A classic test of your understanding of market efficiency and the foundations of quantitative finance.

The Random Walk Theory posits that security prices move randomly, making it impossible to predict future movements from historical data alone. Price changes are independent and identically distributed — technical analysis cannot consistently outperform the market. The practical implication is that passive, diversified investing is superior to active speculation, a position underpinning the Efficient Market Hypothesis (EMH).

Evaluates your command of statistical significance — foundational to model validation and factor selection.

A p-value quantifies the probability of observing data as extreme as the current sample, assuming the null hypothesis is true. If the p-value falls below the significance level (typically 0.05), we reject the null. In financial modelling, p-values drive model selection and predictor significance — though they must be interpreted alongside effect size and economic intuition to avoid overfitting.

Tests your foundation in portfolio theory and risk diversification concepts.

Covariance measures the degree to which two assets' returns move together. In Markowitz portfolio theory, the covariance matrix is the cornerstone of portfolio optimisation: positively covariant assets amplify portfolio risk, while negatively correlated assets reduce it. Covariance is also central to CAPM, where beta is defined as the covariance of an asset's return with the market divided by market variance.

Assesses practical application of statistical methods in forecasting financial market trends.

Regression establishes a quantitative relationship between a stock price (the dependent variable) and independent variables such as historical prices, volume, or macroeconomic indicators. Two key limitations: non-stationarity of price series requires log-differencing or cointegration techniques; structural breaks mean past relationships may not persist. Factor models (Fama-French) are a widely-used extension regressing excess returns on systematic risk factors.

A fundamental theorem underlying most statistical inference in financial modelling.

The CLT states that the sum of a large number of independent, identically distributed random variables tends toward a normal distribution regardless of the underlying distribution. In finance, this justifies normal-distribution assumptions for portfolio returns, underpins confidence intervals for back-test metrics, and supports the Black-Scholes derivation. Its key limitation is fat tails — individual asset returns exhibit excess kurtosis, making EVT models essential complements.

Probes your ability to apply statistical error theory to real trading decisions.

A Type I error (false positive) means concluding a strategy has alpha when it doesn't. A Type II error (false negative) means missing a genuinely profitable strategy. Rigorous back-testing requires balancing these: lowering significance thresholds reduces Type I errors but increases Type II. Multiple testing corrections (Bonferroni, Benjamini-Hochberg) are critical when evaluating many candidate strategies to avoid data snooping, which dramatically inflates false-positive discovery rates.

Tests knowledge of resampling methods when theoretical distributions are intractable.

The bootstrap generates artificial samples by drawing with replacement from observed data, then computes the statistic of interest for each resample to estimate its sampling distribution. In finance it constructs confidence intervals for Sharpe ratios, estimates IRR distributions, and assesses back-test robustness. Block bootstrapping preserves the autocorrelation structure of financial time series — more appropriate than naive i.i.d. resampling.

A gateway to Bayesian inference, increasingly important in modern quant research.

Bayes' Theorem updates prior probability beliefs given new evidence: P(A|B) = P(B|A) × P(A) / P(B). In finance, Bayesian methods incorporate prior economic knowledge into parameter estimation — particularly valuable when data history is short. Applications include the Black-Litterman portfolio framework, credit risk modelling with sparse default data, and regime-switching models updated recursively as new data arrives.

Critical for any time-series modelling task — non-stationary data violates most regression assumptions.

Stationarity requires constant mean, variance, and autocovariance over time. The Augmented Dickey-Fuller (ADF) test tests the null of unit root; the KPSS test has a null of stationarity — using both provides complementary evidence. Most financial price series are non-stationary (I(1)) but log-returns are typically stationary. Regressing non-stationary series directly risks spurious regression, making differencing or cointegration approaches essential.

Assesses practical application of simulation techniques for VaR estimation and stress testing.

Monte Carlo simulation generates thousands of random portfolio paths by drawing correlated asset returns from estimated distributions. This produces an empirical return distribution from which VaR, Expected Shortfall (CVaR), and tail-risk metrics are extracted. Unlike historical simulation, Monte Carlo incorporates correlation stress scenarios and fat-tail distributions. Key inputs are the covariance matrix (estimated via DCC-GARCH), return distribution assumptions, and simulation count.

Probes your knowledge of dimensionality reduction and risk factor decomposition.

