Finance, Probability Theory, and Stochastic Calculus

Comprehensive Course Summary

MODULE 1: FINANCE OVERVIEW

1.1 Basic Financial Literacy

The Story of Money and Markets

Money is a medium of exchange, a unit of account, and a store of value. Financial markets
aggregate information from millions of participants to produce prices. Key evolution: barter
→ commodity money → fiat money → digital assets. Price discovery is the market process by
which buyers and sellers interact to set equilibrium prices.
Information Markets & Efficient Market Hypothesis
Efficient Market Hypothesis (EMH): Asset prices fully reflect all available information at all
times. Three forms:

1. Weak - prices reflect past trading data;

2. Semi-strong - prices reflect all publicly available information;

3. Strong - prices reflect all information, including insider information.

If markets are efficient, no trading strategy consistently beats the market after adjusting for
risk. Prices follow a random walk. EMH implies that abnormal returns are zero on average
(studied via event studies like Q5 of the Midsem).
⋆ Note: Event studies test semi-strong EMH: if a public announcement causes persistent
abnormal returns, the market is not semi-strong efficient.
Company Valuation

• Discounted Cash Flow (DCF): Value = sum of future free cash flows discounted at
the required rate of return (WACC or cost of equity).
X CFt
V0 =
t
(1 + r)t

• Price-to-Earnings (P/E): Market price per share / Earnings per share – measures how
much investors pay per unit of earnings.

• Book Value: Total assets - Total liabilities (accounting-based floor valuation).

• Market Capitalization: Share price × Shares outstanding (market-based total equity

value).

1
 • Dividend Discount Model (Gordon Growth): P0 = D1
r−g where D1 is next dividend,
r is required return, g is growth rate.

ESOPs (Employee Stock Ownership Plans)

ESOPs grant employees rights to buy company stock at a fixed price (the strike price K) after
a vesting period. They function like call options on company stock. They align employee
incentives with shareholder value creation. If ST > K, the employee exercises and gains ST − K
per share. The vesting period creates a lock-in effect encouraging long-term thinking.
Savings and Investment Strategies (Basic)

• Fixed Income: Bonds, FDs – predictable cash flows, lower risk.

• Equity: Stocks – higher expected return, higher volatility.

• Diversification: Combining assets with low correlation reduces portfolio variance with-
out reducing expected return.

• Portfolio Frontier: The set of minimum-variance portfolios for a given expected return
(Markowitz).

• Systematic saving via SIPs (Systematic Investment Plans) exploits rupee-cost averaging.

Different Assets in the Money Market

• T-Bills (Treasury Bills): Short-term government debt, near risk-free, zero coupon,
matures in 91/182/364 days.

• Commercial Paper: Short-term unsecured corporate IOUs, typically 7-270 days, issued
at a discount.

• Certificates of Deposit (CDs): Bank-issued time deposits with fixed tenure and in-
terest rate.

• Repo Agreements: Short-term collateralized borrowing – sell securities now, repurchase

at higher price later.

• Money Market Mutual Funds: Pool investing in short-term instruments, NAV main-
tained near 1.

• Call Money: Overnight interbank lending – the most liquid segment of the money
market.

Idea of Risk
Risk: Uncertainty about future returns. In finance, quantified primarily as variance or standard
deviation of returns.

• Systematic Risk (Market Risk): Market-wide risk that cannot be eliminated through
diversification. Driven by macro factors - interest rates, inflation, GDP. Measured by β.

• Idiosyncratic Risk (Specific Risk): Firm-specific risk – management decisions, prod-

uct failures. Can be eliminated by holding a diversified portfolio.

2
 Total Risk = Systematic Risk + Idiosyncratic Risk
Var(Ri ) = βi2 Var(Rm ) + Var(ϵi )
(from market model Ri = αi + βi Rm + ϵi )
Cov(Ri ,Rm )
⋆ Note: In CAPM: E[Ri ] = Rf + βi (E[Rm ] − Rf ). Here βi = Var(Rm ) . Only systematic risk
is priced.

1.2 Stocks and Derivatives

Stocks and Financial Markets

• Stock (Equity): A share representing fractional ownership in a company. Entitles holder

to proportional dividends and residual claims in liquidation.

• Primary Market: Where new securities are issued – IPOs, FPOs, rights issues.

• Secondary Market: Trading of already-issued securities (NSE, BSE, NYSE, NASDAQ).

• Market Microstructure: Bid-ask spread, order book depth, price impact of large trades
– directly relevant to liquidity discount (Q3 of Quiz 2).

Financial Instruments: How Are They Made?

