliedMachine learningCitadelTwo Sigma
- 977In gradient boosting, what happens if you halve the learning rate (shrinkage)…
AppliedMachine learningCitadelTwo Sigma
- 978Twenty percent of messages are spam.
AppliedMachine learningCitadelTwo Sigma
- 979What does the EM algorithm guarantee when fitting a Gaussian mixture model?
AppliedMachine learningCitadelTwo Sigma
- 980An anomaly detector has sensitivity 90% and specificity 95%.
AppliedMachine learningJane StreetCitadelTwo Sigma
- 981You apply a strictly increasing transformation to a classifier’s scores.
AppliedMachine learningCitadelTwo Sigma
- 982Gradient descent on f(x) = x² starts at x = 1 with learning rate 0.1.
AppliedMachine learningCitadelTwo Sigma
- 983A quadratic loss has Hessian eigenvalues 4 and 0.5.
AppliedMachine learningCitadelTwo Sigma
- 984A convolutional layer uses 3 × 3 kernels, 16 input channels and 32 output…
AppliedMachine learningCitadelTwo Sigma
- 985Bias and variance of nearest neighbours, part 2 of 3
AppliedMachine learning
- 986One step of gradient descent, part 2 of 4
AppliedMachine learning
- 987Cross-validating a return forecaster, part 2 of 3
AppliedMachine learning
- 988Scoring a tree split, part 2 of 3
AppliedMachine learning
- 989What does embargoing do that purging does not?
AdvancedMachine learningTwo SigmaPoint72 and CubistFree solution
- 990A regularised linear model beats your tuned gradient-boosting ensemble out of sample.
AdvancedMachine learningCitadelTwo SigmaAQR
- 991Two features are nearly identical, and a random forest reports low importance for both.
AdvancedMachine learningCitadelTwo SigmaAQR
- 992What does the kernel trick actually buy you?
AdvancedMachine learningJane StreetCitadelJump Trading
- 993You increase the batch size from 32 to 512 without changing anything else. What happens?
AdvancedMachine learningJane StreetCitadelTwo Sigma
- 994Why did rectified linear units largely replace sigmoids in deep networks?
AdvancedMachine learningJane StreetCitadelD. E. Shaw
- 995You cluster a universe of assets from their correlation matrix.
AdvancedMachine learningJane StreetCitadelTwo Sigma
- 996Each tree’s prediction has variance 1 and any two trees’ predictions correlate at 0.3.
AdvancedMachine learningCitadelTwo SigmaAQR
- 997Adam’s first-moment estimate starts at 0 with β₁ = 0.9.
AdvancedMachine learningCitadelTwo Sigma
- 998Bias and variance of nearest neighbours, part 3 of 3
AdvancedMachine learning
- 999One step of gradient descent, part 3 of 4
AdvancedMachine learning
- 1000Cross-validating a return forecaster, part 3 of 3
AdvancedMachine learning