This page is the reference. Nothing on it teaches; it records the conventions every other page assumes, so that a symbol or a unit that trips you has one place to be looked up. Read the prerequisites once, then come back here when you need to.

Units

Time is measured in years. T=1T = 1 is one year, and a step dt\dt is a fraction of a year. Where a page counts days it says which count it is using. The pricer and the library use Actual/365: an expiry in days divided by 365, with nothing rounded in between. Other day counts exist — Actual/360 and 30/360 among them — and the fraction of a year between two dates depends on which one applies; 1.0 prices a deposit with the wrong one and shows the size of the error.

Volatility is annualised and carried as a decimal: σ=0.30\sigma = 0.30 is 30% a year. A slider shows the percent; the code holds the decimal, and the formulas take the decimal. Over TT years the spread is σT\sigma\sqrt{T}, and 2.1 shows why the square root.

Rates are continuously compounded throughout Parts 2 and 3, and carried as decimals. A rate rr means one unit of cash grows to erTe^{rT} over TT years, and a payment of one unit at TT is worth e−rTe^{-rT} now. A rate may be negative. A dividend yield qq is a continuous yield in the same units. Simple and annual compounding, and the conversions between them, are 1.0's.

Prices are in the underlying's currency, and there is one underlying, so no currency ever needs converting. Greeks are raw partial derivatives per unit of their input — vega per 1.00 of volatility, theta per year, rho per 1.00 of rate — and a desk's per-point, per-day and per-basis-point figures are scalings applied at display, once, where you can see them. 2.10 and 3.5 own the Greeks; this page only fixes their units.

Prerequisites

Probability at the level a working developer has: variances add across independent things, and sums of many small independent things come out normal. Both are assumed without proof. The stochastic side — everything from a random walk onwards — is taught in full.

Ordinary calculus to a second-order Taylor expansion. That is the whole of what Itô's lemma adds the rulebook to.

You can read code. The snippets are Python with numpy; the library is C++20 and Part 3 reads it. You do not need to write either to follow the pages.

No finance is assumed. 0.1 says what the instruments are for and Part 1 says what each one is.

Running the code

There are four kinds of code on this site: experiments, snippets, demos and the library. You can directly run all apart from snippets, which are part of the experiments themselves.

The experiments behind Part 2 live in the Touchstone repository under experiments/, one script per page, each writing its JSON result beside it. Every script takes an explicit seed and records it in its output, so the same seed gives the same result, to the tolerance each script asserts. They run under uv, which installs Python, numpy and QuantLib at the pinned versions and needs nothing else.

The snippets shown on a page are extracts of those files, not copies of them, and the site's build runs every file it extracts from. A snippet's caption names the file and the lines. If you want to run a snippet, run the experiment it came from.

The demos run in your browser and need nothing installed.

The library is C++20. It builds with CMake and its tests run under ctest; GCC and Clang both build it with warnings as errors, and the tests read the committed QuantLib reference values rather than linking QuantLib.

The exact commands are in the Touchstone README, which is where they are kept current.

Notation

Bars, brackets and Greek letters carry precise, boring meanings and never decorate. When two look-alike symbols coexist on a page, the page says so at first use.

Not what ++ and −- mean: only the notation that trips working developers.

Derivatives

A function of one variable gets primes: g′(x)g'(x), g′′(x)g''(x). Once time enters, as in f(t,x)f(t, x), we switch to partial derivatives: ∂f/∂t\partial f/\partial t, ∂f/∂x\partial f/\partial x, ∂2f/∂x2\partial^2 f/\partial x^2. The switch itself tells you ff depends on tt as well as xx.

We always differentiate with respect to the placeholder xx, then evaluate along the path: ∂f/∂x\partial f/\partial x at (t,Wt)(t, W_t), never ∂f/∂W\partial f/\partial W.

In this notation, Itô's lemma reads:

df=(∂f∂t+a ∂f∂x+12b2∂2f∂x2)dt+b ∂f∂x dW\begin{aligned} \dd f = {} & \left( \frac{\partial f}{\partial t} + a\,\frac{\partial f}{\partial x} + \tfrac{1}{2} b^2 \frac{\partial^2 f}{\partial x^2} \right) \dt \\ & {} + b\,\frac{\partial f}{\partial x} \,\dW \end{aligned}

where aa is the drift and bb the volatility. Drift-plus-noise lines always put dt\dt before dW\dW.

Random walks and scaling — page 2.1

SymbolSay itHere it meansWatch out
SnS_n"S sub n"position of the walk after nn stepsthe subscript is a time index, never a power
ξi\xi_i"ksee" (xi)the ii-th step, +1+1 or −1-1Greek habit: letters like ξ\xi, ZZ name random draws; not a constant
E[X]\E[X]expectation of XXthe probability-weighted average of XX over all possible outcomes ("worlds")not "the value you expect" — E[ξ]=0\E[\xi] = 0 is a value the step never takes
Var⁡(X), sd⁡(X)\Var(X),\ \sd(X)variance, standard deviationspread of XX; sd⁡=Var⁡\sd = \sqrt{\Var}Var⁡\Var is in squared units and adds across independent terms; sd⁡\sd is in XX's own units and does not
∣x∣\lvert x \rvertabsolute value / magnitudethe size of xx, sign stripped; when xx is a position or a difference, its distance from 0bars mean size, always; "distance" is the reading when the thing inside is a position
E∣X∣\E\lvert X \rvertexpected magnitudeshorthand for E[∣X∣]\E[\lvert X \rvert] — the brackets are elided when bars are present2.1's trap: E[Sn]=0\E[S_n] = 0 exactly while E∣Sn∣≈0.80n\E\lvert S_n\rvert \approx 0.80\sqrt{n}. One glyph, two animals
E[X∣Y]\E[X \mid Y]"expectation of XX given YY"conditioning — arrives properly in 2.2this vertical bar is unrelated to absolute-value bars despite the identical glyph
∝\proptoproportional toequal up to a fixed constantsays how something scales, not what it equals: Var⁡(Sn)∝n\Var(S_n) \propto n
≈\approx vs ∼\sim vs ==approx / asymptotic / exact≈\approx numerically close; ∼\sim "behaves like, for large nn"; == exactE∣Sn∣∼2/π⋅n\E\lvert S_n\rvert \sim \sqrt{2/\pi} \cdot \sqrt{n} is a large-nn statement, not an identity at n=3n = 3
O(dt), O(dt)O(\dt),\ O(\sqrt{\dt})"order dt"shrinks at the same rate as dt\dt (resp. dt\sqrt{\dt}) as dt→0\dt \to 0a rate, not a value — the whole dt\sqrt{\dt} pincer is a fight between orders
W, WtW,\ W_t"W" (for Wiener)the random walk of 2.1 taken to its continuous limit with dt\sqrt{\dt}-sized steps — Brownian motion; WtW_t is its position at time tt. Defined properly in 2.2not new machinery: 2.1's object wearing a name. Var⁡(Wt)=t\Var(W_t) = t is the step dt\sqrt{\dt} column of the pincer
dW\dW"d-W"the increment of WW over the next dt\dt slot: Wt+dt−WtW_{t+\dt} - W_t — a dt\sqrt{\dt}-sized zero-mean kickthe d means "a small increment of", as in dt\dt — NOT a derivative; WW has no derivative (2.2)

Brownian motion — page 2.2

SymbolSay itHere it meansWatch out
Wt, W(t)W_t,\ W(t)"W at t"the position of Brownian motion at time tt; the argument is a time, and Var⁡(Wt)=t\Var(W_t) = tthe number you see is a variance, never an sd⁡\sd; sd⁡\sd is t\sqrt{t} and is never written. W(0.5)−W(0.25)W(0.5) - W(0.25) has sd⁡\sd 0.5 — the time 0.5 in a different hat, coincidence not identity
N(μ,σ2), N(0,t−s)N(\mu, \sigma^2),\ N(0, t - s)"normal with mean mu and variance sigma squared"the second slot of N(⋅,⋅)N(\cdot,\cdot) is a variance, by conventionN(0,0.25)N(0, 0.25) has sd⁡\sd 0.5. numpy's rng.normal(loc, scale) takes scale as the sd, so code says scale=0.5 where the maths says N(0,0.25)N(0, 0.25) — same distribution, opposite slot
X∼N(0,t)X \sim N(0, t)"X is distributed as"∼\sim between a random variable and a distribution means "has this distribution" — exact, not approximatesecond reading of ∼\sim on the course (2.1: "behaves like, for large nn"); which one is meant is fixed by what sits on the right — a distribution, or a formula in nn
ΔW\Delta W"delta W"the increment over a finite step: Wt+dt−WtW_{t+\dt} - W_t, over one step dt\dtthe thing simulation actually sums: ΔW∼N(0,dt)\Delta W \sim N(0, \dt), so the path is ∑dt⋅Z\sum \sqrt{\dt} \cdot Z. dW\dW is the same thing with dt→0\dt \to 0; neither is a derivative
Fs\mathcal{F}_s"F sub s" / "the filtration at s"everything known up to time ss — the whole history of the path so farnot a function and not a set of numbers you compute; E[Xt∣Fs]\E[X_t \mid \mathcal{F}_s] reads "the expectation of XtX_t given all we know at ss". For Brownian motion, WsW_s alone carries all of it (Markov)
E[Xt∣Fs]=Xs\E[X_t \mid \mathcal{F}_s] = X_s"martingale property"the best forecast of the future value, given everything known now, is the current valuethis is conditional: E[W2]=0\E[W_2] = 0 unconditionally, but E[W2∣W1=1.5]=1.5\E[W_2 \mid W_1 = 1.5] = 1.5. Not "the mean is constant" — "the mean, given now, is now"
martingale · Markov—martingale: forecast = present (a statement about the mean). Markov: given the present, the past adds nothing (a statement about information)different properties, both implied by independent increments. Markov ⇏\nRightarrow independent increments: a mean-reverting process is Markov and its increments depend on the level
c⋅W(t/c)\sqrt{c} \cdot W(t/c)"root-c W of t over c"the zoomed path: time stretched by cc, space by c\sqrt{c}; again a Brownian motionthe vertical factor is c\sqrt{c}, not cc (variance goes ×c\times c, sd⁡\sd goes ×c\times \sqrt{c}). For a 1 % zoom c=100c = 100 and the axis stretches ×10\times 10
excess kurtosis—E[(X−μ)4]/σ4−3\E[(X-\mu)^4]/\sigma^4 - 3: zero for a Gaussian, negative for lighter tails, positive for heavierthe −3-3 is what makes it excess; a kk-step sum of ±1\pm 1 coins has exactly −2/k-2/k, which is how 2.2 shows the CLT arriving rather than asserting it
What excess kurtosis measures. All four at mean 0 and variance 1, so only shape differs.
−4−3−2−101234
  • Laplace · +3
  • Gaussian · 0
  • uniform · −1.2
  • 10-step coin sum · −0.2
Positive is a sharper peak and fatter tails than the Gaussian — more mass far out, paid for from the shoulders (Laplace, +3; a Student-t would be higher still). Negative is a flatter top and tails that stop (uniform, −1.2; a single ±1 step, −2, is the floor). The dots are the sum of ten ±1 steps: a lattice, already close to the curve at −0.2, and at −2/k it walks onto the Gaussian as k grows — which is the measurement on 2.2. In markets, daily returns are leptokurtic; that is 2.11’s subject.

