Quantitative Methods for Finance

Random variable is a numerical outcome of a stochastic experiment. In finance it represents quantities such as asset returns, interest rates, or default times. For example, the daily return of a stock can be modeled as a random variable X with a certain probability distribution. Understanding the nature of X is the first step in any quantitative analysis because all subsequent measures—expected value, variance, risk metrics—are derived from its distribution.

The probability distribution describes how probabilities are assigned to the possible values of a random variable. Common distributions used in finance include the normal, log‑normal, Student‑t, and exponential families. The normal distribution is often assumed for short‑term returns due to the Central Limit Theorem, while the log‑normal distribution underlies the geometric Brownian motion model for asset prices because it guarantees positivity.

Expected value, denoted E[X], is the long‑run average of a random variable if the experiment were repeated infinitely. In portfolio theory, the expected return of a portfolio is the weighted sum of the expected returns of its constituent assets. For a simple two‑asset portfolio with weights w1 and w2, the portfolio expected return μp = w1·μ1 + w2·μ2. This linearity is crucial for constructing efficient frontiers.

The variance measures the dispersion of a random variable around its expected value. It is defined as Var(X) = E[(X – E[X])²]. In finance, variance is the foundational component of risk; higher variance implies greater uncertainty about future outcomes. The square root of variance, the standard deviation, is often reported as “volatility” for asset returns.

Covariance captures how two random variables move together and is defined as Cov(X,Y) = E[(X – E[X])(Y – E[Y])]. If Cov(X,Y) > 0, the variables tend to rise together; if negative, they move in opposite directions. In portfolio construction, the covariance matrix of asset returns is essential for evaluating portfolio variance: Σ²p = wᵀΣw, where Σ is the covariance matrix and w the weight vector.

Correlation is the normalized version of covariance: Ρ(X,Y) = Cov(X,Y) / (σX·σY). Correlation values lie between –1 and 1 and provide a dimensionless measure of linear relationship. Practitioners often use correlation to assess diversification benefits; assets with low or negative correlation can reduce overall portfolio risk.

A stochastic process is a collection of random variables indexed by time, {X(t), t ≥ 0}. It models the evolution of financial quantities over continuous or discrete time. The most important stochastic process in quantitative finance is Brownian motion (also called Wiener process), denoted W(t). Brownian motion has independent, normally distributed increments with mean 0 and variance proportional to the time step: W(t+Δt) – W(t) ~ N(0,Δt).

Geometric Brownian motion (GBM) extends Brownian motion to model asset prices that must remain positive. The dynamics are given by dS(t) = μS(t)dt + σS(t)dW(t), where μ is the drift (expected return) and σ the volatility. Solving this stochastic differential equation (SDE) yields S(t) = S(0)·exp[(μ – ½σ²)t + σW(t)], which is log‑normally distributed. GBM underlies the classic Black‑Scholes option pricing model.

Itô’s lemma is the chain rule for stochastic calculus. It allows us to find the differential of a function of a stochastic process. If Y(t) = f(S(t),t) and S(t) follows an SDE, then dY = (∂f/∂t + μS∂f/∂S + ½σ²S²∂²f/∂S²)dt + σS∂f/∂S dW. This result is fundamental for deriving the Black‑Scholes PDE and for transforming SDEs into more tractable forms.

A martingale is a stochastic process whose conditional expected future value equals its current value, given all past information. Formally, E[X(t+Δt) | ℱt] = X(t). In risk‑neutral pricing, discounted asset prices become martingales under the risk‑neutral measure, ensuring no arbitrage opportunities.

Risk‑neutral measure (also called equivalent martingale measure) is a probability measure Q under which the present value of any tradable asset equals the expected discounted payoff under Q. Switching from the real‑world measure P to Q simplifies pricing because we can ignore risk premiums; we only need the risk‑free rate and the asset’s payoff distribution under Q.

Arbitrage refers to a trading strategy that yields a sure profit with zero net investment and no risk. The Fundamental Theorem of Asset Pricing states that a market is arbitrage‑free if and only if there exists an equivalent martingale (risk‑neutral) measure. Detecting arbitrage opportunities is a core challenge in model validation.

The pricing kernel (or stochastic discount factor) is a random variable M that links payoffs to present values: Price = E[M·payoff] under the real‑world measure. In complete markets, M = e^{-rt}·(dQ/dP), linking the pricing kernel to the Radon–Nikodym derivative between Q and P.

