The mathematics underneath every price.

A derivative's price is not a guess. It is the solution to a partial differential equation — and knowing which equation, and why, is what separates pricing from curve-fitting.

Quantitative finance answers a narrower question than time series or causal inference: given an assumed model for how an underlying asset moves, what is the arbitrage-free price of a claim written on it, and how do you hedge that claim as the underlying moves? The answer is a specific, beautiful piece of applied mathematics — stochastic calculus — and almost everything in modern derivatives pricing is a variation on one theorem.

Brownian motion and Itô calculus

A standard Wiener process $W_t$ has independent, normally-distributed increments: $W_t-W_s\sim \mathcal{N}(0,t-s)$. Asset prices are typically modeled as geometric Brownian motion (GBM):

$d{S}_t=\mu S_{t}dt+\sigma S_{t}d{W}_t$

Because $W_t$ is nowhere differentiable, ordinary calculus doesn't apply to functions of $S_t$. Itô's lemma is the stochastic chain rule that fixes this, and it is the single most-used result in the field:

$df(S_t,t)=\left(\frac{\partial f}{\partial t}+\mu S_t\frac{\partial f}{\partial S}+\frac{1}{2}{\sigma }^2{S}_t^2\frac{{\partial }^{2}f}{\partial S^2}\right)dt+\sigma S_t\frac{\partial f}{\partial S}d{W}_t$

That extra second-derivative term — absent from ordinary calculus — comes directly from $(d{W}_t{)}^2=dt$, and it is the reason option prices are convex in the underlying (“gamma”) in a way that has no classical-calculus analogue.

Black-Scholes-Merton

Applying Itô's lemma to a hedged portfolio (long the option, short $\Delta$ shares of the underlying) and requiring the portfolio to earn the risk-free rate — no arbitrage — eliminates the random term entirely and yields a deterministic PDE for the option price $V(S,t)$:

$\frac{\partial V}{\partial t}+\frac{1}{2}{\sigma }^2{S}^2\frac{{\partial }^{2}V}{\partial S^2}+rS\frac{\partial V}{\partial S}-rV=0$

Solved with the appropriate terminal condition (the payoff at expiry), this is the Black-Scholes-Merton equation, and its closed-form solution for a European call is the Black-Scholes formula. The deep insight is not the formula itself but the method: a claim can be priced by constructing a dynamically hedged, riskless replicating portfolio — price is what makes arbitrage impossible, not a forecast of the underlying.

The Feynman-Kac connection

The Feynman-Kac theorem is what makes the PDE and expectation views the same object: it states that the solution to a certain class of parabolic PDEs can be represented as the expected value of a functional of a stochastic process. Black-Scholes is the special case; the same theorem underlies Monte Carlo pricing for path-dependent claims that have no closed-form PDE solution, because the expectation representation lets you simulate paths instead of solving the PDE directly.

Beyond Black-Scholes

Black-Scholes assumes constant volatility — contradicted immediately by the volatility smile observed in real options markets. The extensions each relax one assumption:

• Heston model. Volatility is itself a mean-reverting stochastic process, $d{\nu }_t=\kappa (\theta -{\nu }_t)dt+\xi \sqrt{{\nu }_t}d{W}_t^{\nu }$, correlated with the asset's own Brownian motion — capturing the leverage effect and producing a realistic volatility smile.
• Merton jump-diffusion. Adds a compound Poisson jump process to GBM, capturing the discontinuous moves (earnings surprises, macro shocks) that continuous diffusions cannot represent.
• Local volatility (Dupire). Volatility is a deterministic function of price and time, $\sigma (S_t,t)$, calibrated to exactly match the observed volatility surface — the market-consistent alternative when the goal is pricing exotics off vanilla quotes rather than modeling the true dynamics.

The Fokker-Planck equation

Where Feynman-Kac moves from a PDE to an expectation, Fokker-Planck moves in the other direction: given the stochastic differential equation for $S_t$, it describes how the probability density $p(S,t)$ of the process evolves over time:

$\frac{\partial p}{\partial t}=-\frac{\partial }{\partial S}[\mu (S,t)p]+\frac{1}{2}\frac{{\partial }^2}{\partial S^2}[{\sigma }^2(S,t)p]$

This is the same forward-evolution idea that underlies the state-space models in the time series essay — Fokker-Planck is the continuous-time, continuous-state analogue of a Kalman filter's prediction step.

Portfolio theory and risk management

Markowitz and the efficient frontier

Modern portfolio theory frames allocation as an optimization: minimize portfolio variance $w^T\Sigma w$ subject to a target expected return $w^T\mu =r^{\ast }$. The resulting efficient frontier — and CAPM's further claim that the market portfolio is the unique tangency portfolio — is foundational but fragile: it is highly sensitive to estimation error in $\mu$, which is why financial ML approaches to expected-return estimation matter as much as the optimization itself.

Value at Risk and Expected Shortfall

VaR${}_{\alpha }$ answers “what loss will not be exceeded with probability $\alpha$?” but says nothing about the severity of losses beyond that threshold, and is not sub-additive — diversification can, on paper, increase VaR. Expected Shortfall (CVaR), $\mathbb{E}[L\mid L>{\text{VaR}}_{\alpha }]$, is coherent and is now the Basel-standard risk measure for exactly this reason. Both require backtesting (e.g. Kupiec's proportion-of-failures test) against realized outcomes — a risk model that has never been checked against reality is an assumption wearing a number.

Common pitfalls

• Model risk. Every model in this essay is a simplification; treating its output as ground truth rather than a conditional statement (“if GBM with this $\sigma$ holds, then...”) is the single largest source of blowups in derivatives desks.
• Calibration overfitting. Models with enough free parameters (Heston, local vol) can fit today's volatility surface almost exactly while being badly wrong about tomorrow's dynamics — fit quality on today's surface is not evidence the model is correct.
• Ignoring fat tails. Gaussian-based models systematically underprice tail risk; Mandelbrot flagged this in 1963 and it remains the most under-priced risk in naive implementations.
• Static hedging assumptions. Continuous rebalancing is a modeling convenience, not a market reality — discrete rebalancing under transaction costs introduces hedging error that the frictionless theory doesn't account for.

The path forward

A disciplined pricing and risk workflow, in order:

1. Choose the model to match the question — GBM for intuition, Heston or local vol when the smile matters, jump-diffusion when discontinuities matter.
2. Calibrate to liquid instruments; validate out-of-sample on instruments not used in calibration before trusting exotic prices.
3. Report risk in coherent measures (Expected Shortfall, not just VaR) and backtest the risk model against realized P&L.
4. Treat every price as conditional on the model's assumptions — and know which assumption is most likely to break first.