Quadratic variation — page 2.3

SymbolSay itHere it meansWatch out
∑\sum"sigma" — sumadd up over every grid step, i=1…ni = 1 \ldots n: ∑(ΔW)2=(ΔW1)2+…+(ΔWn)2\sum (\Delta W)^2 = (\Delta W_1)^2 + \ldots + (\Delta W_n)^2a sum over the grid is one number, not a function of tt. nn terms of size dt\dt make order 1; of size dt\sqrt{\dt} make order n\sqrt{n}; of size dt2\dt^2 make order 1/n1/n
n, dt=T/nn,\ \dt = T/nsteps, step sizethe grid: [0,T][0, T] chopped into nn steps of length dt\dtcarry TT: ∑∣ΔW∣≈0.798⋅nT\sum \lvert \Delta W \rvert \approx 0.798 \cdot \sqrt{nT}, which is 0.798⋅n0.798 \cdot \sqrt{n} only because T=1T = 1 in the experiment
Z, ZiZ,\ Z_i"Z"the standard normal, Z∼N(0,1)Z \sim N(0, 1); ZiZ_i is the ii-th independent draw. ΔW=dt⋅Z\Delta W = \sqrt{\dt} \cdot Za specific named variable, not a family: "Z(x,y)Z(x, y)" means nothing. The bracket in N(μ,σ2)N(\mu, \sigma^2) carries the Gaussian family's parameters; other families put other things there (χk2\chi^2_k: a count; U(a,b)U(a, b): endpoints)
Z2, χ12Z^2,\ \chi^2_1"Z squared", "chi-squared with one degree of freedom"the square of a standard normal. Mean 1, variance 2, lives on [0,∞)[0, \infty), median 0.45, piled against 0 with a long right tailnot N(1,2)N(1, 2): an N(1,2)N(1, 2) is negative 24 % of the time, a square never. Variance 2 comes from E[Z4]=3\E[Z^4] = 3: for X∼N(0,σ2)X \sim N(0, \sigma^2), Var⁡(X2)=2σ4\Var(X^2) = 2\sigma^4. sd⁡\sd 2\sqrt{2} is a root-mean-square, not a ±\pm range — the spread is all on the right
χk2\chi^2_k"chi-squared with k degrees of freedom"Z12+…+Zk2Z_1^2 + \ldots + Z_k^2, kk independent standard normals squared. Mean kk, variance 2k2kthe subscript is a count, not a parameter like μ\mu or σ2\sigma^2. ∑(ΔW)2/dt\sum (\Delta W)^2/\dt is χn2\chi^2_n with n=106n = 10^6 — a bell by the CLT, though each term is a spike
[a,b), [0,∞)[a, b),\ [0, \infty)interval notationthe numbers from aa up to but not including bb. Square bracket includes the end, round bracket excludes it. [0,∞)[0, \infty) = zero or any positive number∞\infty always gets a round bracket — it is a direction, not a value reached
[W]T, ⟨W⟩T[W]_T,\ \langle W \rangle_T"the quadratic variation of W up to T"the limit of ∑(ΔW)2\sum (\Delta W)^2 over [0,T][0, T] as the grid is refined; equals TT for Brownian motion, path by pathsquare brackets here are not an interval; angle brackets are not an expectation. Both notations are in use for the same thing
quadratic variation · total variation—quadratic: ∑(Δ⋅)2\sum (\Delta \cdot)^2 in the limit. Total: ∑∣Δ⋅∣\sum \lvert \Delta \cdot \rvert in the limit — a curve's length"variation" is not variance. A smooth curve: total variation finite, quadratic 0. WW: total ∞\infty, quadratic TT. Each object has one and not the other
(dW)2=dt(\dW)^2 = \dt, dW⋅dt=0\dW \cdot \dt = 0, (dt)2=0(\dt)^2 = 0, (dW)3=0(\dW)^3 = 0"the rulebook"statements about sums over the grid as dt→0\dt \to 0: ∑(ΔW)2→T\sum (\Delta W)^2 \to T (what ∑dt\sum \dt sums to), ∑ΔW⋅Δt→0\sum \Delta W \cdot \Delta t \to 0, ∑(Δt)2→0\sum (\Delta t)^2 \to 0, ∑∣ΔW∣3→0\sum \lvert \Delta W \rvert^3 \to 0not about one step: a single (ΔW)2(\Delta W)^2 is a wild χ12\chi^2_1 draw, nothing like dt\dt. "=0= 0" means "vanishes when summed"; "=dt= \dt" means "sums to what dt\dt sums to". Rule: size dt\dt survives, anything smaller dies
f′, f′′, g′′′f',\ f'',\ g'''"f prime", "f double prime", "g triple prime"first, second, third derivatives — each a functionon screen ′′'' looks like a closing quote mark. W′W' does not exist (2.2); that is why 2.3 works with ΔW\Delta W, never W′W'
chain rule—for f(t)=g(h(t))f(t) = g(h(t)): f′(t)=g′(h(t))⋅h′(t)f'(t) = g'(h(t)) \cdot h'(t) — differentiate the outside with the inside left alone, then multiply by the inside's derivativethe inner derivative is the factor people drop: sin⁡(2πt)′=2π⋅cos⁡(2πt)\sin(2\pi t)' = 2\pi \cdot \cos(2\pi t), and that 2π2\pi is squared in every integral of f′f'. For g(W)g(W) the chain rule is short by 12∫g′′(W) dt\tfrac12\int g''(W) \,\dt — 2.4
∫0T…dt\int_0^T \ldots \dt"the integral from 0 to T"the continuous version of ∑…Δt\sum \ldots \Delta t: area under the curve. Over [0,1][0, 1] it is just the average value of the integrandthree different integrals of f′f' on 2.3: ∫f′\int f' = net change (0 for a full sine wave), ∫∣f′∣\int \lvert f' \rvert = distance travelled (4), ∫f′2\int f'^2 = mean squared speed (2π22\pi^2). Only the first is "undoing the derivative"
sin⁡(2πt), cos⁡2\sin(2\pi t),\ \cos^2—one full wave over t∈[0,1]t \in [0, 1]; f′=2π⋅cos⁡(2πt)f' = 2\pi \cdot \cos(2\pi t); cos⁡2\cos^2 averages 12\tfrac12 over a period (cos⁡2θ=12+12cos⁡2θ\cos^2\theta = \tfrac12 + \tfrac12\cos 2\theta)the 2π2\pi makes the period exactly 1 and reappears as 4π24\pi^2 under the integral
φ(z)\varphi(z)"phi of z"the standard normal density, e−z2/2/2πe^{-z^2/2}/\sqrt{2\pi}: the height of the bell at zza height, not a probability; probabilities are areas. E[g(Z)]=∫g(z)⋅φ(z) dz\E[g(Z)] = \int g(z) \cdot \varphi(z) \,\dd z is the expectation recipe — value × how likely, summed
E∣Z∣=2/π\E\lvert Z \rvert = \sqrt{2/\pi}"expected size of Z"0.7979, the average magnitude of a standard normal draw; hence E∣ΔW∣=0.798⋅dt\E\lvert \Delta W \rvert = 0.798 \cdot \sqrt{\dt} and ∑∣ΔW∣≈0.798⋅nT\sum \lvert \Delta W \rvert \approx 0.798 \cdot \sqrt{nT}less than the sd⁡\sd of 1 because sd⁡\sd is a root-mean-square and squaring favours the big draws. The ±dt\pm\sqrt{\dt} coin has constant exactly 1 instead
odd moment—E[Xk]\E[X^k] with kk odd (1, 3, 5, …); zero for any XX symmetric about 0why corr⁡(∑(ΔW)2, WT)\operatorname{corr}(\sum (\Delta W)^2,\ W_T) is exactly 0: the only surviving covariance term is E[(ΔW)3]\E[(\Delta W)^3], an odd moment. max⁡∣W∣\max\lvert W \rvert is not a sum of ΔW\Delta W and the trick fails — its correlation is small, not zero
log-log slope—fit a line to log⁡(y)\log(y) against log⁡(n)\log(n); its slope is the exponent pp in "yy goes like npn^p"−12-\tfrac12 means 1/n1/\sqrt{n}, −1-1 means 1/n1/n, +12+\tfrac12 means n\sqrt{n}, 0 means settles. Every "goes like" claim in the JSON is checked this way
Taylor expansion—g(x+Δ)−g(x)=g′(x)⋅Δ+12g′′(x)⋅Δ2+16g′′′(x)⋅Δ3+…g(x + \Delta) - g(x) = g'(x) \cdot \Delta + \tfrac12 g''(x) \cdot \Delta^2 + \tfrac16 g'''(x) \cdot \Delta^3 + \ldots: the change in a smooth gg over a small step, as a series in the stepfor a smooth path only the Δ\Delta term survives summation (chain rule). For WW the Δ2\Delta^2 term survives too, because ∑(ΔW)2→T\sum (\Delta W)^2 \to T — the whole reason ordinary calculus fails

The Itô integral — page 2.4

Rows marked ↺ revisit an earlier row in its new role.