PCA transforms correlated risk factors into orthogonal principal components ordered by variance explained. In fixed income, PCA applied to yield curve movements typically reveals 3 components — level, slope, and curvature — explaining 95%+ of yield curve variance. In equity portfolio management, PCA identifies dominant risk drivers and enables construction that explicitly controls factor exposures, reducing noise and computational burden.

Tests your understanding of volatility clustering — a ubiquitous feature of financial return series.

The GARCH(1,1) — σ²ₜ = ω + αε²ₜ₋₁ + βσ²ₜ₋₁ — captures volatility clustering by modelling conditional variance as a function of lagged squared residuals and lagged conditional variances. Extensions include EGARCH and GJR-GARCH for leverage effects, and DCC-GARCH for dynamic correlation modelling. Applications span options pricing, VaR computation, risk-adjusted return optimisation, and trading signal generation.

A foundational risk metric — expect probing follow-up questions on its failings.

VaR estimates the maximum loss a portfolio is expected to suffer over a given horizon with a specified confidence level (e.g., 99% 1-day VaR). Key limitations: it ignores tail losses beyond the threshold; assumes stationary return distributions; is not sub-additive — violating the coherent risk measure property. Expected Shortfall (CVaR) remedies many of these weaknesses and has become the preferred measure under Basel FRTB.

Tests core portfolio theory understanding applicable to both buy-side and sell-side roles.

Systematic risk arises from macroeconomic factors affecting all assets — rate changes, recessions, geopolitical events. Measured by beta; cannot be eliminated through diversification and must be managed via hedging instruments. Idiosyncratic risk is company-specific — earnings surprises, management changes. It diminishes as portfolio size grows: 30–50 uncorrelated stocks reduces it to near-zero. Multi-factor models (Barra, Axioma) decompose total risk into systematic factor exposures and idiosyncratic residuals.

Evaluates your understanding of the limits of statistical risk models.

Stress testing evaluates portfolio performance under extreme scenarios — historical events (2008 GFC, 2020 COVID crash) or hypothetical shocks (300bp rate spike, 40% equity drawdown). VaR systematically underestimates tail risk because historical data contains limited extreme events and correlations spike during crises. Stress testing directly addresses this by applying correlated shock vectors reflecting crisis dynamics. Regulatory frameworks (CCAR, DFAST) mandate regular stress testing for systemically important institutions.

A universal performance metric — the interviewer is also looking for nuanced critique.

The Sharpe Ratio is excess return per unit of total volatility: (Rp − Rf) / σp. To improve it: add diversifying assets with low correlation, apply tactical signals that scale positions when Sharpe is high, reduce transaction costs, or use volatility targeting to scale risk dynamically. Important caveat: Sharpe penalises upside volatility equally — the Sortino Ratio (downside only) and Calmar Ratio are better for positively skewed strategies.

An advanced topic especially relevant for derivatives desks and investment banks.

Counterparty credit risk is the risk that a derivatives counterparty defaults before settlement. CVA is the market-implied cost: CVA = (1 − Recovery Rate) × ∫ EE(t) × PD(t) dt, where EE(t) is expected exposure and PD(t) is marginal default probability from CDS spreads. CVA must be simulated because exposure is path-dependent. Post-2008, Basel III introduced capital charges for CVA volatility, and DVA/FVA/KVA extend the framework to bilateral adjustments.

Demonstrates awareness of regulatory evolution (Basel FRTB) and coherent risk measures.

While VaR specifies the threshold loss at a confidence level, Expected Shortfall is the average loss in the worst (1−confidence)% of scenarios — the expected loss given VaR is breached. ES is sub-additive (satisfying the coherent risk measure property VaR lacks), captures tail severity, and is more informative for fat-tailed distributions. Basel FRTB replaced 99% VaR with 97.5% ES as the primary market risk capital metric.

Black-Scholes prices a European option under geometric Brownian motion: C = S₀N(d₁) − Ke^(−rT)N(d₂), where d₁ = [ln(S/K) + (r + σ²/2)T] / (σ√T). Key assumptions — constant volatility, no dividends, log-normal returns — are violated in practice. Volatility exhibits a smile, driving extensions like Heston (stochastic vol), Variance Gamma (jump processes), and local volatility models.

The binomial model builds a discrete-time lattice where at each node the price can move up or down with risk-neutral probabilities. Option values are computed by backward induction. Its primary advantages: naturally handles American options (early exercise checked at every node), accommodates discrete dividends, incorporates time-varying volatility. As time steps approach infinity, it converges to Black-Scholes for European options.