The Black‑Scholes model provides a closed‑form solution for European call and put options on non‑dividend‑paying stocks. The formula C = S₀N(d₁) – Ke^{-rT}N(d₂) uses the cumulative normal distribution N(·) and parameters d₁, d₂ that depend on S₀, K, r, σ, and T. The model assumes GBM dynamics, constant volatility, and frictionless markets. Despite its simplifications, it remains a benchmark for understanding option pricing and for calibrating more sophisticated models.

Monte Carlo simulation is a numerical technique that approximates expectations by averaging many random sample paths. For option pricing, we simulate thousands of price paths under the risk‑neutral dynamics, compute the payoff for each path, discount back, and average. The law of large numbers guarantees convergence as the number of simulations N → ∞, with error decreasing at rate 1/√N.

To improve efficiency, practitioners employ variance reduction methods. Antithetic variates generate paired paths using opposite random draws (e.G., +Z and –Z) to cancel out random noise, reducing variance without extra computational cost. Control variates exploit a known‑expectation variable that is highly correlated with the target payoff; by adjusting the estimator, variance is lowered. Quasi‑Monte Carlo replaces pseudo‑random numbers with low‑discrepancy sequences (Sobol, Halton) that fill the space more uniformly, accelerating convergence for high‑dimensional problems.

When the underlying asset follows a partial differential equation (PDE), finite difference methods provide a grid‑based numerical solution. For the Black‑Scholes PDE, we discretize time and asset price dimensions, replace derivatives with finite differences, and march backwards from the terminal payoff. Stability and convergence depend on the choice of explicit, implicit, or Crank‑Nicolson schemes.

Binomial tree models asset price evolution in discrete time with two possible moves per step: Up (u) or down (d). By calibrating u, d, and the risk‑neutral probability p to match the first two moments of the continuous GBM, the tree converges to the Black‑Scholes price as the number of steps increases. Binomial trees are especially useful for American options because they naturally incorporate early exercise decisions at each node.

A partial differential equation (PDE) arises when applying Itô’s lemma to a derivative payoff and enforcing the martingale condition under the risk‑neutral measure. The Black‑Scholes PDE, ∂V/∂t + ½σ²S²∂²V/∂S² + rS∂V/∂S – rV = 0, must be solved with appropriate boundary conditions to obtain the option price V(S,t). PDE techniques are central to pricing a wide range of derivatives, including exotic options and interest‑rate products.

The Fokker‑Planck equation (also called the forward Kolmogorov equation) describes the time evolution of the probability density function of a stochastic process. For GBM, the density satisfies a diffusion equation whose solution is the log‑normal distribution. Understanding the Fokker‑Planck framework helps in calibrating models to observed market densities and in constructing transition probability matrices for discrete‑time approximations.

In many markets, asset returns exhibit sudden jumps not captured by pure diffusion. Jump‑diffusion models augment GBM with a Poisson jump component: DS/S = (μ – λκ)dt + σdW + J dN, where N is a Poisson process with intensity λ and J is the random jump size with mean κ. The Merton model is a classic example, providing analytic option pricing formulas that incorporate jump risk.

A Poisson process counts the number of events occurring in a fixed interval when events happen independently and at a constant average rate λ. In finance, Poisson processes model default events, order arrivals, or macro‑economic shocks. The inter‑arrival times are exponentially distributed with mean 1/λ.

Lévy processes generalize Brownian motion and Poisson processes by allowing both continuous and jump components with potentially infinite activity. Popular Lévy models include the Variance Gamma, CGMY, and Normal Inverse Gaussian processes. These models produce heavier tails and skewness, matching empirical asset return distributions more closely than Gaussian assumptions.

Volatility itself is time‑varying. ARCH (Autoregressive Conditional Heteroskedasticity) models specify that the variance of a series at time t depends on past squared observations: Σ²t = α₀ + α₁ε²t‑1. GARCH (Generalized ARCH) extends this by adding lagged variance terms: Σ²t = α₀ + α₁ε²t‑1 + β₁σ²t‑1. GARCH models capture volatility clustering – periods of high volatility followed by high volatility – a hallmark of financial time series.