SymbolSay itHere it meansWatch out
∫0TW dW\int_0^T W \,\dW"the integral of WW with respect to WW" — a stochastic integralthe limit as n→∞n \to \infty of the left-endpoint sum ∑W(ti)⋅ΔWi\sum W(t_i) \cdot \Delta W_i. Equals 12WT2−12T\tfrac12 W_T^2 - \tfrac12 Tnot "the area under WW" — that is ∫W dt\int W \,\dt (2.3's kind of integral, with dt\dt). Here the increment is dW\dW, so this is a sum of position × change, a P&L
∫H dW\int H \,\dW, ∫h dS\int h \,\dS"integral of HH against WW"the general form: HH is the integrand (the position), WW or SS what it is held against (the price). ∫W dW\int W \,\dW is the case H=WH = WHH must be adapted — read at the left end of each step. Which letter is "position" and which is "price" is the whole content of the notation
tit_i, mim_i, ti+1t_{i+1}left endpoint, midpoint, right endpointthe three places in step ii where the integrand can be read: ti=i⋅dtt_i = i \cdot \dt, mi=(ti+ti+1)/2m_i = (t_i + t_{i+1})/2, ti+1=ti+dtt_{i+1} = t_i + \dtthe choice changes the answer by 12T\tfrac12 T or TT. On the smooth pages it changed nothing, which is why nobody mentioned it before
δ1\delta_1, δ2\delta_2"delta one, delta two"the two half-step increments of a step: δ1=W(mi)−W(ti)\delta_1 = W(m_i) - W(t_i), δ2=W(ti+1)−W(mi)\delta_2 = W(t_{i+1}) - W(m_i), so ΔW=δ1+δ2\Delta W = \delta_1 + \delta_2; each ∼N(0,dt/2)\sim N(0, \dt/2)lower-case δ\delta for a half-step, capital Δ\Delta for a full step — same "change of" meaning, different grid
∫W∘dW\int W \circ \dW"Stratonovich integral" — the circlethe midpoint limit, 12WT2\tfrac12 W_T^2. Related to Itô by ∫W∘dW=∫W dW+12T\int W \circ \dW = \int W \,\dW + \tfrac12 Trecognise it, do not use it: on this course every ∫…dW\int \ldots \dW is Itô. The circle is not composition of functions
adapted · non-anticipating—a process HH is adapted if H(t)H(t) depends only on what is known at tt — on Ft\mathcal{F}_t, the filtration from 2.2. "Non-anticipating" is the same word from the other sidethe trading meaning: you choose the position before the price moves. WW itself is adapted; W(t+dt)W(t + \dt) is not adapted at tt
I(t)I(t), R(t)R(t)"the running integral"the left (II) or right (RR) sum over the steps before tt — the integral as a process in tt, not one numberI(t)I(t) is a martingale; R(t)R(t) has E[R(t)]=t\E[R(t)] = t. II is for integral (not an identity matrix), RR for right endpoint (not a return)
martingale ↺—2.2's property, now applied to I(t)I(t): E[I(T)∣Fs]=I(s)\E[I(T) \mid \mathcal{F}_s] = I(s). The first thing the Itô integral is, once you have itit is a property of the left-endpoint limit only. Checked on 2.4 as corr⁡(first half, second half)=0\operatorname{corr}(\text{first half},\ \text{second half}) = 0 and E[second half∣sign of first half]=0\E[\text{second half} \mid \text{sign of first half}] = 0
Itô isometry—E[(∫0TH dW)2]=∫0TE[H2] dt\E[(\int_0^T H \,\dW)^2] = \int_0^T \E[H^2] \,\dt: the mean-square of a stochastic integral is an ordinary integral of the mean-square integrand. For H=WH = W: 12T2\tfrac12 T^2"isometry" is geometry's word for "preserves lengths": the size of the integral (root-mean-square) equals the size of the integrand. Not a statement about means
E[X2]\E[X^2] vs Var⁡(X)\Var(X)mean-square vs varianceVar⁡(X)=E[X2]−(E[X])2\Var(X) = \E[X^2] - (\E[X])^2. When E[X]=0\E[X] = 0 they coincide, which is why the isometry gives the variance of ∫W dW\int W \,\dW directlyfor the right sum they differ: E[R2]≠Var⁡(R)\E[R^2] \neq \Var(R) because E[R]=T\E[R] = T
s.e."standard error"sd⁡\sd of a sample mean: sd⁡\sd of the values divided by number of values\sqrt{\text{number of values}}. 20,000 values with sd⁡\sd 0.7 give s.e. 0.005not the sd⁡\sd of the thing itself. "−0.004 ± 0.005" means the mean is 0 to within noise; the values themselves spread ±0.7
limit"the limit as n→∞n \to \infty"the value a grid quantity settles on as the grid is refined without end: the last row of a table like 2.4's path-1 table. lim⁡∑W(ti)⋅ΔWi=12WT2−12T\lim \sum W(t_i) \cdot \Delta W_i = \tfrac12 W_T^2 - \tfrac12 Tnot a bound, a maximum or an extreme path — the everyday sense of the word is the wrong one here. "left − limit" is "this grid's value minus the settled value", the discretisation error, and it can have either sign
== (exact) vs →\to (limit)"equals on every grid" vs "tends to"two kinds of equality on 2.4: left=12WT2−12∑(ΔW)2\text{left} = \tfrac12 W_T^2 - \tfrac12\sum (\Delta W)^2 is algebra, true at n=3n = 3; left→12WT2−12T\text{left} \to \tfrac12 W_T^2 - \tfrac12 T is a limit, and needs 2.32.4's argument is "exact identity, then one limit". Keep the two visibly apart or the reader thinks the −12T-\tfrac12 T is approximate
units of WW"root time"Var⁡(Wt)=t\Var(W_t) = t, so WW has units time\sqrt{\text{time}}; W⋅ΔWW \cdot \Delta W and ∑W⋅ΔW\sum W \cdot \Delta W have units of time, as do 12WT2\tfrac12 W_T^2 and 12T\tfrac12 Tthe check that kills T⋅WTT \cdot W_T (time3/2\text{time}^{3/2}) before any arithmetic. Carry units through every stochastic formula; they catch factor-of-dt\sqrt{\dt} slips
telescoping sum—∑[f(ti+1)−f(ti)]=f(T)−f(0)\sum [f(t_{i+1}) - f(t_i)] = f(T) - f(0): every interior term appears once with each sign and cancelswhy ∑[W(ti+1)2−W(ti)2]=WT2\sum [W(t_{i+1})^2 - W(t_i)^2] = W_T^2 with no limit and no randomness left in it beyond the endpoint — the one exact step in the derivation. Also why ∑ΔWi=WT\sum \Delta W_i = W_T on every grid

Itô’s lemma — page 2.5

Rows marked ↺ revisit an earlier row in its new role.

