I85PREP
⌘ K
CQF study materials
CQF study resources

CQF practice questions

One question at a time. Build your CQF understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Quantitative finance and machine learning

In a Kalman filter applied to a linear state-space model, what does the innovation (measurement residual) sequence represent under correct model specification?

  1. A biased signal whose mean equals the observation matrix H
  2. A white-noise process with zero mean and covariance equal to the innovation covariance S_t Correct answer.
  3. A serially correlated process whose autocorrelation equals the state transition matrix
  4. A deterministic drift term that must be removed before filtering

Explanation

When the model is correctly specified, the one-step-ahead prediction errors are zero-mean and serially uncorrelated, with covariance S = H P H' + R. This property underlies innovation-based diagnostics and the innovation-form likelihood. Systematic autocorrelation or non-zero mean signals model misspecification.

Question 2Quantitative finance and machine learning

A machine-learning model is trained on 10 years of daily returns and tested on a random 20% holdout drawn from the same period. Which flaw most directly inflates the reported out-of-sample performance?

  1. Choice of a non-Gaussian loss function
  2. Use of too few cross-validation folds
  3. Insufficient regularization of the model coefficients
  4. Look-ahead leakage through temporally overlapping samples and features Correct answer.

Explanation

Random holdout on time-series data lets the training set contain observations adjacent to or after test observations, so information leaks across the split. Features built from rolling windows or labels spanning multiple days make this leakage severe. The correct evaluation uses chronological splits with purging and embargo.

Question 3Quantitative finance and machine learning

In a feedforward neural network used for return forecasting, what is the primary role of a nonlinear activation function such as ReLU in a hidden layer?

  1. To normalize input features to unit variance
  2. To reduce the number of parameters in the network
  3. To guarantee the loss surface is globally convex
  4. To allow the network to approximate nonlinear functions of the inputs Correct answer.

Explanation

Stacked linear layers collapse to a single linear map, so nonlinear activations are what let the network represent interactions and curvature in the data. ReLU introduces piecewise-linear nonlinearity that is cheap to differentiate. Convexity and normalization are separate concerns.

Question 4Quantitative finance and machine learning

A portfolio risk model assumes returns are multivariate normal. Which observed feature of daily equity returns most strongly contradicts this assumption in the tails?

  1. Autocorrelation in squared returns
  2. A slightly negative sample skewness in a single asset
  3. Excess kurtosis producing more extreme joint moves than the normal implies Correct answer.
  4. A sample mean close to zero

Explanation

Fat tails mean extreme returns occur far more often than a Gaussian density predicts, so tail risk and joint crash probabilities are understated. Squared-return autocorrelation indicates volatility clustering, a separate violation. Near-zero mean is consistent with the assumption rather than contradicting it.

Question 5Quantitative finance and machine learning

Which statement best describes the bias-variance trade-off when increasing the depth of a regression tree ensemble on noisy financial features?

  1. Deeper trees lower both bias and variance monotonically
  2. Deeper trees raise bias and lower variance
  3. Tree depth affects only training time, not the bias-variance balance
  4. Deeper trees lower bias but raise variance, so out-of-sample error can rise Correct answer.

Explanation

A deeper tree fits finer structure, reducing approximation bias while increasing sensitivity to sampling noise. Beyond an optimum the variance term dominates and generalization degrades. Ensembling and pruning are used to control that variance.

Question 6Quantitative finance and machine learning

A trader computes a 20-day exponentially weighted moving average volatility with decay parameter lambda close to 1. Compared with a simple 20-day moving average of squared returns, what is the main behavioral difference?

  1. It assigns equal weight to all 20 observations
  2. It produces a volatility estimate that is always lower
  3. It removes the need to estimate a mean return
  4. It weights recent squared returns more heavily and responds faster to volatility shifts Correct answer.

Explanation

The EWMA applies geometrically declining weights, so the most recent shocks dominate and the estimate reacts more quickly to regime changes. A simple moving average gives equal weights and drops old data abruptly. Neither estimator is uniformly lower.

Question 7Quantitative finance and machine learning

In principal component analysis of a covariance matrix of asset returns, what does the first principal component typically capture for an equity universe?