Historical volatility is computed from past asset returns, typically as the annualized standard deviation over a rolling window. It is used for risk management and as an input to pricing models when an implied volatility surface is unavailable. In contrast, implied volatility is the volatility parameter that, when plugged into a pricing model (e.G., Black‑Scholes), reproduces the observed market price of an option. Implied volatility reflects market expectations and risk premia.

Risk management uses metrics such as Value‑at‑Risk (VaR), which quantifies the maximum expected loss over a given horizon at a specified confidence level. For a portfolio with loss distribution L, VaR₍α₎ satisfies P(L ≤ VaR₍α₎) = α. VaR can be estimated parametricly (assuming normality), via historical simulation (re‑ordering past losses), or through Monte Carlo simulation of the portfolio’s stochastic dynamics.

Conditional Value‑at‑Risk (CVaR), also known as Expected Shortfall, measures the average loss beyond the VaR threshold: CVaR₍α₎ = E[L | L > VaR₍α₎]. CVaR is coherent, addressing some of VaR’s shortcomings (e.g., lack of subadditivity). It is increasingly used in regulatory frameworks such as Basel III.

Stress testing and scenario analysis complement VaR by evaluating portfolio performance under extreme but plausible market conditions. A stress test might assume a 30 % equity market drop, a sudden spike in credit spreads, or a sharp currency devaluation. Scenario analysis helps identify hidden concentrations and tail risk not captured by statistical measures.

Dimensionality reduction techniques are valuable when dealing with large factor models. Principal Component Analysis (PCA) transforms a set of correlated variables into orthogonal components that capture most of the variance. In fixed‑income markets, the first few principal components often correspond to level, slope, and curvature of the yield curve, enabling parsimonious modeling of interest‑rate dynamics.

The eigenvalue associated with each principal component measures the amount of variance explained. Selecting components with eigenvalues above a certain threshold (e.G., Kaiser’s rule) yields a reduced factor set that retains the majority of information while simplifying calibration and simulation.

In time‑series modeling, stationarity is a key property: A stationary process has constant mean, variance, and autocorrelation structure over time. Many statistical tests, such as the Augmented Dickey‑Fuller (ADF) test, assess whether a series contains a unit root (non‑stationary) or is stationary. Non‑stationary series often require differencing or transformation before applying models like ARMA.

Cointegration occurs when a linear combination of two or more non‑stationary series is stationary. This concept underlies pairs trading strategies: Two assets whose price spread is mean‑reverting can be traded on the basis that deviations will revert, generating profit. The Engle‑Granger two‑step method and Johansen test are standard tools for detecting cointegration.

State‑space models provide a flexible framework for representing dynamic systems with hidden (latent) states. The Kalman filter recursively estimates the hidden state vector given noisy observations, assuming linear dynamics and Gaussian noise. In finance, Kalman filtering is used to estimate time‑varying parameters such as stochastic volatility, term‑structure factors, or unobservable risk premiums.

When the linear‑Gaussian assumptions are violated, the Extended Kalman Filter (EKF) or the Unscented Kalman Filter (UKF) can handle nonlinearities. However, for highly nonlinear or multimodal problems, Particle filters (Sequential Monte Carlo) approximate the posterior distribution with a set of weighted particles, providing a non‑parametric solution.

Parameter estimation techniques include Maximum Likelihood Estimation (MLE), which finds the parameter values that maximize the likelihood of observed data under a chosen model. For instance, fitting a GARCH(1,1) model involves maximizing the likelihood of the observed return series given the conditional variance recursion. MLE yields asymptotically efficient estimates under regularity conditions.

Bayesian inference treats parameters as random variables with prior distributions, updating them with data via Bayes’ theorem to obtain posterior distributions. Markov Chain Monte Carlo (MCMC) methods, such as the Metropolis‑Hastings algorithm and Gibbs sampling, generate draws from complex posteriors. Bayesian approaches provide full uncertainty quantification, enabling credible intervals for forecasts and risk measures.

Markov chains describe stochastic processes with the memoryless property: The future state depends only on the current state, not on the path taken to arrive there. Transition probability matrices govern the evolution. In credit risk, a discrete‑time Markov chain models rating migrations, where each rating class is a state and transition probabilities are estimated from historical data.

Markov Chain Monte Carlo (MCMC) leverages Markov chains to sample from high‑dimensional probability distributions. By constructing a chain whose stationary distribution equals the target posterior, MCMC approximates expectations that are analytically intractable. Convergence diagnostics (e.G., Gelman‑Rubin statistic) assess whether the chain has reached equilibrium.