SymbolSay itHere it meansWatch out
Itô's lemma—dg(Wt)=12g′′(Wt) dt+g′(Wt) dWt\dd g(W_t) = \tfrac12 g''(W_t) \,\dt + g'(W_t) \,\dW_t: the chain rule plus a second-order term, for a smooth gg of Brownian motion. Taylor to second order with (dW)2=dt(\dW)^2 = \dtit is a limit statement (the dt\dt-form), not an identity; the identity behind it is Taylor summed over the grid, the (ΔW)2(\Delta W)^2-form. 2.4's ∫W dW=12WT2−12T\int W \,\dW = \tfrac12 W_T^2 - \tfrac12 T is the case g=x2g = x^2
a dt+b dWa \,\dt + b \,\dW — order of terms"drift first"every line with both a dt\dt and a dW\dW is written with the certain part first and the random part second: dg=12g′′ dt+g′ dW\dd g = \tfrac12 g'' \,\dt + g' \,\dW, dS=μS dt+σS dW\dS = \mu S \,\dt + \sigma S \,\dW, the general formthe two orders say the same thing; the course fixes one so the eye can find the drift in the same place every time. Taylor expansions stay in power order (ΔW\Delta W, (ΔW)2(\Delta W)^2, (ΔW)3…(\Delta W)^3 \ldots): that is an expansion, not a drift-plus-noise line
12g′′(W) dt\tfrac12 g''(W) \,\dt"the Itô term", "the dt term", "the second-order term"half the curvature of gg at the current WW, times dt\dt. The survivor of 2.3: the one Taylor term whose sum is dt\dt-sized and does not diethe 12\tfrac12 is Taylor's 1/2!1/2!, not a convention — drop it and eW−12te^{W - \frac12 t} stops being a martingale. Its sign follows g′′g'': positive for exe^x and x2x^2 (means grow), 3W3W for x3x^3 (either sign)
(ΔW)2(\Delta W)^2-form · dt\dt-form—the two ways of writing the second-order term on a grid: 12∑g′′(Wi)(ΔWi)2\tfrac12\sum g''(W_i)(\Delta W_i)^2 (exact) and 12∑g′′(Wi) dt\tfrac12\sum g''(W_i) \,\dt (a limit). Their difference is the (ΔW)2(\Delta W)^2-vs-dt\dt noise, sd⁡∝1/n\sd \propto 1/\sqrt{n}== vs →\to again (2.4 ↺). On a grid the lemma is off by the noise, which comes from replacing (ΔW)2(\Delta W)^2 by dt\dt, not from Taylor — the remainder is 1/n1/n, ten to a thousand times smaller
third order and beyond · remainder—16∑g′′′(Wi)(ΔWi)3+(1/24)∑g′′′′(Wi)(ΔWi)4+…\tfrac16\sum g'''(W_i)(\Delta W_i)^3 + (1/24)\sum g''''(W_i)(\Delta W_i)^4 + \ldots: everything after the second-order term. For x3x^3 it is exactly ∑(ΔW)3\sum (\Delta W)^3; for x2x^2 it is emptyrms ∝1/n\propto 1/n, faster than 2.3's ∑∣ΔW∣3∝1/n\sum \lvert \Delta W \rvert^3 \propto 1/\sqrt{n} bound, because the signed cubes cancel (odd moment ↺). "Vanishes" here means "sums to 0 in the limit", not "each term is small"
∑(ΔW)3\sum (\Delta W)^3 vs ∑∣ΔW∣3\sum \lvert \Delta W \rvert^3signed vs unsigned cubes∑(ΔW)3\sum (\Delta W)^3 has mean 0 and sd⁡\sd 15⋅T3/2/n\sqrt{15} \cdot T^{3/2}/n; ∑∣ΔW∣3\sum \lvert \Delta W \rvert^3 has mean 22/π⋅T3/2/n2\sqrt{2/\pi} \cdot T^{3/2}/\sqrt{n}. On the coin walk: dt⋅WT\dt \cdot W_T exactly, and dt\sqrt{\dt} exactlythe bars change the rate by a factor n\sqrt{n}. 2.3 measured the unsigned one; the lemma's remainder is the signed one
E[eWt]=et/2\E[e^{W_t}] = e^{t/2}"the lognormal mean"the mean of the exponential of a Brownian motion grows at a constant rate 12\tfrac12 per unit time: all of it from the 12eW dt\tfrac12 e^W \,\dt term, none from eW dWe^W \,\dWnot ete^{\sqrt{t}}: WtW_t's size t\sqrt{t} sets the spread of eWte^{W_t}, and its variance tt sets the mean. "Mean of the exponential ≠ exponential of the mean" — the mean is held up by rare large paths while a typical path sags
eWt−12te^{W_t - \frac12 t}"the exponential martingale"the 12t\tfrac12 t in the exponent cancels the 12eW dt\tfrac12 e^W \,\dt term exactly, leaving df=f dW\dd f = f \,\dW: constant mean 1, second half independent of the firstc=12c = \tfrac12 is forced, not chosen: c=1c = 1 gives a mean falling like e−t/2e^{-t/2}, c⋅t=tc \cdot t = \sqrt{t} is not a constant and falls too. The first appearance of 2.6's −12σ2-\tfrac12\sigma^2
ftf_t, fxf_x, fxxf_{xx}, ftxf_{tx}"f sub t", "f sub x", "f sub x x", "f sub t x" — partial derivativesthe derivative of f(t,x)f(t, x) with respect to tt holding xx fixed; with respect to xx holding tt fixed; the second in xx; ftf_t then differentiated in xx (read a two-letter subscript left to right as the order of differentiation; for a smooth ff, ftx=fxtf_{tx} = f_{xt})a subscript that is a derivative, on a page full of subscripts that are indices (WiW_i, tit_i). fxf_x is not "ff at xx". Tell them apart by the letter: a variable name (tt, xx) is a derivative, a counter (ii, nn) is an index. And xx names ff's second slot, not a moving thing: WtW_t is what occupies the slot, so fx⋅ΔWf_x \cdot \Delta W pairs "sensitivity to the slot" with "how far the occupant moved" — as g′(W)⋅ΔWg'(W) \cdot \Delta W does. 2.7 names the slot after its occupant, V(t,S)V(t, S), ∂V/∂S\partial V/\partial S
dX=a dt+b dW\dd X = a \,\dt + b \,\dW"drift a, volatility b"the general one-step recipe for a process XX: a certain part a⋅dta \cdot \dt (aa is the drift, the average rate of change of XX, units of XX per unit time) plus a random part b⋅dWb \cdot \dW (bb is the volatility or diffusion coefficient, the size of the random step per time\sqrt{\text{time}}; sd⁡(ΔX)=bdt\sd(\Delta X) = b\sqrt{\dt}). WW itself is a=0a = 0, b=1b = 1; 2.6's GBM is a=μSa = \mu S, b=σSb = \sigma Sthe general Itô form is df=(∂f/∂t+a ∂f/∂x+12b2 ∂2f/∂x2) dt+b ∂f/∂x dW\dd f = (\partial f/\partial t + a\,\partial f/\partial x + \tfrac12 b^2\,\partial^2 f/\partial x^2) \,\dt + b\,\partial f/\partial x \,\dW — the b2b^2 comes from (dX)2=b2(dW)2=b2 dt(\dd X)^2 = b^2(\dW)^2 = b^2 \,\dt, the rulebook applied to XX. Everything else is the chain rule
smooth (ff, gg)—ff has continuous derivatives up to the order Taylor uses: two in xx, one in tt. x2x^2, exe^x, x3x^3, log⁡x\log x, ex−12te^{x - \frac12 t} and every pricing function on the coursean assumption about the function, never about WW, which is nowhere differentiable (2.2): the lemma is "a smooth function of a rough path". Where gg has a kink — an option payoff max⁡(S−K,0)\max(S - K, 0) at expiry — g′′g'' is a spike and the lemma needs an extra term; 2.10 meets it as gamma at the strike
(dW)2=dt(\dW)^2 = \dt ↺the rulebook, as an operation2.3's line, now used as a substitution inside a Taylor expansion: wherever (ΔW)2(\Delta W)^2 appears under a ∑\sum, replace it by dt\dt and let n→∞n \to \inftyit is licensed only under a sum: a single (ΔW)2(\Delta W)^2 is a χ12\chi^2_1 draw, and the substitution costs the (ΔW)2(\Delta W)^2-vs-dt\dt noise on any finite grid. The same substitution on a smooth xx — (Δx)2(\Delta x)^2 by dt\dt — is wrong by 12∫g′′ dt\tfrac12\int g'' \,\dt
Taylor's 12\tfrac12 and 16\tfrac16"one over two factorial, one over three factorial"the coefficients of the second- and third-order terms in g(x+Δ)−g(x)=g′Δ+12g′′Δ2+16g′′′Δ3+…g(x + \Delta) - g(x) = g'\Delta + \tfrac12 g''\Delta^2 + \tfrac16 g'''\Delta^3 + \ldotsthe 12\tfrac12 of Itô's lemma, of 2.4's −12T-\tfrac12 T, and of the lognormal's −12σ2-\tfrac12\sigma^2 are all this one 12\tfrac12. Check it on x2x^2: (x+Δ)2−x2=2xΔ+Δ2(x + \Delta)^2 - x^2 = 2x\Delta + \Delta^2, and 12⋅g′′=12⋅2=1\tfrac12 \cdot g'' = \tfrac12 \cdot 2 = 1 is the coefficient of Δ2\Delta^2. In two variables the second-order block has one 12\tfrac12 in front of four products, and the two equal cross terms ∂2f/∂x ∂t⋅dt⋅ΔW+∂2f/∂t ∂x⋅ΔW⋅dt\partial^2 f/\partial x\,\partial t \cdot \dt \cdot \Delta W + \partial^2 f/\partial t\,\partial x \cdot \Delta W \cdot \dt eat it: (a+b)2=a2+2ab+b2(a + b)^2 = a^2 + 2ab + b^2
heavy tail · s.e. of a lognormal mean—a mean that is set by rare large values: E[e2Wt]=e2t\E[e^{2W_t}] = e^{2t} is carried by paths with W≈3W \approx 3 (e6≈400e^6 \approx 400 against a typical 4). 400 paths give such a mean to tens of percent, 20,000 to a few"± s.e." from the sample sd⁡\sd undercounts when the rare values are missing from the sample. The JSON checks exe^x's levels loosely (30 %) and its rates tightly (slopes ± 0.05) for this reason; x3x^3's levels are tight (15 %)

Partial derivatives — page 2.5

These rows replace the subscript partials of the batch above.

SymbolSay itHere it meansWatch out
∂f/∂t\partial f/\partial t"partial f by partial t"the derivative of f(t,x)f(t, x) with respect to tt, holding xx fixed: how fast ff changes as the clock moves with the slot's contents left where they arethe curly ∂\partial marks a partial derivative: ff has more than one slot and the others are held fixed. It is not the d of dt\dt and dW\dW, which is an increment. In the lemma it multiplies dt\dt alone — the clock's own contribution
∂f/∂x\partial f/\partial x"partial f by partial x"the derivative of f(t,x)f(t, x) with respect to its second slot, holding tt fixedxx names ff's second slot, not a moving thing: differentiate in the placeholder, then evaluate along the path — ∂f/∂x\partial f/\partial x at (t,Wt)(t, W_t), never ∂f/∂W\partial f/\partial W. ∂f/∂x⋅ΔW\partial f/\partial x \cdot \Delta W pairs "sensitivity to the slot" with "how far the occupant moved", as g′(W)⋅ΔWg'(W) \cdot \Delta W does
∂2f/∂x2\partial^2 f/\partial x^2"second partial of f in x"∂f/∂x\partial f/\partial x differentiated in xx again, tt still held fixed: the curvature of ff in its second slotthe two 2^2s count differentiations and square nothing — one sits on ∂\partial, the other on xx. It enters the lemma in the one term the chain rule does not have, 12b2 ∂2f/∂x2\tfrac12 b^2 \,\partial^2 f/\partial x^2
ftf_t, fxf_x, fxxf_{xx}, ftxf_{tx} — retired—the subscript partials of the batch above: ftf_t is ∂f/∂t\partial f/\partial t, fxf_x is ∂f/∂x\partial f/\partial x, fxxf_{xx} is ∂2f/∂x2\partial^2 f/\partial x^2. ftxf_{tx}, read left to right, is ∂2f/∂x ∂t\partial^2 f/\partial x\,\partial t, read right to left; for a smooth ff the order does not matterno page uses them. A function of one variable keeps its primes (g′g', g′′g''); a function of time and xx takes ∂\partial. Subscripts are left to indices (WiW_i, tit_i) and times (WtW_t)

SDEs and geometric Brownian motion — page 2.6

Rows marked ↺ revisit an earlier row in its new role.