  1. The bid-ask spread component of returns
  2. Idiosyncratic noise specific to each asset
  3. A common market factor affecting most assets with the same sign Correct answer.
  4. The term structure of interest rates

Explanation

The leading eigenvector of an equity covariance matrix usually has loadings of similar sign across assets, representing a market-wide level factor. Subsequent components capture sector and style contrasts. Idiosyncratic risk is spread across the many small eigenvalues.

Question 8Quantitative finance and machine learning

A logistic regression is fitted to predict the sign of next-day returns. What does the fitted linear combination inside the logit represent?

  1. The expected return in basis points
  2. The log-odds of a positive return conditional on the features Correct answer.
  3. The variance of the return distribution
  4. The probability of a positive return directly

Explanation

Logistic regression models the log-odds as a linear function of features, and the sigmoid maps that to a probability. The output is a classification probability, not a magnitude of return. Converting to expected return requires an additional assumption about payoff sizes.

Question 9Quantitative finance and machine learning

A researcher uses k-fold cross-validation with shuffling to tune a model on daily returns where each label is the sum of the next 5 days' returns. Why does this produce optimistic error estimates?

  1. Overlapping label windows cause training and validation observations to share future information Correct answer.
  2. The model cannot be fit on shuffled data
  3. Shuffling reduces the effective sample size below the number of folds
  4. Five-day labels are always non-stationary

Explanation

Labels spanning multiple days overlap in time, so a validation observation's outcome partly appears in neighboring training labels. This leakage inflates measured accuracy. Purging and embargo around the validation block remove the overlap.

Question 10Quantitative finance and machine learning

In a hidden Markov model for market regimes with two states, the filtered probability of the high-volatility state is 0.7 today. The transition matrix has persistence 0.9 on the diagonal. Ignoring today's new observation, what is the predicted probability of the high-volatility state tomorrow?

  1. 0.66 Correct answer.
  2. 0.70
  3. 0.34
  4. 0.90

Explanation

The prediction step applies the transition matrix: 0.7 times 0.9 plus 0.3 times 0.1 equals 0.63 plus 0.03, giving 0.66. This is the Chapman-Kolmogorov prediction before incorporating the next observation. The filtered value 0.7 is not carried forward unchanged.

Question 11Quantitative finance and machine learning

A gradient-boosted tree model and a linear ridge model are compared on the same features. The boosted model has much lower training error but slightly worse test error. Which conclusion is best supported?

  1. The features contain no predictive information
  2. The boosted model is overfitting relative to the simpler model on this sample Correct answer.
  3. The ridge model has higher bias and should be discarded
  4. The boosted model needs a larger learning rate

Explanation

A large gap between training and test error is the signature of overfitting, and the simpler model generalizes at least as well here. Discarding the ridge model is unwarranted since it performs better out of sample. Increasing the learning rate would typically worsen instability.

Question 12Quantitative finance and machine learning

A quant fits an AR(1) model to a return series and finds the estimated autoregressive coefficient is 0.03 with a standard error of 0.02. What is the most defensible interpretation for trading purposes?

  1. The model is misspecified because coefficients must exceed one
  2. The coefficient implies a half-life of about 23 days
  3. The evidence for exploitable linear predictability is weak and likely within noise Correct answer.
  4. The series is strongly mean-reverting and warrants a large contrarian position

Explanation

The t-statistic of roughly 1.5 is not significant at conventional levels, so the apparent predictability could easily be sampling noise. Transaction costs would further erode any tiny edge. Claiming strong mean reversion or a precise half-life overstates the evidence.

Question 13Quantitative finance and machine learning

A risk manager estimates a 99% one-day Value at Risk of 2.5% using a historical simulation on 500 days of data. What is a key limitation of this approach for tail risk?

  1. It systematically overstates VaR during calm periods
  2. The estimate is driven by the few most extreme historical days and cannot extrapolate beyond them Correct answer.
  3. It cannot incorporate correlations among assets
  4. It requires assuming returns are normally distributed

Explanation

Historical simulation is purely empirical, so the 99% quantile is interpolated from a handful of worst observations and cannot exceed the worst historical loss. It also reacts slowly to regime change. It makes no normality assumption, which is one of its selling points.