Regression analysis is a fundamental tool for modeling relationships between variables. The ordinary least squares (OLS) estimator minimizes the sum of squared residuals, providing unbiased and efficient estimates under the Gauss‑Markov assumptions (linearity, exogeneity, homoskedasticity, no autocorrelation). In finance, regression underlies factor models such as the Capital Asset Pricing Model (CAPM) and Fama‑French three‑factor model.

Multicollinearity occurs when explanatory variables are highly correlated, inflating variance of coefficient estimates and making inference unstable. Variance Inflation Factor (VIF) diagnostics detect multicollinearity; remedial actions include dropping variables, combining them, or applying principal component regression.

Heteroskedasticity violates the constant variance assumption, leading to inefficient OLS estimates and biased standard errors. The Breusch‑Pagan and White tests detect heteroskedasticity. Robust (Huber‑White) standard errors provide consistent inference without correcting the model structure, while generalized least squares (GLS) reweights observations to restore efficiency.

Autocorrelation (serial correlation) in residuals indicates model misspecification, especially in time‑series regressions. The Durbin‑Watson statistic tests for first‑order autocorrelation; the Ljung‑Box Q‑test assesses higher‑order dependence. If autocorrelation is present, one may adopt ARIMA models, incorporate lagged dependent variables, or use Newey‑West adjusted standard errors.

Model selection balances goodness‑of‑fit against complexity. Information criteria such as the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) penalize the log‑likelihood by the number of parameters, discouraging over‑parameterization. Cross‑validation, where the data set is split into training and validation subsets, provides an empirical assessment of out‑of‑sample performance.

Overfitting occurs when a model captures noise rather than the underlying signal, resulting in poor predictive performance on new data. Techniques to mitigate overfitting include regularization (Ridge, Lasso), dimensionality reduction (PCA), and pruning in tree‑based methods. In financial contexts, overfitting is especially dangerous because backtested strategies may appear profitable but fail in live trading.

Bootstrap resampling creates many pseudo‑samples by randomly drawing with replacement from the original data. It estimates the sampling distribution of a statistic (e.G., Mean, regression coefficient) without relying on parametric assumptions. Bootstrapped confidence intervals are widely used for risk measures like VaR, where analytical distributions are unknown.

In high‑frequency finance, the microstructure noise arising from bid‑ask bounce, discrete price ticks, and order‑flow dynamics contaminates price series. Estimators such as the realized variance must be adjusted (e.G., Using subsampling or two‑scale estimators) to obtain unbiased volatility estimates.

Stochastic volatility models treat volatility itself as a latent stochastic process, often correlated with the asset price. The Heston model specifies dynamics: DS = μSdt + √v S dW₁, dv = κ(θ – v)dt + ξ√v dW₂, with correlation ρ between dW₁ and dW₂. Closed‑form solutions exist for European options via characteristic functions and Fourier inversion, enabling fast calibration to implied volatility surfaces.

Local volatility models, such as the Dupire framework, infer a deterministic volatility surface σ_loc(S,t) that exactly reproduces observed option prices under risk‑neutral dynamics. The local volatility function is derived by differentiating market option prices with respect to strike and maturity. While locally consistent, the model may generate unrealistic dynamics for path‑dependent derivatives.

In credit risk, the structural approach (Merton model) views default as the event where a firm’s asset value falls below its debt threshold at maturity. Asset value follows GBM, and equity is modeled as a call option on assets. The model yields a closed‑form default probability and credit spread, linking equity volatility to credit risk.

Conversely, the reduced‑form (intensity) approach models default as an exogenous Poisson process with stochastic intensity λ(t). The survival probability is exp(–∫₀ᵗ λ(s) ds). Calibration to credit default swap (CDS) spreads provides market‑consistent default intensities. Extensions include jump‑diffusion intensity processes and regime‑switching models.

Liquidity risk captures the cost and difficulty of executing large trades without significantly moving market prices. Market impact models, such as the Almgren‑Chriss framework, decompose execution cost into temporary and permanent impact components, providing optimal trade schedules that balance market risk against impact cost.