SymbolSay itHere it meansWatch out
SDE · dX=a(t,X) dt+b(t,X) dW\dd X = a(t, X) \,\dt + b(t, X) \,\dW"stochastic differential equation"2.5's one-step recipe dX=a dt+b dW\dd X = a \,\dt + b \,\dW, with the drift aa and the volatility bb now allowed to depend on where XX is and on the time. It says what happens over the next dt\dt; it is not yet a formula for XTX_Tthe d's are increments, not derivatives: WW has none (2.2). Read it as "ΔX=a Δt+b ΔW\Delta X = a \,\Delta t + b \,\Delta W for a small step", which is exactly the Euler step
dS=μS dt+σS dW\dS = \mu S \,\dt + \sigma S \,\dW · GBM"geometric Brownian motion"the price model: drift μS\mu S and volatility σS\sigma S, both proportional to the price, so the percentage step is the same whatever the price. μ\mu is the drift rate (per year), σ\sigma the volatility (per year\sqrt{\text{year}})"geometric" = multiplicative: SS can never cross zero, and log⁡S\log S is the natural variable. Not "Brownian motion of the price" — WW is the driver, SS is the exponential of a Brownian motion with drift. Written out as "Geometric Brownian motion (GBM)" at first use on every page
solve (an SDE)—write XTX_T as a formula in WW (no stepping). The route: find a function of XX whose SDE has no XX on the right, sum the steps exactly, invert. log⁡S\log S for GBMmost SDEs cannot be solved; GBM can, which is why 2.9 has a closed form at all. "Solve" never means "simulate"
ST=S0e(μ−12σ2)T+σT⋅ZS_T = S_0 e^{(\mu - \frac12\sigma^2)T + \sigma\sqrt{T} \cdot Z}"the closed-form sample", "exact sampling", "draw STS_T directly"the solution of GBM with WT=T⋅ZW_T = \sqrt{T} \cdot Z. One standard normal per path gives STS_T exactly, with no grid and no error. Decision 1: use it whenever the payoff only looks at STS_Tone draw gives STS_T only, nothing in between; but the same line holds between any two grid points, Si+1=Si⋅e(μ−12σ2)dt+σΔWiS_{i+1} = S_i \cdot e^{(\mu - \frac12\sigma^2)\dt + \sigma\Delta W_i}, so a path-dependent payoff on GBM steps exactly too. Forget the −12σ2-\tfrac12\sigma^2 and every draw is e12σ2Te^{\frac12\sigma^2 T} too high — 8 % at σ=0.4\sigma = 0.4, T=1T = 1
Euler–Maruyama · Si+1=Si(1+μ dt+σdt⋅Zi)S_{i+1} = S_i(1 + \mu \,\dt + \sigma\sqrt{\dt} \cdot Z_i)"Euler step", "Euler"the SDE read literally, one step at a time, ΔWi=dt⋅Zi\Delta W_i = \sqrt{\dt} \cdot Z_i. nn steps, nn normals per path. Decision 2: the fallback when the model has no solution (Part 4); GBM itself steps exactly, and is where Euler's cost can be measuredits STS_T differs from the exact one on the same normals by −12σ2(∑(ΔW)2−T)-\tfrac12\sigma^2(\sum (\Delta W)^2 - T) to leading order: 1 % rms at n=100n = 100, shrinking like 1/n1/\sqrt{n}. The error is 2.5's (ΔW)2(\Delta W)^2-vs-dt\dt noise, not a bias — its mean is 0
the stepping error · means ignore it, spreads keep it"strong order ½"the pathwise gap ∣STE−STexact∣\lvert S^E_T - S^{\text{exact}}_T \rvert shrinks like dt\sqrt{\dt} (measured slope −0.501) and has mean zero. A mean over paths — a price — hardly feels it; a spread over paths — a P&L distribution, a hit-rate — keeps it, because a zero-mean noise adds to a variancethe convergence of E[ST]\E[S_T] is a different, faster rate ("weak order", dt\dt; E[STE]=S0(1+μ dt)n\E[S^E_T] = S_0(1 + \mu \,\dt)^n exactly). Not "a hedge follows one path": a hedging simulator follows many and reads their spread, which is a spread
−12σ2-\tfrac12\sigma^2 · volatility drag"the drag", "the Itô correction"the lemma's 12b2⋅∂2f/∂x2\tfrac12 b^2 \cdot \partial^2 f/\partial x^2 on f=log⁡Sf = \log S with b=σSb = \sigma S: 12σ2S2⋅(−1/S2)\tfrac12\sigma^2 S^2 \cdot (-1/S^2). The mean log-return is (μ−12σ2)T(\mu - \tfrac12\sigma^2)T, not μT\mu T; the median price is S0e(μ−12σ2)TS_0 e^{(\mu - \frac12\sigma^2)T} while the mean price is S0eμTS_0 e^{\mu T}it moves the median, not the mean. With μ=0\mu = 0 and σ=0.4\sigma = 0.4 the mean price stays at S0S_0 and the median falls to 0.92S00.92 S_0; 58 % of paths end below the mean. 2.5's E[eWt]=et/2\E[e^{W_t}] = e^{t/2} is the same fact from the mean's side
mean vs median (lognormal)—for ST=S0em+sZS_T = S_0 e^{m + sZ}: median S0emS_0 e^m, mean S0em+12s2S_0 e^{m + \frac12 s^2}. The ratio e12s2e^{\frac12 s^2} is the lognormal's mean-over-median; here e0.08=1.083e^{0.08} = 1.083. In log space there is one centre: mean = median = mm; the median carries over through exp, the mean does notthe mean is held up by the tail — the 99th percentile is 2.3S02.3 S_0 for μ=0\mu = 0 — and the fraction of paths below the mean is Φ(12σT)=0.579\Phi(\tfrac12\sigma\sqrt{T}) = 0.579, not 12\tfrac12. Nobody "sees" the median: it is the point a single path is as likely to finish above as below. The expected price and the odds for one path are different things
log-return · log⁡(ST/S0)\log(S_T/S_0)—the Gaussian one: mean (μ−12σ2)T(\mu - \tfrac12\sigma^2)T, variance σ2T\sigma^2 T, skewness and excess kurtosis 0. STS_T is lognormal because its log is normalthe simple return ST/S0−1S_T/S_0 - 1 is not Gaussian and its mean is eμT−1e^{\mu T} - 1. The course's σ\sigma is the sd⁡\sd of the log-return per year\sqrt{\text{year}}, and 2.11's "returns are not lognormal" is a statement about this quantity. Write log⁡(ST/S0)\log(S_T/S_0), not log⁡ST\log S_T, unless S0=1S_0 = 1
dt\sqrt{\dt} in the step ↺—2.1's dt\sqrt{\dt}-sized kick, now inside the Euler step: ΔWi=dt⋅Zi\Delta W_i = \sqrt{\dt} \cdot Z_i, so nn kicks have total variance σ2T\sigma^2 T whatever nn isσZ⋅dt\sigma Z \cdot \dt instead of σZ⋅dt\sigma Z \cdot \sqrt{\dt} gives a volatility of σdt\sigma\sqrt{\dt} — too small by n\sqrt{n}, 100× at n=104n = 10^4, and worse the finer the grid. The tell: the drift μ dt\mu \,\dt becomes the same size as the kick
realised volatility · the three tests—n/T×sd⁡\sqrt{n/T} \times \sd of a path's per-step log-returns — the σ\sigma you would measure from the path; 0.400 on every grid for the correct step. With mean ST=S0eμTS_T = S_0 e^{\mu T} and median ST=S0e(μ−12σ2)TS_T = S_0 e^{(\mu - \frac12\sigma^2)T}, the three assertions a GBM simulator must pass (Decision 3)a per-path number with its own noise (±0.003 at n=104n = 10^4, σ/2n\sigma/\sqrt{2n}), so assert to a few percent. The mean test catches a missing −12σ2-\tfrac12\sigma^2; the volatility test catches σZ⋅dt\sigma Z \cdot \dt; the mean passes the bug, so it is not enough on its own. 2.10 compares realised with implied σ\sigma; 2.11 asks whether it is constant at all
Φ(x)\Phi(x) · N(x)N(x)"phi of x", "N of x"the probability that a standard normal is below xx — the cumulative normal. On 2.6, Φ(12σT)=Φ(0.2)=0.579\Phi(\tfrac12\sigma\sqrt{T}) = \Phi(0.2) = 0.579 is the fraction of paths below the meanthe same function is written N(x)N(x) from 2.9 on, as the index writes N(d1)N(d_1), N(d2)N(d_2) and as most option texts do; Φ\Phi appears only on 2.6. Not a density: Φ(0)=12\Phi(0) = \tfrac12, Φ(x)→1\Phi(x) \to 1 as x→∞x \to \infty

Delta hedging and the Black–Scholes PDE — page 2.7

Rows marked ↺ revisit an earlier row in its new role.

SymbolSay itHere it meansWatch out
V(t,S)V(t, S)"the option value", "VV"the price of the option at time tt when the stock is at SS — a function of two variables, the first the course has met. Taken on 2.7 from the closed form as a candidate function (2.9 derives it), licensed by the check that it satisfies the PDE — which is what any VV must satisfyVV is a surface over (t,S)(t, S), not a number; "the price" is V(0,S0)=18.02V(0, S_0) = 18.02 here. Its derivatives are partials, ∂\partial throughout
Δ=∂V/∂S\Delta = \partial V/\partial S — delta, hedge ratio"delta"how much VV moves per unit move in SS; the number of shares to short against one option so that the first-order exposure to SS cancels. 0.63 at inception here, 1 for a forward, 0 or 1 for the call at expirya function of (t,S)(t, S), so it changes along the path — that is why the hedge is rebalanced. Held fixed within a step, reset between steps. Not "the derivative with respect to time". Because Δ\Delta is taken, a change over one grid step is written δS\delta S, δV\delta V on 2.7 (2.3–2.6 wrote increments ΔW\Delta W)
Γ=∂2V/∂S2\Gamma = \partial^2 V/\partial S^2 — gamma"gamma"the curvature of VV in SS; the rate at which Δ\Delta changes. 12Γ(δS)2\tfrac12\Gamma(\delta S)^2 is the P&L a delta-hedged position makes on a move of δS\delta S, in either direction, and 2.7's numbers are all this termzero for a forward, positive for a call, largest near the strike and growing like 1/τ1/\sqrt{\tau} towards expiry. 2.10 names the rest of the Greeks; the P&L needs this one now
Θ=∂V/∂t\Theta = \partial V/\partial t — theta"theta"how VV changes as time passes with SS held fixed; negative for a call (−9.8 per year at inception here): the option decays. The PDE reads Θ=−12σ2S2Γ−rSΔ+rV\Theta = -\tfrac12\sigma^2 S^2\Gamma - rS\Delta + rV: theta pays for gamma∂/∂t\partial/\partial t at fixed SS, not d/dt\dd/\dt along a path. Often quoted per day; here per year, like rr and σ\sigma
Π=V−ΔS\Pi = V - \Delta S — the hedged portfolio"pi", "the hedged book", "the portfolio"long one option, short Δ\Delta shares; the cash difference at rr. dΠ=dV−Δ dS\dd\Pi = \dd V - \Delta \,\dS over a step on which Δ\Delta is held fixed. Self-financing: nothing goes in or out after inception; rebalances pay for themselves from cashthe money tied up is V−ΔSV - \Delta S (negative here: the shares you sold more than paid for the option), and it is that amount that earns rr. "Π\Pi" is not a product sign
rebalance — NN—resetting the share position to the current Δ\Delta at a grid time, NN times over the life. Decision 1: a static hedge (N=1N = 1) takes the spread from 29.6 to 11.8, and from there it falls like 1/N1/\sqrt{N} — 6.50 at N=4N = 4, 0.22 at N=4096N = 4096 on an 18 optiona hedge is never "done"; between rebalances it is stale by Γ⋅δS\Gamma \cdot \delta S. The 1/N1/\sqrt{N} is 2.3's ∑(ΔW)2−T\sum (\Delta W)^2 - T noise again: NN terms of sd⁡∝dt\sd \propto \dt
rr — the risk-free rate"rr", "the risk-free rate"the rate cash earns with certainty; 0.05 here. A portfolio with no dW\dW term must earn exactly rr, or there is an arbitrage. The only way rr enters the derivation, and the rate at which a price is discounted (2.8)not the stock's expected return μ\mu. Nothing on 2.7 says the stock drifts at rr; 2.8 will say that pricing as if it did gives the same VV
KK, τ=T−t\tau = T - t"strike", "time to expiry"the call pays max⁡(ST−K,0)\max(S_T - K, 0) at TT; τ\tau is how long is left. VV depends on tt only through τ\tau. K=100=S0K = 100 = S_0 here (at the money)∂V/∂t=−∂V/∂τ\partial V/\partial t = -\partial V/\partial \tau: the sign flips. The forward struck at KK has V=S−Ke−rτV = S - Ke^{-r\tau}, Δ=1\Delta = 1, Γ=0\Gamma = 0
the Black–Scholes PDE"the PDE", "Black–Scholes"∂V/∂t+12σ2S2⋅∂2V/∂S2+rS⋅∂V/∂S−rV=0\partial V/\partial t + \tfrac12\sigma^2 S^2 \cdot \partial^2 V/\partial S^2 + rS \cdot \partial V/\partial S - rV = 0, with V(T,S)V(T, S) = the payoff. What VV must satisfy for a delta-hedged position to earn rr; a property any pricer can be tested against (residual 10−1510^{-15} on the closed form, h2h^2 by finite differences)μ\mu is not in it. σ\sigma and rr are. "The formula" (2.9) is one solution of it, for one payoff; the PDE holds for every European payoff on this stock. 3.3 solves it on a grid
μ\mu vanishes ↺—2.6's drift μ\mu, cancelled from dΠ\dd\Pi by the same choice of Δ\Delta that cancels dW\dW, because both enter through (∂V/∂S−Δ)(\partial V/\partial S - \Delta). Decision 2: a hedger needs no view on direction. Measured: the hedged mean and sd⁡\sd at μ=−0.2\mu = -0.2, 0, +0.3 agree within five s.e. from N=64N = 64. Not true of a static hedge: at N=1N = 1 the spread at μ=0.3\mu = 0.3 is 17.6 against 11.8 — drift-blindness is bought by rebalancingit vanishes from the hedged book, not from the stock: the unhedged mean swings from −11.7 to +22.1. On a coarse grid (N=4N = 4) an O(dt)O(\dt) trace of it survives, +0.7 on a spread of 7
realised vs implied volatility ↺"σrealised\sigma_{\text{realised}}", "σimplied\sigma_{\text{implied}}"2.6's realised volatility is the σ\sigma the stock actually moves with; the implied is the σ\sigma the price and the hedge were built from. Decision 3: hedged at σi\sigma_i while the stock moves at σr\sigma_r, the mean P&L is (σr2−σi2)×∑e−rt⋅12ΓS2 dt(\sigma_r^2 - \sigma_i^2) \times \sum e^{-rt} \cdot \tfrac12\Gamma S^2 \,\dt — +3.74 for 0.5 against 0.4, −3.76 for 0.3the gap in variances, not in vols, times a path-dependent gamma integral. "Implied" is named for inverting 2.9's formula, which 3.4 does; here it only means "the σ\sigma in VV". That P&L keeps a spread (1.3–1.7) no rebalancing removes
realised quadratic variation (of the stock) ↺—2.3's ∑(ΔW)2\sum (\Delta W)^2 with the stock in place of WW: ∑(δS)2\sum (\delta S)^2 along the path, the activity the hedge is left exposed to. Per step its mean is σr2S2 dt\sigma_r^2 S^2 \,\dt, so the gamma P&L's mean is 12ΓS2(σr2−σi2) dt\tfrac12\Gamma S^2(\sigma_r^2 - \sigma_i^2) \,\dtdirection is ∑δS\sum \delta S, which the hedge removes; activity is ∑(δS)2\sum (\delta S)^2, which it does not. The hook's "gamma acting on realised quadratic variation" is this row
gamma P&L — 12ΓS2((δS/S)2−σ2 dt)\tfrac12\Gamma S^2((\delta S/S)^2 - \sigma^2 \,\dt)"the gamma P&L", "the hedging error"what a delta-hedged position makes on one step: the move squared against what the price assumed for it. Mean zero when σr=σi\sigma_r = \sigma_i, spread ∝dt\propto \dt, summed over NN steps ∝1/N\propto 1/\sqrt{N}. 2.5's (dW)2−dt(\dW)^2 - \dt noise with a price on it; 2.10's "a delta-hedged option earns 12ΓS2(σr2−σi2) dt\tfrac12\Gamma S^2(\sigma_r^2 - \sigma_i^2) \,\dt" is its meana second-order quantity: it exists only because VV is curved. A linear payoff has none, and that is the course's whole distinction. Path by path it reproduces the P&L with a residual that falls faster than the P&L (ratio 0.54 → 0.06 across the grids)