Question 14Quantitative finance and machine learning

In a long-short equity backtest, a signal is standardized cross-sectionally each day and portfolios are formed on the top and bottom deciles. Which control most directly addresses the concern that results come from unintended sector bets?

  1. Extending the backtest to a longer history
  2. Winsorizing the signal at three standard deviations
  3. Increasing the rebalancing frequency to daily
  4. Neutralizing the signal to sector exposures before forming portfolios Correct answer.

Explanation

Sector-neutralizing the signal removes the component explained by industry membership, so the strategy's returns reflect stock selection rather than sector rotation. The other choices affect estimation noise or sample length but do not isolate sector exposure.

Question 15Quantitative finance and machine learning

A Gaussian mixture model with three components is fitted to daily returns. The fitted mixture has two components with large variances and small weights and one with a small variance and large weight. What does this structure typically represent?

  1. A linear trend in the mean return
  2. Measurement error in the return data
  3. A calm regime plus a rare crash regime, approximating fat tails Correct answer.
  4. A deterministic seasonal pattern in returns

Explanation

A dominant low-variance component plus rare high-variance components is a parsimonious way to generate leptokurtic returns. The mixture mimics fat tails and volatility clustering at the unconditional level. It does not encode seasonality or trend.

Question 16Quantitative finance and machine learning

When training a neural network on financial features, why is early stopping with a temporal validation set preferred over training to convergence on the training loss?

  1. It limits overfitting by halting when validation performance stops improving Correct answer.
  2. It guarantees the global optimum is reached
  3. It removes the need for a learning-rate schedule
  4. It makes the loss function convex

Explanation

Training loss keeps falling as the network memorizes noise, while validation performance peaks and then degrades. Early stopping selects the epoch at that peak. It does not alter convexity or guarantee global optimality.

Question 17Quantitative finance and machine learning

A researcher computes the information coefficient as the cross-sectional correlation between a signal and next-month returns, obtaining a mean IC of 0.04 with a standard deviation of 0.10 over 120 months. What does the information ratio of the IC series approximately equal?

  1. 4.0
  2. 0.04
  3. 0.4 Correct answer.
  4. 0.10

Explanation

The information ratio of the IC series is the mean divided by the standard deviation, 0.04 over 0.10, giving 0.4. Annualizing would multiply by the square root of the periods per year. The raw ratio is a scale-free measure of signal consistency.

Question 18Quantitative finance and machine learning

In a random forest applied to a large set of correlated technical features, which statement about feature importance is correct?

  1. Impurity-based importance is unbiased for correlated predictors
  2. Impurity-based importance can favor high-cardinality or correlated features, so permutation importance on held-out data is often preferred Correct answer.
  3. Permutation importance requires retraining the model for each feature
  4. Feature importance directly measures causal effect on returns

Explanation

Mean decrease in impurity is computed on training data and is biased toward features with many split points and correlated substitutes. Permutation importance measures the drop in out-of-sample performance when a feature is shuffled, avoiding retraining. Importance is predictive, not causal.

Question 19Quantitative finance and machine learning

A trader builds a mean-reversion strategy on a spread and estimates a half-life of 10 days from an Ornstein-Uhlenbeck fit. If the spread is currently two standard deviations from its mean, which statement best reflects the model's implication?

  1. The spread will widen further with probability one
  2. The expected time to revert halfway to the mean is about 10 days, but the path is stochastic Correct answer.
  3. The spread will reach the mean in exactly 10 days
  4. The half-life equals the volatility of the spread

Explanation

In an OU process the half-life is the time for the expected deviation to halve, not a deterministic crossing time. Individual paths remain random and can move further from the mean first. Half-life is a speed parameter, not a volatility measure.

Question 20Quantitative finance and machine learning

A portfolio optimization uses a sample covariance matrix estimated from 100 assets with 250 daily observations. Why might the resulting mean-variance weights be unstable?