Term structure models describe the evolution of interest rates across maturities. The Vasicek model posits a mean‑reverting Ornstein‑Uhlenbeck process for the short rate: Dr = a(b – r)dt + σdW. The Cox‑Ingersoll‑Ross (CIR) model adds a square‑root diffusion to ensure positivity: Dr = a(b – r)dt + σ√r dW. Both yield analytical bond pricing formulas.

The Heath‑Jarrow‑Morton (HJM) framework specifies the entire forward rate curve dynamics directly, ensuring arbitrage‑free evolution by imposing a drift condition on the volatility structure. While flexible, HJM often requires infinite‑dimensional modeling; practical implementations adopt finite‑factor specifications.

Affine term structure models (e.G., The Duffie‑Kaneko model) assume that bond yields are affine functions of state variables, enabling closed‑form solutions for zero‑coupon bond prices. These models are popular for calibrating to the full yield curve and for pricing interest‑rate derivatives such as caps, floors, and swaptions.

In portfolio optimization, mean‑variance analysis solves for the weight vector that minimizes portfolio variance for a given expected return, or maximizes expected return for a given variance. The efficient frontier represents the set of optimal portfolios. The solution requires the inverse of the covariance matrix; numerical stability can be enhanced by shrinkage estimators or factor‑based approximations.

Utility theory extends mean‑variance by incorporating investor preferences via a utility function, commonly the constant‑relative‑risk‑aversion (CRRA) form U(W) = W^{1–γ}/(1–γ). Maximizing expected utility leads to optimal allocations that depend on the investor’s risk aversion γ and the distribution of returns. In practice, utility‑based approaches guide dynamic asset allocation and consumption decisions.

Dynamic programming solves multi‑period optimization problems by breaking them into stages and applying Bellman’s principle of optimality. The Hamilton‑Jacobi‑Bellman (HJB) equation characterizes the value function of a continuous‑time control problem, such as optimal consumption‑investment decisions under stochastic wealth dynamics.

In algorithmic trading, execution algorithms (e.G., VWAP, TWAP, Implementation Shortfall) schedule orders to achieve specific objectives: Matching the volume‑weighted average price, minimizing market impact, or reducing timing risk. Quantitative models estimate the trade‑off between price risk (variance) and impact cost, often using stochastic control techniques.

Machine learning methods, such as random forests, gradient boosting, and deep neural networks, are increasingly applied to financial prediction, risk assessment, and anomaly detection. Feature engineering—creating informative variables from price, volume, macro data, and sentiment—remains critical. Overfitting is a major concern; techniques like cross‑validation, early stopping, and regularization are essential safeguards.

Reinforcement learning frames trading as an agent interacting with a market environment, learning optimal policies through trial‑and‑error. The agent receives rewards based on profit and risk metrics, and updates its policy using algorithms like Q‑learning or policy gradients. Challenges include non‑stationary markets, exploration‑exploitation balance, and ensuring interpretability.

High‑dimensional optimization arises in large‑scale portfolio construction, where the number of assets can exceed thousands. Convex optimization solvers (e.G., Interior‑point methods) efficiently handle problems with linear constraints and quadratic objectives. For non‑convex problems, heuristic methods such as genetic algorithms or simulated annealing are employed, though they lack guarantees of global optimality.

Risk budgeting allocates risk contributions across portfolio components, ensuring that each asset or factor contributes a predefined portion of total risk (e.G., Volatility or Expected Shortfall). The risk contribution of asset i is w_i·(∂σ_p/∂w_i), where σ_p is portfolio volatility. Budgeting promotes diversification and aligns with institutional risk limits.

Scenario analysis often uses historical or hypothetical shocks to stress test portfolios. For instance, a “severe recession” scenario may involve a 40 % equity drawdown, a 200 bp increase in credit spreads, and a 30 % depreciation of the domestic currency. By re‑pricing all positions under these shocks, managers quantify potential losses and identify vulnerable exposures.

Liquidity‑adjusted VaR (L‑VaR) incorporates market depth and price impact into the VaR framework. Instead of assuming linear price changes, L‑VaR simulates trade execution paths that consume order book depth, yielding a more realistic loss estimate for large, illiquid portfolios.

Credit valuation adjustment (CVA) reflects the expected loss due to counterparty default on derivative contracts. CVA = (1 – R) ∫₀ᵀ E[EE(t)]·dPD(t), where R is recovery rate, EE(t) expected exposure, and PD(t) default probability. Computing CVA requires Monte Carlo simulation of exposure profiles and calibrated default intensity models.