The Carr-Madan FFT approach leverages analytically tractable characteristic functions in models like Heston and Variance Gamma. By transforming the pricing problem to the frequency domain, option prices are expressed as integrals over the characteristic function of log-returns, evaluated via the Fast Fourier Transform. This enables rapid calibration of complex models to the volatility surface for real-time trading applications.

Delta (∂C/∂S) is the hedge ratio. Gamma (∂²C/∂S²) measures delta's sensitivity to price — high gamma requires frequent rebalancing. Theta (∂C/∂t) is time decay. Vega (∂C/∂σ) is sensitivity to implied volatility — critical for volatility trading. Rho (∂C/∂r) measures interest rate sensitivity. Delta-gamma-vega hedging is standard desk practice; risk systems monitor all Greeks in real-time against position limits.

The volatility smile describes implied volatility varying across strike prices for options with the same expiry. In equity markets, OTM puts command higher implied vol, reflecting demand for downside protection and empirical negative skew. This directly contradicts Black-Scholes' constant-volatility assumption and motivates stochastic volatility models (Heston, SABR), local volatility models (Dupire), and jump-diffusion models, each calibrated to reproduce the observed surface.

Put-call parity states for European options: C − P = S − Ke^(−rT). This no-arbitrage relationship holds because long call, short put with the same strike replicates a forward contract. If violated, risk-free profit is available. In practice, apparent violations reflect dividend expectations, borrow costs for short selling, and bid-ask spreads. This principle underpins synthetic position construction and option replication strategies in structured product hedging.

A vanilla IRS exchanges fixed cash flows for floating (SOFR-based). The par swap rate is set so PV of fixed payments equals PV of expected floating payments. Procedure: (1) bootstrap the OIS discount curve; (2) derive the forward rate curve to project floating cash flows; (3) discount all legs using OIS. Post-2008 multi-curve frameworks are standard — separate curves for discounting (OIS) and projecting floating rates, as these diverged materially during the credit crisis.

Heston replaces constant volatility with a mean-reverting stochastic variance: dV = κ(θ − V)dt + ξ√V dW_v, correlated with the asset process. Parameters κ (reversion speed), θ (long-run variance), ξ (vol of vol), and ρ (correlation) allow the model to generate a volatility skew. It has a semi-analytical characteristic function enabling fast FFT-based pricing. Compared to Black-Scholes, it better captures the implied vol term structure and models volatility clustering.

Delta hedging holds a quantity of the underlying equal to the option's delta, making the portfolio instantaneously insensitive to small price moves. Challenges include: discrete rebalancing introduces gamma P&L; transaction costs make frequent hedging expensive; hedging at incorrect implied vol misestimates replication cost; and gap risk from overnight price jumps. Sophisticated desks hedge delta-gamma and use vega hedging via vanilla options to manage volatility exposure.

A Markov Chain's future state depends only on the current state. In finance they model credit rating transitions, interest rate regime-switching, and default intensity processes. Hidden Markov Models extend this where the Markov state is unobserved but drives observable returns — used for regime detection (bull/bear/crisis) and tactical asset allocation. The transition matrix eigenstructure determines long-run stationary behaviour and convergence speed.

Two non-stationary I(1) series are cointegrated if a linear combination is stationary, implying a long-run equilibrium. In pairs trading, cointegrated stock pairs are identified, the spread modelled as a mean-reverting Ornstein-Uhlenbeck process, and trades executed when the spread deviates significantly from its mean. Key risks include cointegration breakdown due to regime shifts and synchronisation risk — the spread may diverge further before converging.

Standard Brownian Motion W(t) satisfies: W(0) = 0; increments W(t) − W(s) ~ N(0, t−s); increments are independent; paths are continuous but nowhere differentiable. Geometric Brownian Motion models log-price as a Brownian motion with drift, ensuring non-negative prices. Itô calculus — built on Brownian motion — provides the machinery for derivatives pricing via Itô's lemma, Girsanov's theorem, and the Feynman-Kac formula.

The OU process: dX = κ(θ − X)dt + σ dW. Stationary, Gaussian, and mean-reverts to θ at speed κ. In fixed income it underpins the Vasicek short-rate model. In statistical arbitrage, pairs-trading spreads are modelled as OU processes to derive optimal entry/exit thresholds. The half-life (ln(2)/κ) is a key practical parameter — too slow means sluggish trades, too fast means noise dominates. Maximum likelihood and Kalman filter methods estimate OU parameters from discrete observations.