  1. The number of observations exceeds the number of assets, which invalidates the estimator
  2. The sample covariance is always singular in this case
  3. The covariance estimate has substantial estimation error and its inverse amplifies that error Correct answer.
  4. Mean-variance optimization ignores expected returns

Explanation

Although 250 observations exceed 100 assets, the sample covariance is still noisy relative to its dimension, and inversion magnifies small eigenvalues. This produces extreme, unstable weights. Shrinkage or factor models regularize the estimate.

Question 21Quantitative finance and machine learning

A binary classifier for credit default has a precision of 0.3 and a recall of 0.8 at a chosen threshold. What does this combination imply?

  1. The classifier is well calibrated by construction
  2. The threshold should be raised to increase recall
  3. It captures most true defaults but many predicted defaults are false alarms Correct answer.
  4. It captures few true defaults but almost all predictions are correct

Explanation

High recall means most actual defaults are flagged, while low precision means a large share of flagged cases are non-defaults. This is a common trade-off when defaults are rare. Raising the threshold generally lowers recall further.

Question 22Quantitative finance and machine learning

A researcher applies a Kalman filter to estimate a time-varying hedge ratio between two assets. The process noise covariance Q is set very small and the measurement noise R is set large. What is the likely effect on the estimated hedge ratio?

  1. The estimate becomes extremely volatile and tracks noise
  2. The filter diverges immediately
  3. The estimate equals the ordinary least squares constant beta
  4. The estimate becomes smooth and slow to adapt to structural change Correct answer.

Explanation

Small Q tells the filter the state changes little, and large R discounts new observations, so the gain is small and updates are gradual. The estimate lags genuine regime shifts. Volatility would result from the opposite configuration.

Question 23Quantitative finance and machine learning

Two machine-learning models are compared using a Diebold-Mariano test on their out-of-sample loss differentials, and the test yields a p-value of 0.03. Which interpretation is most appropriate?

  1. The difference in forecast accuracy is statistically significant at the 5% level, subject to the test's assumptions Correct answer.
  2. The better model will outperform in all future periods
  3. The models are equally accurate because the p-value exceeds 0.01
  4. The test measures whether the models are correctly calibrated

Explanation

The DM test evaluates whether the expected loss differential is zero using the autocovariance structure of the loss series. A p-value of 0.03 rejects equality at 5% but says nothing about future certainty. It also does not assess calibration.

Question 24Quantitative finance and machine learning

A quant trains a deep network to forecast returns and observes that adding more features steadily improves validation performance while the training loss is already near zero. Which explanation is most consistent with this pattern?

  1. The network is underfitting and needs more capacity
  2. The learning rate is too high
  3. The loss function is non-differentiable
  4. Validation leakage through feature construction that uses future information Correct answer.

Explanation

Near-zero training loss with improving validation performance is suspicious in finance because genuine out-of-sample gains are hard. Features built with forward-looking windows or full-sample normalization leak future information into validation. The fix is to construct features using only past data within each split.

Question 25Quantitative finance and machine learning

A portfolio uses a machine-learning alpha model retrained monthly. Backtested Sharpe ratios decline sharply when realistic transaction costs and a one-day execution lag are added. Which change is most likely to recover net performance?

  1. Removing the execution lag to match the original backtest
  2. Shortening the rebalancing interval to capture more signal
  3. Reducing turnover by smoothing the signal and trading only when the forecast change exceeds a threshold Correct answer.
  4. Increasing the model's complexity to raise gross alpha

Explanation

Costs scale with turnover, so smoothing forecasts and applying a no-trade band cuts trading without discarding the core signal. Adding complexity often raises turnover and overfitting. Removing the execution lag would reintroduce an unrealistic assumption.

Question 26Quantitative finance and machine learning

A quant computes the sample autocorrelation of daily returns at lag 1 and obtains a value of -0.02 with a 95% confidence band of roughly plus or minus 0.04 under the white-noise assumption. Which conclusion is most defensible?

  1. The series must be non-stationary because the autocorrelation is negative.
  2. The confidence band is invalid because autocorrelations are always positively biased.
  3. The negative sign proves the series is mean-reverting with a well-defined half-life.
  4. The lag-1 autocorrelation is statistically indistinguishable from zero, so there is no evidence of first-order linear dependence. Correct answer.