Debit valuation adjustment (DVA) accounts for the company’s own credit risk, effectively reducing the fair value of liabilities. While theoretically sound under bilateral valuation, DVA raises accounting and risk‑management challenges because it can create profit from deteriorating credit quality.

Funding valuation adjustment (FVA) captures the cost of funding uncollateralized positions. It is computed as the discounted cost of borrowing the cash needed to finance the derivative’s exposure over its life. Inclusion of FVA in pricing aligns the valuation with the institution’s actual funding costs.

Initial margin requirements for non‑centrally cleared OTC derivatives are determined using models such as the Standard Initial Margin Model (SIMM). SIMM aggregates risk factors (e.G., Delta, vega, curvature) across asset classes, applying concentration and correlation rules to produce a margin amount that reflects potential future exposure.

Backtesting evaluates the performance of a model or strategy against historical data. For risk models, backtesting VaR involves counting the number of exceptions (realized losses exceeding VaR) and comparing to the expected frequency using the Kupiec test. For predictive models, backtesting examines out‑of‑sample accuracy, turnover, and profit‑and‑loss characteristics.

Model risk arises from incorrect model specifications, parameter mis‑estimation, or implementation errors. Effective model risk management includes independent model validation, benchmarking against alternative approaches, sensitivity analysis, and documentation of assumptions. Regulatory frameworks (e.G., SR 11‑7) mandate rigorous oversight of quantitative models.

Parameter uncertainty can be addressed through Bayesian posterior distributions, allowing risk measures to incorporate the full distribution of possible parameter values rather than point estimates. For instance, a Bayesian GARCH model yields a posterior over volatility paths, leading to a distribution of VaR estimates that reflects estimation risk.

Regime‑switching models capture structural changes in market behavior by allowing parameters (e.G., Drift, volatility) to switch between discrete states according to a Markov chain. The Hamilton model for business cycles is a classic example. In finance, regime‑switching GARCH models can differentiate between tranquil and turbulent periods, improving volatility forecasts.

Jump‑diffusion calibration typically involves matching both the implied volatility surface and higher‑order moments (skewness, kurtosis) observed in market option prices. The calibration problem is nonlinear and may require global optimization techniques such as differential evolution or particle swarm optimization to avoid local minima.

Factor models decompose asset returns into exposures to a set of common risk factors plus idiosyncratic noise: R_i = α_i + β_i′f + ε_i. The Fama‑French three‑factor model adds size (SMB) and value (HML) factors to the market factor. Multi‑factor models are used for risk attribution, performance evaluation, and constructing factor‑tilted portfolios.

Eigen‑value decomposition of the covariance matrix yields principal components that serve as orthogonal risk factors. By retaining the top K components, one reduces dimensionality while preserving the bulk of variance. This approach underpins risk‑parity strategies that allocate capital equally across principal component risk contributions.

Liquidity‑adjusted pricing incorporates the expected cost of trading into asset valuation. For example, the liquidity‑adjusted discount rate r_L = r_f + λ·L, where λ reflects the market’s illiquidity premium and L is a liquidity measure (e.G., Bid‑ask spread). This adjustment aligns pricing with investors’ required compensation for bearing liquidity risk.

Option Greeks quantify sensitivities of option prices to underlying parameters: Delta (∂V/∂S), gamma (∂²V/∂S²), vega (∂V/∂σ), theta (∂V/∂t), rho (∂V/∂r). Accurate Greek computation is essential for hedging. Analytical formulas exist for vanilla options under Black‑Scholes; for exotic options, Greeks are estimated via finite differences or adjoint algorithmic differentiation.

Adjoint algorithmic differentiation (AAD) computes gradients efficiently by propagating sensitivities backward through the computational graph, reducing the cost of obtaining all Greeks from O(N) to O(1) relative to the number of input parameters. AAD is widely used in large‑scale Monte Carlo simulations where thousands of sensitivities are required.

Dynamic hedging involves continuously adjusting a portfolio of underlying assets and derivatives to maintain a target exposure (e.G., Delta‑neutral). In practice, hedging is performed discretely, leading to hedging error. Stochastic control techniques can optimize rebalancing frequency, balancing transaction costs against hedging error.