Itô's Lemma: for f(t, X) where dX = μ dt + σ dW, then df = (∂f/∂t + μ∂f/∂X + ½σ²∂²f/∂X²)dt + σ∂f/∂X dW. The ½σ²∂²f/∂X² Itô correction arises because Brownian increments are order √dt rather than dt. Applied to an option price as a function of the underlying, Itô's Lemma directly yields the Black-Scholes PDE. It is also the starting point for computing Greeks analytically and transforming SDEs into tractable forms.

Jump-diffusion models augment GBM with a Poisson jump process to capture discrete, sudden price movements during earnings, geopolitical events, or crises. Merton (1976) adds normally distributed jumps; Kou (2002) uses a double-exponential jump distribution. Key benefit: jump models generate steep short-maturity implied vol smiles that pure stochastic-vol models struggle to fit. The challenge is that jump risk cannot be perfectly hedged, introducing market incompleteness and non-unique risk-neutral pricing.

ARIMA(p,d,q) combines autoregression, differencing (order d for stationarity), and moving average components. Box-Jenkins methodology — identify (ACF/PACF plots), estimate, diagnose residuals, forecast — is the standard process. In finance: short-horizon macro forecasting (GDP, inflation), yield spread prediction, and ML benchmarking. Its linear nature limits handling of non-linear dynamics, motivating hybrid ARIMA-GARCH specifications that jointly model conditional mean and variance.

The Kalman filter recursively estimates unobserved state variables in a linear Gaussian state-space model, balancing model predictions against noisy observations via predict and update steps. Finance applications: estimating time-varying betas, fitting term structure models, tracking unobserved signals in stat arb, and computing hidden states in stochastic volatility models. Extended and Unscented Kalman Filters extend the approach to non-linear systems.

CAPM: E(R) = Rf + β(E(Rm) − Rf). Applied to compute required returns for capital budgeting, identify mispriced securities via alpha, construct market-neutral portfolios (beta = 0), and decompose returns into systematic beta contribution and manager skill. CAPM's empirical failures — explaining only ~70% of cross-sectional return variation — motivated multi-factor models: Fama-French (value, size), Carhart (momentum), and Quality factor.

MVO selects weights that minimise variance for a given expected return: w* = (1/λ)Σ⁻¹(μ − rf·1). Practical limitations: extreme sensitivity to return estimates ("error maximisation"); static covariance assumptions; ignoring higher moments. Practitioners address these through Ledoit-Wolf shrinkage for Σ, Black-Litterman for return estimates, position constraints, and alternative risk measures (CVaR optimisation, risk parity).

Kelly determines the optimal fraction of capital to maximise long-run geometric growth. For a binary bet: f* = (pb − (1−p)) / b. The multi-asset Kelly solution is proportional to Σ⁻¹μ — identical to MVO with log utility. Full Kelly produces extreme volatility and catastrophic drawdowns under estimation error. Practitioners use fractional Kelly (commonly half-Kelly), sacrificing some growth for significantly lower variance.

Fama-French augments CAPM with SMB (Small Minus Big — size premium) and HML (High Minus Low — value premium): R − Rf = α + βm(Rm − Rf) + βs·SMB + βv·HML + ε. Extended to five factors (2015) with Profitability (RMW) and Investment (CMA). These style factors are the backbone of systematic equity factor investing — long/short implementations generate significant information ratios when properly constructed and risk-adjusted.

Risk parity allocates capital so each asset class contributes equally to total portfolio risk. Traditional 60/40 portfolios have 90%+ risk concentrated in equities. Risk parity significantly overweights less-volatile assets and uses leverage on low-volatility components. The construction minimises ∑(RCi − RCj)² where RCi = wi·(Σw)i. Core critique: leverage amplifies losses when bond-equity correlation turns positive (as in 2022), eliminating diversification benefits.

Black-Litterman is a Bayesian framework combining market equilibrium returns (implied by market-cap weights via reverse optimisation) with investor views, producing a posterior expected return vector. Starting from equilibrium rather than arbitrary forecasts dramatically reduces error-amplification. An investor expresses views with confidence levels — BL blends these with equilibrium, producing intuitive tilts rather than concentrated positions. The standard approach for systematic incorporation of analyst forecasts in institutional asset management.