Financial instruments are contracts with payoff profiles. They are constructed to transfer risk
between counterparties. Building blocks: underlying assets (stocks, bonds, currencies, com-
modities) + contractual terms (strike, expiry, delivery mechanism).

• Plain vanilla instruments: Forward, futures, calls, puts.

• Structured products: Combinations e.g., straddle (call + put at same strike), collar
(buy put + sell call).

The key construction principle: No-Arbitrage. Any derivative payoff must be replicable
by a portfolio of traded assets, otherwise arbitrage exists.
Futures and Options

• Forward Contract: OTC agreement to buy/sell an asset at a future date T at forward

price F agreed today. Zero initial cost. No daily settlement. Credit risk exists.

• Futures Contract: Exchange-traded, standardized forward with daily mark-to-market

settlement (variation margin). No credit risk. Forward price ≈ Futures price when interest
rates are constant.

• Option: Gives the holder the RIGHT (not obligation) to buy (call) or sell (put) an asset
at strike price K on or before expiry T . In exchange for this right, the holder pays the
option premium upfront.

American option: exercisable any time up to T . European option: exercisable only at T .

(Black-Scholes applies to European options.)
Call and Put Options - Payoffs and Diagrams

3
 • Call: Right to BUY at K. Profitable when stock price ST > K (stock went up beyond
your cost).
Call payoff at expiry = (ST − K)+ = max(ST − K, 0)
Call profit = (ST − K)+ − C · erT (where C = premium paid)

• Put: Right to SELL at K. Profitable when stock price ST < K (stock fell below your
floor).
Put payoff at expiry = (K − ST )+ = max(K − ST , 0)
Put profit = (K − ST )+ − P · erT (where P = premium paid)

⋆ Note: Put-Call Parity for European options: C − P = S0 − Ke−rT . This is a model-free

no-arbitrage result.
Idea of Hedging
Hedging: Taking an offsetting position to reduce exposure to price risk. A perfect hedge elimi-
nates ALL price risk.

• Example 1: An airline (fears rising jet fuel costs) buys oil futures → locks in fuel price
→ hedges oil price risk.

• Example 2: A portfolio manager holds stock worth S0 and buys a put option at strike
K. This is a protective put. Payoff floor = K. This is also the payoff of a call + bond
(by put-call parity).

Delta Hedging: A dynamic hedging strategy. Hold ∆ = ∂V ∂S shares of stock per option to make
the portfolio instantaneously risk-free. The BSM equation arises from requiring a delta-hedged
portfolio to earn the risk-free rate.
Arbitrage
Arbitrage: A trading strategy that: (1) requires zero initial investment, (2) has non-negative
payoff in all scenarios, (3) has strictly positive payoff with positive probability. Riskless profit
from zero cost.
No-Arbitrage Principle: In efficient markets, arbitrage opportunities disappear immediately
as traders exploit them. All consistent pricing formulas (risk-neutral pricing, BSM, binomial
model) are derived from No-Arbitrage.

• Example: If Gold = $1800 in New York and $1820 in London: buy in NY, simultaneously
sell in London → $20 risk-free. In practice, transaction costs and execution speed limit
real arbitrage.

EXAM ALERT: No-arbitrage in binomial model: d < 1 + r < u. If this fails, one can
construct a riskless portfolio with positive return.
Information Asymmetry
When one party to a transaction has more/better information than the other. Leads to:

• Adverse Selection (pre-contract): The “lemons problem” - the uninformed party

selects from a pool biased toward bad risks (e.g., only sellers of bad stocks offer to trade).

• Moral Hazard (post-contract): Once insured/financed, party takes more risk (e.g., a
CEO with a salary cap has no upside incentive to take risk, but ESOPs fix this).

4
In options markets: insider trading exploits information asymmetry. Market microstructure
models (Kyle, Glosten-Milgrom) model how informed trading affects bid-ask spreads.
Discounted Price & Risk-Free Rate
Discount Factor: The present value of $1 received at time t: β(t) = e−rt in continuous com-
pounding, or (1 + r)−t in discrete time. The risk-free rate r is the rate of return on a zero-risk
investment.
Discounted price: S̃t = e−rt St
The discounted stock price S̃t is a martingale under the risk-neutral measure Q. This is the
cornerstone of derivative pricing - it means the expected discounted future price equals today’s
price under Q.
Continuous vs. Discrete: erT ≈ (1 + nr )nT as n → ∞ (continuous compounding limit). In BSM,
continuous compounding is used throughout.
How to Quantify Risk?

