CQF key terms

The chain rule for functions of a stochastic process driven by Brownian motion. It adds a second-order correction term, producing the drift adjustment seen in the Black-Scholes derivation.

A continuous-time model in which log returns are normally distributed and prices stay positive. Its constant proportional volatility makes it the textbook baseline for equity prices.

Valuing a claim as the discounted expectation of its payoff under a measure where every tradable asset earns the risk-free rate. Requires no-arbitrage and complete markets.

A process whose conditional expected future value equals its current value. In finance, discounted asset prices are martingales under the pricing measure.

A result describing how changing the probability measure alters the drift of a Brownian motion while leaving volatility unchanged. It formalizes the move from real-world to risk-neutral dynamics.

A closed-form option pricing framework assuming lognormal prices, constant volatility and continuous hedging. Its delta-hedging replication argument underpins modern derivatives pricing.

The volatility input that makes a pricing model reproduce an observed market price. Quoting options by implied volatility rather than price standardizes across strikes and maturities.

The pattern in which implied volatility varies with strike, typically higher for deep out-of-the-money options. It signals that lognormal constant-volatility assumptions are too simple.

A deterministic function of price and time chosen so the model reprices all observed European options exactly. It captures the smile but usually misprices forward-smile dynamics.

A modeling approach where variance itself follows a random process, often correlated with the asset. It generates realistic smile dynamics and volatility clustering.

Sensitivities of an option's value to underlying factors: delta to price, gamma to delta, vega to volatility, theta to time and rho to rates. Used for hedging and risk limits.

Offsetting an option's directional exposure by trading the underlying in proportion to delta. In continuous time under the model assumptions it replicates the option payoff.

Estimating prices or risk measures by averaging outcomes over many simulated random paths. Accuracy improves with sample size at a rate proportional to the inverse square root.

Techniques such as antithetic variates, control variates and importance sampling that lower the sampling error of a Monte Carlo estimate without increasing path count.

A quantile of the loss distribution over a stated horizon and confidence level. It is simple to report but ignores the size of losses beyond the threshold.

The average loss conditional on exceeding the Value at Risk threshold. It is a coherent risk measure, rewarding diversification, unlike VaR.

Estimating a risk distribution by reapplying past observed factor returns to the current portfolio. It avoids distributional assumptions but depends heavily on the sample window.

Comparing model-generated risk forecasts with realized outcomes to check calibration, for example counting VaR breaches against the expected rate. Persistent failures trigger model review.

Deriving zero-coupon discount factors sequentially from prices of coupon bonds or swaps, each maturity solved using previously determined shorter rates.

The interest rate agreed now for borrowing or lending over a future period. It is implied by the ratio of discount factors at the period's start and end dates.

The present value today of one unit of currency received at a future date. It encodes the entire term structure and converts future cash flows into present terms.

A linear transformation finding orthogonal directions of maximal variance in data. In rates it typically shows level, slope and curvature as the dominant curve factors.

Factoring a square matrix into eigenvectors and eigenvalues, revealing directions that scale rather than rotate under the transformation. Central to PCA and covariance analysis.

A symmetric matrix of pairwise covariances between variables. It must be positive semi-definite, a property portfolio optimization and simulation routines rely on.

Factoring a positive-definite matrix into a lower triangular matrix times its transpose. Commonly used to generate correlated random draws in Monte Carlo simulation.

Rescaling inputs, for example to unit variance or a fixed range, so that no variable dominates distance or regularization. Essential before PCA and penalized regression.

Adding a penalty on model complexity, such as L1 or L2 norms on coefficients, to reduce overfitting. The strength is chosen by cross-validation.

Least squares with an L2 penalty on coefficients, shrinking them toward zero and stabilizing estimates when predictors are collinear. It rarely sets coefficients exactly to zero.

Regression with an L1 penalty that both shrinks coefficients and forces some exactly to zero, performing variable selection. Useful for sparse factor or signal construction.

Fitting noise in the training sample so that out-of-sample performance degrades. Detected by comparing training and validation error and mitigated by simpler models or more data.

Splitting data into folds, training on some and validating on others, then averaging. It estimates generalization error and tunes hyperparameters; time series need blocked splits.

Fitting a mapping from inputs to labeled targets using examples. In finance, labels are often noisy returns or default indicators, making validation design critical.

Predicting a discrete category, such as default or rating downgrade, rather than a continuous value. Performance is judged with confusion matrices and related scores.

A linear model passing a weighted sum through the logistic function to output probabilities. Coefficients are estimated by maximum likelihood and are interpretable as log-odds effects.

A plot of true positive rate against false positive rate across classification thresholds. The area beneath it summarizes ranking ability independent of any single cutoff.

An ensemble of decision trees each trained on bootstrapped samples and random feature subsets, then averaged. It reduces variance relative to a single tree.

Sequentially adding weak learners, each fitted to the residual errors of the current ensemble. Strong predictive power but sensitive to learning rate and early stopping.

A layered composition of linear transformations and nonlinear activations whose weights are learned by gradient descent on a loss function. Depth enables complex feature interactions.

Computing gradients of a loss with respect to every network weight by applying the chain rule backward through the layers. It makes gradient-based training computationally feasible.

An optimization method updating parameters using gradients from small random data batches. Mini-batches trade gradient noise for speed and help escape poor regions.