Skip to main content
This page collects interview questions on statistical modeling, time series and backtesting. The key test is whether you spot the common traps: look-ahead bias, overfitting, multiple testing and non-stationarity.
Status of this page. Last checked on 15 November 2026. These are standard econometrics and quantitative research frameworks, not any firm’s specific methods.

Regression

  • Linearity and correct specification: Omitted variables or the wrong functional form bias the coefficients.
  • Exogeneity (errors uncorrelated with regressors): If it fails (endogeneity), coefficients are biased and inconsistent. This is the most serious problem.
  • Homoskedasticity: If it fails, coefficients remain unbiased but standard errors are wrong. Use robust standard errors.
  • No autocorrelation: Often fails in time series and also affects standard errors. Use Newey-West (HAC) corrections.
  • No perfect multicollinearity: If it fails, coefficients are not identified.
The key distinction: heteroskedasticity and autocorrelation affect standard errors (inference), while endogeneity affects the coefficients themselves.
It means the explanatory variables are highly correlated. Consequences: coefficient estimates have higher variance, are sensitive to small data changes and are hard to interpret individually, but overall predictive power is usually unaffected.Detection: VIF and the condition number. Remedies: drop or combine variables, use PCA, or use regularization such as ridge. If the goal is prediction rather than interpreting coefficients, the impact is smaller.
  • Ridge (L2): Shrinks all coefficients toward zero but never exactly to zero. Good with many small, correlated effects, and it stabilizes collinearity.
  • Lasso (L1): Can set some coefficients exactly to zero, which performs variable selection. Good for sparse signals, but selection is unstable among highly correlated variables.
  • Elastic net: A mix of the two.
Both trade a little bias for lower variance (the bias-variance tradeoff). Choose the penalty strength by cross-validation, and for time series use splits that do not shuffle.

Time series

Weak stationarity means the mean and variance do not change over time and the autocovariance depends only on the lag. Most statistical inference, including regression and forecasting, assumes stationarity.Regressing one random walk (non-stationary) on another independent random walk often gives a high R² and significant coefficients that mean nothing. This is spurious regression. The fix is to model returns or differenced series, or to test for unit roots (for example the ADF test).
If some linear combination of two non-stationary series is stationary, they are cointegrated. The spread fluctuates around a long-run equilibrium and tends to revert after deviations.Tests: Engle-Granger (regress, then run a unit root test on the residual) or Johansen. Application: trade the spread rather than a single price, such as a stock against an index or an ETF against its constituent basket.Risk: the cointegrating relationship can break (structural change), so monitor the spread’s half-life and the stability of the relationship, and use stop losses. Estimation error and the choice of sample period can make a backtest look better than reality.
  • Close-to-close: The sample standard deviation of returns. Simple, but ignores intraday information, and the window length is a tradeoff.
  • EWMA: σ²_t = λ σ²_{t−1} + (1 − λ) r²_{t−1}. Recent observations get more weight so it reacts quickly, with λ usually close to 1.
  • GARCH(1,1): σ²_t = ω + α r²_{t−1} + β σ²_{t−1}. It captures volatility clustering and mean reversion to a long-run level. EWMA is a special case (ω = 0, α + β = 1).
There are also range-based estimators using intraday high and low (for example Parkinson and Garman-Klass) that are more efficient than close-to-close, but they are sensitive to jumps and discrete quotes.
  • VaR(α): The loss threshold that is not exceeded at confidence level α. It says nothing about how bad losses are beyond the threshold.
  • Expected Shortfall: The average loss given that the loss exceeds VaR.
ES is a coherent risk measure (it is subadditive), while VaR is not always, so with diversification VaR can suggest risk increases. ES is more sensitive to the tail shape but harder to backtest than VaR.

Backtesting and research process

  • Look-ahead bias: Using data that was not known at the time. Use point-in-time data and confirm that the signal time precedes the trade time.
  • Survivorship bias: Including only names that still exist. Use data that includes delisted securities.
  • Data snooping and overfitting: Tuning repeatedly on the same data. Hold out out-of-sample data and use walk-forward validation.
  • Understated costs and market impact: Use realistic spreads, fees, borrow costs and tradable volume.
  • Corporate action errors: Dividends, splits and ticker changes must be adjusted correctly.
Not necessarily. With multiple testing, even if none has a real effect, about 5 of 100 strategies are expected to be significant at the 5% level, so a maximum t well above 2 is unsurprising.Responses:
  • Adjust the threshold (Bonferroni, or Benjamini-Hochberg to control the false discovery rate)
  • Compute a deflated Sharpe ratio that accounts for the number of trials
  • Validate on genuinely unused out-of-sample data
  • Require an economic rationale, not just statistical significance
Under Bonferroni with 100 tests, the per-test threshold is about 0.05 / 100 = 0.0005, which corresponds to a t-statistic of roughly 3.5.
Annualizing a daily Sharpe by multiplying by √252 assumes i.i.d. returns. Limits:
  • With autocorrelation, √252 scaling over- or understates the annual figure.
  • It ignores skew and kurtosis. A strategy that sells options can show a high Sharpe while carrying crash tail risk.
  • Estimates are noisy on short samples and inflated by multiple trials.
  • It can be manipulated: smoothed valuations and illiquid assets understate volatility.
Complementary measures: Sortino, maximum drawdown, Calmar and tail risk measures.
You cannot shuffle randomly, otherwise future information leaks into the training set.
  • Walk-forward or expanding window: Roll forward in time, train on the past and test on what follows.
  • Purged and embargoed CV: When labels span a period (for example a 5-day forward return), remove samples near the train-test boundary that overlap, and add an embargo period.
Also check the time alignment of features and labels so nothing uses data that was not yet available at time t.
It is an eigendecomposition of the return covariance (or correlation) matrix. The first principal component is usually the market factor (all stocks moving together), and later components often correspond to sector or style factors.Uses: dimension reduction, identifying risk factors, building factor hedges, and cleaning the covariance matrix. Limits: components are statistical constructs without guaranteed economic meaning, they are sensitive to scaling and sample period, and when the number of observations is small relative to the number of stocks, the covariance estimate is very noisy.