Explanation

Under the white-noise null the sample autocorrelation at lag 1 has standard error approximately 1/sqrt(N), and the estimated value lies inside that band. Failing to reject zero dependence at one lag says nothing about the sign being economically meaningful or about stationarity. A single insignificant autocorrelation cannot establish a half-life.

Question 27Quantitative finance and machine learning

In supervised learning for return forecasting, what distinguishes a classification setup from a regression setup?

  1. Classification predicts a discrete label such as sign of return, while regression predicts a continuous target such as the return itself. Correct answer.
  2. Classification needs no train-test split, while regression does.
  3. Classification minimizes mean squared error, while regression minimizes cross-entropy.
  4. Classification requires panel data, while regression requires only a single time series.

Explanation

The target variable type defines the task: discrete class labels call for classification losses such as cross-entropy, and continuous targets call for regression losses such as squared error. The choice of data layout and the need for out-of-sample splitting are independent of this distinction.

Question 28Quantitative finance and machine learning

A quant fits a linear regression of a stock's excess returns on the market's excess returns and obtains a slope of 1.3 and an intercept of 0.1% per month. How should the intercept be interpreted?

  1. As the risk-free rate over the sample period.
  2. As the stock's beta with respect to the market.
  3. As the residual standard deviation of the regression.
  4. As the average monthly excess return not explained by market exposure, i.e. an estimated alpha. Correct answer.

Explanation

In a single-factor market model the slope is beta and the intercept captures average performance relative to the market-adjusted benchmark, conventionally called alpha. The intercept is a return quantity, not a risk measure or a rate taken from outside the regression.

Question 29Quantitative finance and machine learning

A risk analyst wants a model-free estimate of the 95% one-day VaR of a portfolio using 1,000 historical daily P&L observations. Which approach is most consistent with historical simulation?

  1. Take the 50th worst loss in the sorted sample of 1,000 P&L outcomes. Correct answer.
  2. Use the average of all losses exceeding the 99th percentile.
  3. Take the single largest observed loss in the sample.
  4. Fit a normal distribution and use 1.645 times the sample standard deviation.

Explanation

Historical simulation reads the quantile directly off the empirical distribution, and the 5% tail of 1,000 observations corresponds to the 50th worst outcome. Distributional fitting and tail-averaging are different methods, and the maximum loss estimates a far more extreme quantile.

Question 30Quantitative finance and machine learning

A quant builds a signal that is the z-score of each stock's earnings yield within its sector on each date. What is the main purpose of standardizing within sector rather than across the whole universe?

  1. To convert the signal into a probability of outperformance.
  2. To guarantee the signal has no look-ahead bias.
  3. To remove persistent sector-level differences in valuation levels so the signal ranks stocks against comparable peers. Correct answer.
  4. To make the signal independent of the return distribution.

Explanation

Sector-neutral standardization compares each stock with its peer group, stripping out level differences that reflect industry structure rather than relative attractiveness. It does not by itself address look-ahead bias, which depends on data timing, nor does it convert a score into a calibrated probability.

Question 31Quantitative finance and machine learning

Why is the Sharpe ratio of a strategy typically computed from periodic returns and then annualized rather than computed from total cumulative return?

  1. Because volatility is undefined for cumulative returns.
  2. Because annualization removes the effect of transaction costs.
  3. Because cumulative returns cannot be computed for strategies with short histories.
  4. Because the Sharpe ratio is defined as mean excess return divided by return volatility, which requires a return series. Correct answer.

Explanation

The Sharpe ratio scales average excess return by the dispersion of returns, so it is inherently a property of a return series. Annualizing multiplies by the square root of the number of periods under an i.i.d. assumption and does not adjust for trading costs.

Question 32Quantitative finance and machine learning

A quant estimates a covariance matrix and finds one eigenvalue close to zero. What does this imply about the corresponding eigenvector portfolio?

  1. It is the portfolio with the highest expected return.
  2. It is a combination of assets whose return variance is nearly zero, so it is a near-arbitrage direction. Correct answer.
  3. It has the largest variance among all unit-norm portfolios.
  4. It must consist of a single asset with zero volatility.