Overfitting captures noise rather than signal — fitting idiosyncratic patterns that don't generalise. In finance this is pervasive: strategies back-tested on historical data often fail dramatically out-of-sample. Countermeasures: strict train-validation-test splits; walk-forward cross-validation; regularisation (Ridge, LASSO); Bayesian model comparison; and the deflated Sharpe ratio, which adjusts for multiple testing when selecting the best strategy from many candidates.

Random Forests build many decision trees on bootstrapped subsamples with random feature subsets and aggregate predictions. In finance: equity return prediction using fundamental, technical, and alternative data features; credit default prediction; market microstructure classification. Key advantages: handles non-linearities and interactions, robust to irrelevant features, built-in feature importance. Key limitation: requires purged k-fold cross-validation per Lopez de Prado's methodology to avoid look-ahead bias in time-series applications.

LASSO (L1 penalty) minimises ∑(yi − Xβ)² + λ∑|βj|; Ridge uses ∑βj² as penalty. LASSO's absolute-value penalty creates a non-differentiable constraint forcing some coefficients exactly to zero, producing sparse solutions ideal for variable selection in high-dimensional factor models. Ridge shrinks all coefficients but retains all variables. In practice, LASSO identifies the active set of predictors from hundreds of candidate factors; Elastic Net (combining both) handles highly correlated predictors.

Text-based alpha sources include earnings call transcripts, news sentiment processed with BERT-family models, social media, and SEC filings (10-K risk factor changes). Key NLP techniques: Named Entity Recognition, domain-specific fine-tuned transformers, and BERTopic for emerging theme tracking. Alternative data beyond text includes credit card transactions, satellite imagery of retail locations and oil tanks, and app usage data. Critical challenges: establishing causality, avoiding look-ahead bias, and data vendor reliability.

Expected prediction error decomposes into bias² + variance + irreducible noise. Too-simple models produce high bias but low variance. Too-complex models produce low bias but high variance (overfitting). In finance, with low signal-to-noise ratios and non-stationarity, high variance is often more damaging: an overfit model's out-of-sample Sharpe collapses. This explains why regularised linear models and tree ensembles often outperform deep neural networks in tabular financial prediction tasks.

RL trains an agent to maximise cumulative reward through trial and error. In trading, state = portfolio positions + market features; actions = buy/sell/hold; rewards = risk-adjusted returns. Suited to sequential decision-making: optimal execution, dynamic hedging, and portfolio rebalancing with transaction costs. Key algorithms: Q-learning, DDPG, PPO. Challenges: non-stationarity of financial environments, sparse rewards, overfitting to historical regimes, and the difficulty of simulation-to-live transfer as strategy scale changes the environment.

I would approach this in stages: (1) EDA — statistical summaries, distribution checks, time-series plots; (2) feature engineering — transaction frequency per account, rolling z-scores, velocity measures, and graph features; (3) anomaly detection — Mahalanobis distance, Isolation Forest, Autoencoder reconstruction error, DBSCAN, and XGBoost (with SMOTE for class imbalance) if labelled data is available; (4) model interpretation via SHAP values to explain flagged transactions for compliance review.

A volatility forecasting model showed anomalously low realised-to-forecast ratios during market stress. Rather than dismissing the discrepancy, I isolated whether the issue was data (feed errors, adjusted vs unadjusted prices), coding logic (incorrect weighting, look-ahead bias), or model specification (regime change the model couldn't adapt to). The root cause was survivorship bias — delisted stocks had been silently dropped, biasing the model toward lower-volatility firms. The fix involved reconstructing the training set with proper point-in-time data.

For academic research: SSRN's quantitative finance section, Journal of Portfolio Management, Journal of Financial Economics, and AQR/Man Institute practitioner research. For technical methodology: arXiv q-fin and stat.ML sections, and following researchers like Marcos Lopez de Prado. For market developments: risk.net for derivatives, Bloomberg Intelligence, and publicly available research from Two Sigma. I also implement new techniques in personal projects — building things from scratch is the most reliable way to identify whether a published method is genuinely robust or relies on overly favourable assumptions.

💡

Finance firms — from hedge funds to investment banks — increasingly offer compressed schedules as retention tools for high-value analytical talent. Understanding the benefits, trade-offs, and how to make the case for a shorter workweek gives you a meaningful edge when evaluating offers and negotiating your next role.