Transaction cost analysis (TCA) evaluates the explicit (commissions, fees) and implicit (market impact, opportunity cost) costs of trading. Models such as Almgren‑Chriss provide a framework to predict impact as a function of trade size, volatility, and market depth. TCA informs optimal execution strategies and performance attribution.

Risk‑adjusted performance metrics include Sharpe ratio (excess return divided by volatility), Sortino ratio (excess return divided by downside deviation), and Information ratio (active return divided by tracking error). These metrics standardize performance across strategies with differing risk profiles, facilitating manager evaluation.

Portfolio insurance replicates a protective floor on portfolio value using dynamic trading or options. The Constant Proportion Portfolio Insurance (CPPI) method allocates a multiple of the cushion (portfolio value minus floor) to risky assets, rebalancing as the cushion changes. CPPI can be implemented with futures, reducing transaction costs.

Credit risk models for portfolio credit exposures often employ the Gaussian Copula framework, where default times are linked through a common latent factor. The model derives portfolio loss distribution, enabling calculation of portfolio VaR and Expected Loss. Limitations include unrealistic tail dependence, prompting the use of t‑copulas or more sophisticated credit contagion models.

Systemic risk measures such as CoVaR (conditional VaR) assess the spillover effect of one institution’s distress on the system. CoVaR is defined as the VaR of the financial system conditional on an institution being in distress. Empirical estimation uses quantile regression, highlighting institutions that contribute disproportionately to systemic risk.

Network models represent financial interconnections (e.G., Interbank loans, derivative exposures) as graphs. Metrics like degree centrality, eigenvector centrality, and clustering coefficient identify systemically important nodes. Simulation of shock propagation on the network helps regulators evaluate contagion risk and design macroprudential policies.

Monte Carlo Greeks compute sensitivities by differentiating the Monte Carlo estimator. Pathwise differentiation works when the payoff is smooth; likelihood ratio (Malliavin) methods apply to discontinuous payoffs. These techniques allow simultaneous estimation of option price and Greeks within a single simulation run.

Policy iteration and value iteration are algorithms for solving discrete‑time Markov decision processes (MDPs) in optimal trading problems. Policy iteration alternates between evaluating a policy and improving it, converging to the optimal policy. Value iteration updates the value function directly using the Bellman optimality equation.

Stochastic differential equation solvers such as the Euler‑Maruyama and Milstein schemes discretize continuous‑time dynamics for simulation. The Milstein method adds a correction term to improve strong convergence order, beneficial when simulating diffusion processes with high accuracy requirements.

Variance‑gamma (VG) process is a pure‑jump Lévy process obtained by time‑changing Brownian motion with a Gamma subordinator. VG captures leptokurtic return distributions and asymmetric skewness, improving option pricing for assets with pronounced smile effects. Calibration involves matching moments or implied volatilities across strikes.

Fourier transform methods (e.G., Carr‑Madan, COS method) compute option prices by evaluating the characteristic function of the underlying process. These methods are fast and accurate for models with known characteristic functions (e.G., Heston, VG). They bypass path simulation, enabling efficient calibration to large option datasets.

Adjoint Monte Carlo combines Monte Carlo simulation with adjoint differentiation to obtain gradients of expected payoffs with respect to model parameters. This is crucial for calibrating complex models where each iteration would otherwise require a full re‑simulation.

Risk factor modeling identifies the fundamental drivers of portfolio risk, such as equity indices, interest rate curves, commodity spot prices, and credit spreads. Factor sensitivities (betas) are estimated via regression or historical mapping, allowing aggregation of risk across heterogeneous positions.

Scenario generation for multi‑asset risk simulations often employs copula‑based approaches to preserve dependence structures while sampling marginal distributions. Gaussian copulas are simple but may underestimate tail dependence; t‑copulas or vine copulas provide richer dependence modeling.

Liquidity‑driven price dynamics incorporate order‑book depth and arrival rates of market orders into stochastic models. The Avellaneda‑Stoikov framework models optimal quoting for market makers, balancing inventory risk against profit from the spread, and is used to simulate realistic high‑frequency price paths.

Risk‑adjusted return on capital (RAROC) measures the expected return of a business unit relative to its economic capital, defined as the capital needed to cover a chosen risk metric (e.G., VaR). RAROC = (Expected Profit – Cost of Capital) / Economic Capital, guiding allocation decisions and performance assessment.