Explanation

Eigenvalues of a covariance matrix equal the variances of the corresponding unit-norm eigenvector portfolios. A near-zero eigenvalue therefore identifies a direction with almost no risk, which in practice often reflects near-collinearity among assets.

Question 33Quantitative finance and machine learning

In a machine-learning pipeline for financial prediction, what is the role of a validation set separate from the training set?

  1. To select hyperparameters and compare models using data not used to fit the parameters. Correct answer.
  2. To increase the number of observations available for fitting.
  3. To replace the need for a test set entirely.
  4. To provide the final unbiased estimate of live performance.

Explanation

A validation set is used for model selection, such as tuning regularization or choosing among candidate models, while leaving the parameters untouched by that data. The final honest performance estimate should come from a test set that played no role in selection.

Question 34Quantitative finance and machine learning

A quant estimates a GARCH(1,1) model and finds alpha = 0.05 and beta = 0.92. What is the implied unconditional long-run variance relative to the model's constant term omega?

  1. omega divided by 0.03 Correct answer.
  2. omega divided by 0.97
  3. omega multiplied by 0.97
  4. omega divided by 0.05

Explanation

For GARCH(1,1) the unconditional variance is omega/(1 - alpha - beta). With alpha + beta = 0.97 the denominator is 0.03, so persistence is high and the long-run variance is about 33 times omega.

Question 35Quantitative finance and machine learning

A researcher runs a regression of next-month returns on a momentum signal and reports a Newey-West t-statistic of 2.4, while the ordinary t-statistic is 3.1. What is the most likely reason for the difference?

  1. The Newey-West estimator uses fewer observations than the ordinary estimator.
  2. The Newey-West correction accounts for serial correlation and heteroskedasticity, which inflates the ordinary standard error estimate. Correct answer.
  3. The ordinary t-statistic is always biased downward in finance.
  4. The Newey-West correction removes the effect of outliers from the residuals.

Explanation

Newey-West standard errors are robust to autocorrelation and heteroskedasticity and typically exceed the homoskedastic OLS standard errors when residuals are positively autocorrelated. A larger standard error lowers the t-statistic, which is why the robust statistic is smaller here.

Question 36Quantitative finance and machine learning

A quant fits a ridge regression to highly collinear predictors of returns. How does the ridge penalty change the coefficient estimates relative to ordinary least squares?

  1. It increases coefficient magnitudes to offset collinearity.
  2. It shrinks coefficients toward zero, trading some bias for a reduction in variance. Correct answer.
  3. It sets exactly zero coefficients for the least useful predictors.
  4. It leaves coefficients unchanged but reduces the residual variance.

Explanation

Ridge adds an L2 penalty on coefficient size, which shrinks estimates toward zero and stabilizes them under collinearity. The shrinkage introduces bias but can substantially lower estimation variance and improve out-of-sample prediction.

Question 37Quantitative finance and machine learning

A quant uses a Kalman filter to estimate a time-varying beta. The state equation has process noise variance Q and the observation equation has measurement noise variance R. What happens qualitatively as Q increases relative to R?

  1. The filtered beta becomes more responsive to recent observations and more volatile. Correct answer.
  2. The filter stops updating because the model becomes unidentified.
  3. The measurement noise is absorbed into the state estimate, eliminating it.
  4. The filtered beta becomes smoother and closer to its long-run average.

Explanation

The ratio Q/R governs how much the filter trusts new data versus its prior state estimate. A larger process noise signals a faster-changing state, so the Kalman gain rises and the estimate tracks recent observations more closely.

Question 38Quantitative finance and machine learning

A portfolio manager combines two alpha signals with correlation 0.8 and similar individual information coefficients. Why does the combined signal typically have a higher information ratio than either component alone?

  1. Because combining imperfectly correlated signals diversifies their idiosyncratic noise, raising the ratio of signal to noise. Correct answer.
  2. Because correlations below one guarantee a higher Sharpe ratio for any weighting.
  3. Because the information coefficient of the combination equals the sum of the individual coefficients.
  4. Because the combined signal has a higher expected return than the sum of the components.

Explanation

When signals are not perfectly correlated, their noise terms partially cancel while the common predictive component adds, improving the information ratio. The benefit depends on the correlation and on appropriate weighting, not on an automatic return increase.