Risk-neutral pricing — page 2.8

Rows marked ↺ revisit an earlier row in its new role.

SymbolSay itHere it meansWatch out
PP — the real-world measure"PP", "the real-world measure", "the physical measure"the probabilities 2.6's model assigns to paths when the stock drifts at its actual expected return μ\mu; the measure every page up to 2.7 worked in without naming it. EP[ST]=S0eμT\E^P[S_T] = S_0 e^{\mu T}named only now that there is a second one. Nothing about PP changed on 2.8; what changed is which measure the price is an expectation under. Not "the real distribution" — still the model's
QQ — the risk-neutral measure"QQ", "the risk-neutral measure", "the pricing measure"the same model with the drift set to rr: dS=rS dt+σS dW~\dS = rS \,\dt + \sigma S \,\dd\tilde{W}. Under QQ every discounted traded price is a martingale — the stock, the forward, the call, the hedged book — and prices are discounted expectations. Measured: e−rtSte^{-rt}S_t has mean S0S_0 at every checkpoint; the QQ-run's average is the pricea device, not a belief: nobody thinks the stock drifts at rr, and EP[ST]\E^P[S_T] stays S0eμTS_0 e^{\mu T}. "Risk-neutral" names the investor who would believe it, and no such investor enters the argument — the hedge does. That there is exactly one QQ here is 2.11's to show (one noise, one traded stock)
EQ[⋅]\E^Q[\cdot] — expectation under QQ"the QQ-expectation", "the risk-neutral expectation"the average over paths drawn with the rr-drift — or over paths drawn with any drift, each weighted by LL. The quantity a Monte Carlo pricer estimatesnot EP\E^P. The two differ by e−rT⋅EP[payoff]−V=3.34e^{-rT} \cdot \E^P[\text{payoff}] - V = 3.34 at μ=0.1\mu = 0.1, 21.0 at 0.3. The superscript is not decoration
V=e−rT⋅EQ[payoff]V = e^{-rT} \cdot \E^Q[\text{payoff}] — the pricing formula"discount the risk-neutral expectation"2.8's result: the price is the QQ-expectation of the payoff, discounted at rr. For V(t,S)V(t, S), condition on St=SS_t = S and discount over τ\tau. Measured: 18.010 ± 0.031 against 18.023; the surface over (t,S)(t, S) within 1.7 s.e. at 32 pointsboth the drift inside the expectation and the discount rate are rr, for different reasons: the discount is the time value of the payoff; the drift is what makes the hedge break even. EP\E^P discounted at rr is not a price
the rr-drift — "as if the stock drifted at rr""simulate under QQ", "the rr-drift"what a simulator does to price: ST=S0⋅e(r−12σ2)T+σT⋅ZS_T = S_0 \cdot e^{(r - \frac12\sigma^2)T + \sigma\sqrt{T} \cdot Z}. The only change from 2.6's exact draw is the number in the drift slotσ\sigma is untouched. The −12σ2-\tfrac12\sigma^2 is 2.6's drag, unchanged; the QQ-median of STS_T is S0e(r−12σ2)T=97.0S_0 e^{(r - \frac12\sigma^2)T} = 97.0, below S0S_0 — "drifts at rr" is about the mean
W~\tilde{W} — the QQ-Brownian motion"W-tilde", "the Brownian motion under QQ"W~t=Wt+λt\tilde{W}_t = W_t + \lambda t: 2.2's Brownian motion with a tilt added, which under QQ is a Brownian motion with no tilt. The noise the stock has under the rr-drift: dS=rS dt+σS dW~\dS = rS \,\dt + \sigma S \,\dd\tilde{W}the same path as WW — the same random numbers — read with a different drift absorbed into it. Under PP it is "Brownian motion with drift λ\lambda" (2.2); under QQ it is Brownian motion. Which one it is depends on the measure, not on the path
λ=(μ−r)/σ\lambda = (\mu - r)/\sigma — the market price of risk"lambda", "the market price of risk"how far the real-world drift sits above rr, in units of σ\sigma: excess return per unit of volatility. 0.625 at μ=0.3\mu = 0.3, −0.625 at −0.2, 1.375 at 0.6. The size of the tilt Girsanov removesnot needed to price — it is exactly what QQ eliminates, and the price does not contain it. It sets the cost of reweighting (e−λ2Te^{-\lambda^2 T} of the paths) and nothing else on 2.8. The name is finance's, not 2.8's argument
L=e−λWT−12λ2TL = e^{-\lambda W_T - \frac12\lambda^2 T} — the likelihood ratio, Girsanov's weight"the weight", "the likelihood ratio", "dQ/dP\dd Q/\dd P"the factor that turns a PP-average into a QQ-average on the same paths: EQ[X]=EP[L⋅X]\E^Q[X] = \E^P[L \cdot X]. 2.5's exponential martingale eaW−12a2Te^{aW - \frac12 a^2 T} with a=−λa = -\lambda, so mean L=1L = 1. Measured: the weighted μ\mu-paths average to the price at every μ\mu in the table; mean LL = 1.000 ± 0.001 at every μ\mu up to 0.3a weight per path, not per average: dividing the μ=0.3\mu = 0.3 average by e(μ−r)Te^{(\mu-r)T} gives 30.3, not 18.0, because the division corrects only the part of the payoff proportional to STS_T, not the strike. Mean L2=eλ2TL^2 = e^{\lambda^2 T}, so the weights have a spread — the ESS row
Girsanov's theorem"Girsanov", "the change of measure"adding a drift λ\lambda to a Brownian motion is undone by reweighting its paths by LL: the tilted process is a Brownian motion under the reweighted measure. Here: the μ\mu-stock and the rr-stock are the same paths with different weights. Named on 2.8 and measured, not proved: weighted mean log-return (r−12σ2)T(r - \tfrac12\sigma^2)T, weighted realised vol σ\sigmachanges the drift and only the drift — "reweighting paths, not changing volatility". A change of volatility is not a change of measure; no reweighting turns a σ=0.4\sigma = 0.4 path into a 0.5 one
Feynman–Kac"Feynman–Kac"the discounted QQ-expectation of the payoff, as a function of (t,S)(t, S), solves 2.7's PDE with the payoff as its condition at TT — because e−rtV(t,St)e^{-rt}V(t, S_t) is a QQ-martingale and Itô's lemma says its dt\dt term is the PDE. Named on 2.8, sketched in the "why" section, measured: VMCV_{\text{MC}} agrees with the closed form at 32 grid points within 1.7 s.e. and its PDE residual is the size of its Monte Carlo noisethe link between three pricers — Monte Carlo (2.8), the closed form (2.9), finite differences (3.3) — and the reason the course can ask them to agree. It says the expectation solves the PDE; it does not say how to evaluate either
s.e. = sd⁡/n\sd/\sqrt{n} — the Monte Carlo standard error ↺"standard error", "s.e."2.1's n\sqrt{n} in its working clothes: the sd⁡\sd of the mean of nn draws is the sd⁡\sd of one draw over n\sqrt{n}. ±0.95 at 10310^3 paths, ±0.03 at 10610^6. Measured, not assumed: 100 pricers of 10410^4 paths scatter by 0.298 and each reported 0.312a Monte Carlo price is never quoted without it. 100× the paths for 10× the precision. "Within standard error" means within one; the checks use five
ESS = n⋅e−λ2Tn \cdot e^{-\lambda^2 T} — effective sample size"effective sample size", "ESS"(∑L)2/∑L2(\sum L)^2/\sum L^2: how many equally-weighted paths a reweighted average of nn is worth. 68 % of nn at ∣λ∣=0.625\lvert\lambda\rvert = 0.625, 15 % at 1.375 (0.158 measured). The cost of simulating at the wrong drift and fixing it with weightscounts the weights, not the estimator: paths drawn towards the payoff can give a smaller s.e. than the QQ-run (0.015 against 0.031 at μ=0.3\mu = 0.3) — importance sampling, which 3.2 takes up. The reason to simulate under QQ is that it needs no weights for any payoff
e−rtSte^{-rt}S_t — the discounted stock, a QQ-martingale ↺"the discounted stock"2.5's martingale, with the thing that is a martingale now named: under QQ, e−rtSt=S0⋅eσW~t−12σ2te^{-rt}S_t = S_0 \cdot e^{\sigma\tilde{W}_t - \frac12\sigma^2 t}, 2.5's exponential martingale with a=σa = \sigma, mean S0S_0 at every tt. Under PP it drifts at μ−r\mu - r: a submartingale for μ>r\mu > r (128.6 at TT from 100), a supermartingale for μ<r\mu < r (77.8)a martingale is not a constant: its sd⁡\sd grows to 41.8 by TT. "Martingale" is a statement about the mean given the past; "martingale" needs "under QQ" after it to be a sentence