Question 39Quantitative finance and machine learning

A researcher estimates the probability of default with a logistic model and evaluates it using the area under the ROC curve, obtaining 0.55. What does this indicate?

  1. The model ranks defaulters only slightly better than random chance. Correct answer.
  2. The model's predicted probabilities are well calibrated.
  3. The model has a 55% false positive rate at the optimal threshold.
  4. The model classifies 55% of defaults correctly.

Explanation

The AUC measures rank discrimination and equals 0.5 for a random classifier, so 0.55 reflects very weak separation between defaulters and non-defaulters. AUC says nothing directly about calibration or about accuracy at any particular threshold.

Question 40Quantitative finance and machine learning

A quant computes a 60-day rolling correlation between two assets and notices it jumps from 0.3 to 0.7 in a single day because one very large observation enters and another leaves the window. What is the most appropriate diagnosis?

  1. The estimate is dominated by window composition effects, so a shorter-memory or robust estimator may be preferable. Correct answer.
  2. The true correlation has permanently shifted, so the full-sample estimate should be discarded.
  3. The jump proves the two assets have become cointegrated.
  4. Correlation is undefined when returns are large.

Explanation

Equal-weighted rolling windows give full weight to a single observation for the entire window length, so a large return entering or leaving can cause abrupt jumps. Exponentially weighted or robust estimators reduce this sensitivity. A one-day jump is not evidence of a permanent regime change or of cointegration.

Question 41Quantitative finance and machine learning

In a decision tree trained on financial features, what does the depth of the tree primarily control?

  1. The maximum number of sequential splits, and hence the model's capacity to fit complex interactions. Correct answer.
  2. The learning rate applied to each leaf value.
  3. The fraction of observations used to fit each tree.
  4. The number of features considered at each split.

Explanation

Tree depth bounds how many splits can occur along any path from root to leaf, directly limiting interaction order and model capacity. Feature subsampling, row sampling, and shrinkage are separate hyperparameters.

Question 42Quantitative finance and machine learning

A quant computes the covariance matrix of 200 assets from 150 daily observations and inverts it for mean-variance optimization. What problem arises and how is it commonly mitigated?

  1. The matrix is well conditioned, so the optimizer is unaffected by the sample size.
  2. The matrix is positive definite but biased upward, and adding noise to returns fixes it.
  3. The matrix has too few assets for the sample size, and dropping assets is the only remedy.
  4. The matrix is singular or nearly so, and shrinkage toward a structured target such as a diagonal or factor model stabilizes the inverse. Correct answer.

Explanation

With more assets than observations the sample covariance is singular, and even near that regime its inverse amplifies estimation error. Shrinkage estimators such as Ledoit-Wolf pull the sample matrix toward a structured target, improving conditioning and out-of-sample behavior.

Question 43Quantitative finance and machine learning

A quant fits a hidden Markov model and observes that the transition matrix has diagonal entries of 0.98 and 0.95. What do these entries imply about regime behavior?

  1. Regimes switch almost every period.
  2. Regimes are persistent, with expected durations of roughly 50 and 20 periods respectively. Correct answer.
  3. The model is degenerate because the diagonals exceed 0.9.
  4. The stationary distribution places equal weight on both states.

Explanation

For a two-state chain the expected duration in a state is 1/(1 - p_ii), giving about 50 periods for the first state and 20 for the second. High diagonal probabilities therefore indicate persistent regimes rather than rapid switching.

Question 44Quantitative finance and machine learning

A quant evaluates a strategy with a backtest Sharpe ratio of 2.0 over 5 years of daily data. A colleague argues the result may be inflated by data snooping. Which adjustment is most directly relevant?

  1. Computing the Sharpe ratio on weekly rather than daily returns.
  2. Annualizing the Sharpe ratio using 252 trading days.
  3. Deflating the Sharpe ratio for the number of strategies or parameter combinations tried. Correct answer.
  4. Replacing the Sharpe ratio with the maximum drawdown.

Explanation

When many configurations are searched, the best in-sample Sharpe ratio is a maximum over noisy estimates and is upward biased. Multiple-testing corrections such as deflated Sharpe ratios adjust for the effective number of trials. Annualization and return frequency do not address selection bias.