The Black–Scholes formula — page 2.9

Rows marked ↺ revisit an earlier row in its new role.

SymbolSay itHere it meansWatch out
N(x)N(x) — the cumulative normal ↺"N of x"the probability that a standard normal is below xx: 2.6's Φ\Phi under the name the formula is always written with. N(d2)=0.470N(d_2) = 0.470, N(d1)=0.627N(d_1) = 0.627 on 2.9N(−x)=1−N(x)N(-x) = 1 - N(x) is what turns the call's two pieces into the put's; NN is the area, not the density φ\varphi
1{ST>K}\mathbf{1}\{S_T > K\} — the indicator of exercise"one if STS_T is above KK"1 on the paths that finish above the strike, 0 elsewhere; max⁡(ST−K,0)=(ST−K)⋅1{ST>K}\max(S_T - K, 0) = (S_T - K) \cdot \mathbf{1}\{S_T > K\} is what splits the price into two piecesits expectation is a probability, Q(ST>K)=N(d2)Q(S_T > K) = N(d_2); multiplied by STS_T it picks out the stock's value on those paths, not their average price
d2d_2 — the exercise threshold"d two"ST>KS_T > K exactly when Z>−d2Z > -d_2: the distance from the strike to the QQ-median of ln⁡ST\ln S_T, in standard deviations. d2=(ln⁡(S0/K)+(r−12σ2)T)/(σT)=−0.075d_2 = (\ln(S_0/K) + (r - \tfrac12\sigma^2)T)/(\sigma\sqrt{T}) = -0.075 on 2.9 (the median 97.0 sits just below KK)born from the inequality, not chosen; negative when the median is below the strike
d1=d2+σTd_1 = d_2 + \sigma\sqrt{T} — the threshold after the stock-weighting shift"d one"completing the square in eσT⋅z−12z2e^{\sigma\sqrt{T} \cdot z - \frac12 z^2} moves the normal's centre by σT\sigma\sqrt{T}; d1=(ln⁡(S0/K)+(r+12σ2)T)/(σT)=0.325d_1 = (\ln(S_0/K) + (r + \tfrac12\sigma^2)T)/(\sigma\sqrt{T}) = 0.325 on 2.9the 12σ2\tfrac12\sigma^2 enters with a plus in d1d_1 and a minus in d2d_2; swapping them is the commonest implementation error, and the oracle catches it
N(d2)N(d_2) — the risk-neutral probability of exercise"N d two"Q(ST>K)=0.470Q(S_T > K) = 0.470: the share of rr-drift paths that finish above the strike, 2.8's 47.0 % now named. Measured 0.4707 ± 0.0005not the real-world probability (70.8 % at μ=0.3\mu = 0.3) and not the delta
N(d1)N(d_1) — the share-measure probability of exercise; the delta"N d one"the probability of exercise when each path is weighted by ST/EQ[ST]S_T/\E^Q[S_T]: EQ[ST⋅1{ST>K}]/EQ[ST]=0.627\E^Q[S_T \cdot \mathbf{1}\{S_T > K\}]/\E^Q[S_T] = 0.627, the share of the stock's forward value that comes from the exercising paths. Also ∂C/∂S0\partial C/\partial S_0, the delta 2.7 hedged with. Measured 0.6278 ± 0.0005; bump 0.62741always above N(d2)N(d_2); not "the probability of finishing in the money". EQ[ST∣ST>K]=140.3\E^Q[S_T \mid S_T > K] = 140.3 is a different number and is not in the price
S0N(d1)S_0 N(d_1) and Ke−rTN(d2)K e^{-rT} N(d_2) — the two pieces"the stock piece", "the strike piece"the discounted stock received on exercise, 62.74, and the discounted strike paid on exercise, 44.72; their difference is the priceeach is measurable on the draws (62.79 ± 0.07, 44.77 ± 0.05); the price is not "KK times a probability"
C=S0N(d1)−Ke−rTN(d2)C = S_0 N(d_1) - K e^{-rT} N(d_2) — the Black–Scholes call price"the formula", "Black–Scholes"2.8's expectation evaluated: 18.023 on 2.9. The function 2.7 checked against the PDE and 2.8 borrowed, now derived; the function Part 3 implements and testsexact — the Monte Carlo estimate is the approximation. Exact for this payoff and this model; 2.11 is where it breaks
P=Ke−rTN(−d2)−S0N(−d1)P = K e^{-rT} N(-d_2) - S_0 N(-d_1) — the put price"the put"the same integral over the other side of the strike: 13.146 on 2.9; measured 13.122 ± 0.017a European put can sit below its intrinsic value K−S0K - S_0 deep in the money, because Ke−rT<KK e^{-rT} < K; that is not a bug
C−P=S0e−qT−Ke−rTC - P = S_0 e^{-qT} - K e^{-rT} — put–call parity"parity"the call and the put differ by the forward, path by path and in price: 4.877 on 2.9, measured 4.889 ± 0.042needs no model — the payoffs differ by ST−KS_T - K on every path and the discounted stock is a QQ-martingale; the formula obeys it, it does not create it
qq — the continuous dividend yield"q", "the yield"cash the stock pays its holder at the rate qq; under QQ the price drifts at r−qr - q and S0S_0 becomes S0e−qTS_0 e^{-qT} in the formula. The call falls (18.02 → 16.21 at q=0.03q = 0.03), the put risesqq is not μ\mu: the hedge removes the real-world drift, not the yield. Discrete dividends are 2.11's
F=S0e(r−q)TF = S_0 e^{(r - q)T} — the forward price"F", "the forward"EQ[ST]\E^Q[S_T]: 105.13 at q=0q = 0, 102.02 at q=0.03q = 0.03. The one number the market quotes for delivery at TT, whoever receives the dividendsthe forward's price, not the forward contract's value (S0e−qT−Ke−rTS_0 e^{-qT} - K e^{-rT}, 4.88); 2.8's forward was the contract
Black-76 — the formula written in FF"Black seventy-six"C=e−rT[F⋅N(d1)−K⋅N(d2)]C = e^{-rT}[F \cdot N(d_1) - K \cdot N(d_2)], d1=(ln⁡(F/K)+12σ2T)/(σT)d_1 = (\ln(F/K) + \tfrac12\sigma^2 T)/(\sigma\sqrt{T}): rr only in the discount, qq folded into FF. The same number as the other two spellings to 2×10−142 \times 10^{-14}not a second model: the lognormal with its mean named differently. Prices options on forwards, futures, FX and swap rates because FF can be any quoted forward

Greeks — page 2.10

Rows marked ↺ revisit a 2.7 row in its new role: computed from the formula rather than taken as a black box.

SymbolSay itHere it meansWatch out
Greek — a partial derivative of the price"the Greeks", "a Greek"∂V/∂(one input)\partial V/\partial(\text{one input}), every other input held still: the first-order answer to "if this moves and nothing else does, how much does my book move?" Five on 2.10 — Δ\Delta, Γ\Gamma, vega, Θ\Theta, ρ\rho — all from 2.9's formula by one identityfirst order only: the change on a finite move is Δ⋅δS+12Γ(δS)2+…\Delta \cdot \delta S + \tfrac12\Gamma(\delta S)^2 + \dots, and the gap (0.632 against 0.627) is the next Greek's job. A Greek is a function of (t,S)(t, S), not a number; "the delta" is its value now
Δ=∂C/∂S0=N(d1)\Delta = \partial C/\partial S_0 = N(d_1) — delta ↺"delta"2.7's hedge ratio, now differentiated out of the formula: the S0φ(d1)S_0 \varphi(d_1) and Ke−rTφ(d2)K e^{-rT} \varphi(d_2) terms the chain rule produces cancel, leaving N(d1)=0.627N(d_1) = 0.627. Put: N(d1)−1=−0.373N(d_1) - 1 = -0.373moves with the stock — 0.637 at S=101S = 101 — which is Γ\Gamma. Not the probability of exercise (that is N(d2)N(d_2), 2.9), though it is one under the stock-weighted measure
Γ=∂2C/∂S2=φ(d1)/(S0σT)\Gamma = \partial^2 C/\partial S^2 = \varphi(d_1)/(S_0 \sigma\sqrt{T}) — gamma ↺"gamma"the delta's own sensitivity to spot; 0.00946 here. The normal density in ln⁡S\ln S: peaks below the strike at long expiry (74.8 at T=1T = 1), at it as T→0T \to 0, growing like 1/T1/\sqrt{T}. The same for the put"largest at the strike" is a short-expiry statement. Long the option is long gamma: 12Γ(δS)2≥0\tfrac12\Gamma(\delta S)^2 \ge 0 on a move either way — the term the hedge does not remove
S0φ(d1)=Ke−rTφ(d2)S_0 \varphi(d_1) = K e^{-rT} \varphi(d_2) — the identity—the two density terms are equal (37.84 both), because the exponents differ by exactly ln⁡(Ke−rT/S0)\ln(K e^{-rT}/S_0). Every input but KK moves d1d_1 and d2d_2 together, so the chain-rule pair cancels — delta, vega, rho and theta each fall out in one linenot a coincidence of the base case: an identity in every input. The reason rho is "discount only" even though rr sits in d1d_1 and d2d_2
vega=∂C/∂σ=S0φ(d1)T\text{vega} = \partial C/\partial \sigma = S_0 \varphi(d_1) \sqrt{T} — vega"vega"how much the price moves per 1.00 of volatility: 37.84 here, 0.378 per vol point. Positive, no direction in it, equal for call and put, largest near the strike (103.0 at T=1T = 1), growing like T\sqrt{T}not a Greek letter; the vol trader's Greek, because an option is a position in σ\sigma. Per unit on 2.10 and in the oracle; a desk's "vega" is usually per point — ask which
Θ=∂V/∂t\Theta = \partial V/\partial t — theta ↺"theta"computed: −S0φ(d1)σ/(2T)−rKe−rTN(d2)=−7.568−2.236=−9.804-S_0 \varphi(d_1) \sigma/(2\sqrt{T}) - r K e^{-rT} N(d_2) = -7.568 - 2.236 = -9.804 per year, −0.027 per day. Most negative near the strike (−10.20 at S=114S = 114); the carry alone deep in the money (−4.76); zero deep out. Put: −5.05∂/∂t\partial/\partial t at fixed SS = −∂/∂T-\partial/\partial T. The first term is −12σ2S2Γ-\tfrac12\sigma^2 S^2 \Gamma: theta is gamma's bill. Per year here; per day on a desk
ρ=∂C/∂r=KTe−rTN(d2)\rho = \partial C/\partial r = K T e^{-rT} N(d_2) — rho"rho"the discount factor's derivative, the dd-terms having cancelled: +44.72 per 1.00 of rr, 0.0045 per basis point. The strike piece's TT-weight. Put: −KTe−rTN(−d2)=−50.40-K T e^{-rT} N(-d_2) = -50.40positive for the call (paying KK later is cheaper when rates are higher), negative for the put. Not zero — the hedge removes μ\mu, not rr. Small for short equity options; the rates desk's Greek
bump-and-revalue"bump", "finite-difference Greek"(V(x+h)−V(x−h))/2h(V(x + h) - V(x - h))/2h, and Γ\Gamma by the second difference. The check on every analytic form here (10−710^{-7} at 216 points, call and put); the only Greek there is when there is no formula (3.5)two errors in tension: truncation ∝h2\propto h^2, round-off ∝ε/h\propto \varepsilon/h (ε/h2\varepsilon/h^2 for Γ\Gamma). Γ\Gamma's best hh here is 10−4×S10^{-4} \times S, not the 10−210^{-2} a hand estimate gives; the study is 3.5's
theta pays for gamma — the PDE as P&L"theta pays for gamma", "long gamma, short theta"2.7's Θ+12σ2S2Γ+rSΔ−rV=0\Theta + \tfrac12\sigma^2 S^2 \Gamma + rS\Delta - rV = 0 read per unit time on the hedged book: −Θ=12σ2S2Γ+carry-\Theta = \tfrac12\sigma^2 S^2 \Gamma + \text{carry} — the theta paid is what the curvature is priced to earn at σi\sigma_i. Measured per step on 4,000 paths: a gamma gain of 7.58 ± 0.05 against a theta cost of 9.71 less carry 2.14an identity in expectation, not per step: the gamma gain depends on how the stock actually moves, the theta cost does not. Their difference is 2.7's gamma P&L, (σr2−σi2)×(\sigma_r^2 - \sigma_i^2) \times the gamma integral — −3.80 / +3.73 at σr=0.3\sigma_r = 0.3 / 0.5
gamma gain · theta cost · carry — the step attribution—∑e−rt12Γ(δS)2\sum e^{-rt} \tfrac12\Gamma(\delta S)^2, ∑e−rtΘ δt\sum e^{-rt} \Theta \,\delta t, ∑e−rt(rSΔ−rV) δt\sum e^{-rt} (rS\Delta - rV) \,\delta t along a path: the three pieces of the hedged book's P&L once Δ δS\Delta \,\delta S is removed, plus Taylor's remainder (rms 0.014 at N=4096N = 4096, 0.11 at 256)the remainder falls about eightfold from N=256N = 256 to 4096; the theta cost's spread (sd⁡\sd 3.3) does not — it is a path integral, and how much theta you pay depends on where the stock spent the year
raw partials — the units"per unit", "raw"vega per 1.00 of σ\sigma, Θ\Theta per year, ρ\rho per 1.00 of rr: what the formula gives, what the golden file stores, what the C++ returnsa desk scales at display — ÷100, ÷365, ÷10410^4 — and only there; a convention factor inside the oracle would have to match in three places. "Vega 37.8" and "vega 0.378" are the same number
∂V/∂q=−TS0e−qTN(d1)\partial V/\partial q = -T S_0 e^{-qT} N(d_1) — the dividend-rho"dividend rho"the stock piece's TT-weight, as ρ\rho is the strike piece's: −62.74 here. Carried by the oracle and checked against the closed form at 7,234 rowsnot one of 2.10's five; named once so the golden file's seventh column has a home

Where it breaks — page 2.11

Rows marked ↺ revisit an earlier row in its new role.

SymbolSay itHere it meansWatch out
σi\sigma_i — implied volatility ↺"the implied vol", "implied"the σ\sigma that makes 2.9's formula return the market's price: 0.400 for 18.02. Unique because vega>0\text{vega} > 0 everywhere; found by Newton on σ\sigma in four steps here. 2.7's σi\sigma_i, now backed out of a price rather than assumeda price in different units, not a forecast of σr\sigma_r. Defined only for prices strictly inside the bounds; not recoverable where vega has vanished (the deep and one-day rows)
(S0−Ke−rT)+<C<S0(S_0 - K e^{-rT})^+ < C < S_0 — the bounds"the no-arbitrage bounds"the call's range as σ\sigma runs from 0 to ∞\infty: 4.877 to 100 at the base case (put: (Ke−rT−S0)+(K e^{-rT} - S_0)^+ to Ke−rTK e^{-rT}). Every price between has one implied vol; a price outside has none and the solver refuses it687 golden rows sit on the lower bound (within 10−1210^{-12}): no time value at double precision, so no vol
the smile"the smile", "the vol smile"σi\sigma_i against KK at one expiry when it is not flat: for the fair mixture a U with its bottom at the forward (0.398), 0.477 at K=60K = 60, 0.452 at K=160K = 160its minimum is at the forward, not the spot; symmetric in ln⁡(K/F)\ln(K/F), not in KK. A single-σ\sigma model has none
the skew"skew", "the skew"a tilted smile: higher σi\sigma_i at lower strikes (0.448 at K=60K = 60 against 0.372 at 160) when the high-vol state carries the lower forwardthe fair mixture has none; direction comes from asymmetry between the states, not from fat tails alone
σi(K,T)\sigma_i(K, T) — the volatility surface"the surface", "the vol surface"the smile at every expiry, as one object: what a production pricer takes in instead of a σ\sigma. Named on 2.11; built on 3.4, drawn by D92.11 measures one expiry. The surface's own Greeks (vanna, volga) are named, not computed
the lognormal mixture — the synthetic market"the mixture"STS_T drawn from one of two 2.6 models, σ=0.2\sigma = 0.2 or 0.6 with equal chance, the state fixed for the life of the option and unknown when it is sold; forwards tied so E[ST]=F\E[S_T] = F. Every price is 12C(0.2)+12C(0.6)\tfrac12 C(0.2) + \tfrac12 C(0.6)a market built to break one assumption at a time, not a model of any stock. Each component alone passes the kurtosis test on 2.6
excess kurtosis — fat tails"kurtosis", "fat tails"the fourth central moment over the variance squared, less the normal's 3: 1.80 for the mixture's log⁡ST\log S_T, 1.92 for its daily returns, 0 for any lognormal. Mass beyond 3 sd⁡\sd: 1.23 % against 0.27 %the check that catches the mixture where mean and variance do not. Within each vol state the returns are Gaussian: the fat tails are the mixing
p(1−p)⋅∣C(σhigh)−C(σlow)∣\sqrt{p(1 - p)} \cdot \lvert C(\sigma_{\text{high}}) - C(\sigma_{\text{low}}) \rvert — the between-state spread"the unhedgeable piece", "incompleteness"7.54 here: the sd⁡\sd of a coin toss between the two states' prices — uncertainty about which state the option was sold in, fixed before the state became known. What is left of the hedged book's spread when the Greeks are right (the oracle hedge: 7.542 at N=4096N = 4096)incompleteness, not a hedge failing: with only the stock to trade, the state is a risk nothing spans. The Black–Scholes hedge adds a within-state 2.83 in quadrature: 2.7's gamma P&L with the wrong σi\sigma_i
kk — the proportional cost"ten basis points", "the fee"the charge per unit of stock traded: 0.001 here, on every share at setup, each rebalance and the unwind. The rebalancing cost is k⋅∑∣δΔ∣⋅Sk \cdot \sum \lvert \delta\Delta \rvert \cdot S and grows like N\sqrt{N}: 0.485 at N=256N = 256, 1.94 at 4096a fee, not a bid–ask model; spread + cost is smallest at an interior NN (256 on the grid). Leland (1985) folds it into an adjusted σ\sigma — v2
"vol is the unit of price""forty vol", "quoted in vol"quoting σi\sigma_i instead of CC: the stock moves 100 → 110 and the price quote 18.02 is stale (the call is worth 24.74 at 40 vol) while the vol quote is not. Vol removes the inputs that are usually observable (SS, KK, rr, TT) and leaves the quantity options are quoted and compared incomparable across strikes, expiries and days; the reason the surface, not a price list, is the market's object
vanna, volga — the surface's Greeks"vanna", "volga"∂Δ/∂σ\partial \Delta/\partial \sigma and ∂vega/∂σ\partial \text{vega}/\partial \sigma: how the hedge and the vega exposure move when the surface moves. Named on 2.10's "Into 2.11" and on 2.11 as where the Greeks go nextnot computed on 2.11; 3.4 and v2

Check

A page says E[Sₙ] = 0 and, two lines later, E|Sₙ| ≈ 0.8√n. Which is right?

In E[W₂ | W₁ = 1.5], what does the vertical bar mean?

dW appears in dS = μS dt + σS dW. What is dW?

A demo slider reads "Vol 30%". What number does the code carry, and for what period?

The site says r = 0.05, continuously compounded. What is £100 worth in one year?