Question 45Quantitative finance and machine learning

A quant applies a neural network to predict returns and uses dropout during training. What is the primary mechanism by which dropout reduces overfitting?

  1. It permanently removes the least important neurons after training.
  2. It reduces the learning rate as training progresses.
  3. It randomly deactivates units during training, preventing the network from relying on any single pathway. Correct answer.
  4. It normalizes the input features to have zero mean and unit variance.

Explanation

Dropout randomly masks activations on each training step, which forces redundant representations and acts like an ensemble of subnetworks. It is not a pruning step, a learning-rate schedule, or a feature normalization method.

Question 46Quantitative finance and machine learning

A researcher computes the information coefficient for a signal and finds a mean IC of 0.04 with a standard deviation across periods of 0.10 over 240 periods. What is the approximate information ratio of the IC series?

  1. About 0.4 times the square root of 240, roughly 6.2 Correct answer.
  2. About 0.10 divided by 0.04, roughly 2.5
  3. About 0.04 times 240, roughly 9.6
  4. About 0.04 divided by 240, roughly 0.0002

Explanation

The information ratio of the IC series is mean IC divided by its standard deviation, giving 0.4 per period, and annualizing or scaling by the square root of the number of periods gives about 6.2. This measures the consistency of the signal's predictive power over time.

Question 47Quantitative finance and machine learning

A quant trains a model on a panel of stocks and evaluates it with 5-fold cross-validation that randomly assigns stock-date rows to folds. Why can this produce an overly optimistic error estimate?

  1. Random assignment always creates imbalanced class proportions.
  2. Five folds are too few for any reliable estimate.
  3. Rows from the same date or stock can appear in both training and validation folds, leaking common information. Correct answer.
  4. Cross-validation cannot be applied to panel data under any scheme.

Explanation

Random row splitting breaks the dependence structure of panel data, so near-duplicate observations from the same date or firm leak across folds. Grouped or time-aware splits that keep dates or stocks together give a more honest estimate.

Question 48Quantitative finance and machine learning

A quant estimates a value-at-risk model and performs a Kupiec proportion-of-failures test. The observed exception rate is 3.1% against a 1% target over 500 days. What does the test assess?

  1. Whether the magnitude of losses on exception days matches the expected shortfall.
  2. Whether the return distribution is symmetric.
  3. Whether the number of exceedances is consistent with the stated coverage probability. Correct answer.
  4. Whether the VaR estimates are serially independent across days.

Explanation

The Kupiec test is an unconditional coverage test comparing the observed exception frequency with the nominal tail probability. It does not examine the size of tail losses, the clustering of exceptions, or distributional symmetry.

Question 49Quantitative finance and machine learning

A quant fits a gradient boosting model with a very small learning rate and many trees, and observes that training loss keeps falling while validation loss plateaus. Which interpretation is most accurate?

  1. Additional boosting rounds are fitting noise, so early stopping at the validation minimum is appropriate. Correct answer.
  2. The learning rate is too large and should be increased further.
  3. The validation loss plateau proves the features have no predictive power.
  4. The model is underfitting and needs deeper trees or more features.

Explanation

A widening gap between falling training loss and flat or rising validation loss is the classic signature of overfitting during boosting. Early stopping at the validation minimum keeps the model at the point of best out-of-sample performance. Underfitting would show both losses still decreasing.

Question 50Quantitative finance and machine learning

A quant builds an ensemble by averaging many bootstrapped regression trees and finds the out-of-bag error is much lower than the error on a later holdout period. Which explanation is most plausible?

  1. The bootstrap sample size is too small for OOB to be computed.
  2. The OOB estimate shares the same time period as training, so it misses structural change that affects the later holdout. Correct answer.
  3. OOB error is always biased upward relative to holdout error.
  4. Averaging trees cannot reduce variance, so the holdout result must be a data error.

Explanation

Out-of-bag samples are random subsets of the same historical period, so they reflect the training-era relationships rather than later market conditions. When the data-generating process shifts, a forward holdout reveals degradation that OOB cannot detect.