• Variance/Standard Deviation: σ 2 = E[(X − µ)2 ] – the classical measure. Problem:

penalizes upside and downside equally.

• Value at Risk (VaR): The maximum loss at confidence level α over horizon T . P(Loss >
VaR) = 1 − α. Problem: not sub-additive (does not always reward diversification).

• Expected Shortfall (CVaR): E[Loss | Loss > VaR] – average loss in the worst (1 − α)
fraction. Coherent risk measure.

• Beta (β): Cov(Ri , Rm )/Var(Rm ) - measures systematic risk. β = 1 moves with market,
β > 1 amplifies market moves, β < 1 dampens.

• Sharpe Ratio: (E[R] − Rf )/σ - reward per unit of total risk. Higher = better risk-
adjusted performance.

• Market Model: Ri,t = αi + βi Rm,t + ϵi,t where ϵi,t is idiosyncratic noise. Used in event
studies (Midsem Q5).

MODULE 2: PROBABILITY THEORY

2.1 Discrete Probability

Basic Probability Theory (Classical/Traditional Definition)

Classical Probability: For a finite equally-likely sample space Ω with n outcomes: P(A) =
|A|/|Ω|. Requires all outcomes to be equally likely.
Kolmogorov Axioms (the foundation of modern probability):

1. Non-negativity: P(A) ≥ 0 for all events A

2. Normalization: P(Ω) = 1

3. Countable
P Additivity (σ-additivity): For pairwise disjoint events A1 , A2 , · · · : P(∪i Ai ) =
i P(A i )

Derived results:

• P(Ac ) = 1 − P(A)

5
 • P(A ∪ B) = P(A) + P(B) − P(A ∩ B) (inclusion-exclusion)

• Independence: P(A ∩ B) = P(A) · P(B)

• Conditional Probability: P(A | B) = P(A∩B)

P(B) for P(B) > 0

• Bayes’ Theorem: P(A | B) = P(B|A)P(A)

P(B) = P(B|A)P(A)
P(B|A)P(A)+P(B|Ac )P(Ac )

Infinite Probability Space - Why Classical Definition Fails

The classical definition P(A) = |A|/|Ω| breaks down when Ω is infinite or uncountable. For
example: Pick a random real number X in [0, 1]: P(X = x) = 0 for any specific x, yet
P(X ∈ [0, 1]) = 1. We cannot sum uncountably many zeros to get 1 using classical probability.
The Vitali Set construction (assuming Axiom of Choice) shows that some subsets of [0, 1] cannot
be assigned any consistent probability. So we cannot allow ALL subsets of R to be events –
we need a restricted class of “measurable” sets. This motivates the axiomatic framework: we
specify exactly which subsets are “events” via a sigma-algebra.

2.2 Continuous Probability & Axiomatic Construction

The Role of Available Information in Probability Theory

A key insight: probability theory models not just randomness but also the state of available
information. Different agents may have different information sets. The sigma-algebra F formal-
izes “what is known.” A larger F means more information is available. This is why filtrations
(sequences of sigma-algebras) model information flow over time – crucial for stochastic processes
in finance.
Axiomatic Construction: The Triple (Ω, F , P)

• Outcome Space Ω: The set of ALL possible outcomes of the random experiment. E.g.,
Ω = {H, T }∞ for infinite coin tosses (each ω is an infinite sequence of H and T), or Ω = R
for a Gaussian random variable, or Ω = C([0, T ]) (continuous paths) for Brownian motion.

• Sigma-Algebra F: A collection F of subsets of Ω satisfying three axioms: (i) Ω ∈ F

(the whole space is measurable); (ii) A ∈ F ⇒ Ac ∈ F (closed under complements); (iii)
If A1 , A2 , · · · ∈ F, then ∪i Ai ∈ F (closed under countable unions). Elements of F are
called “events” or “measurable sets.”

• Probability Measure P: A function P : F → P [0, 1] satisfying: P(∅) = 0, P(Ω) = 1, and

for pairwise disjoint A1 , A2 , · · · ∈ F : P(∪i Ai ) = i P(Ai ). This is the Kolmogorov axiom
of countable additivity.

Probability Space: The triple (Ω, F , P).

Why Do We Need an Axiomatic Construction of Probability?

1. Handle uncountable spaces: Classical definition fails for continuous distributions (e.g.,
uniform on [0, 1], Gaussian). We cannot assign positive probability to each point.

2. Avoid paradoxes: Non-measurable sets (Vitali sets) exist in R. If we tried to assign

probability to ALL subsets, we get contradictions. The sigma-algebra F restricts us to
“well-behaved” events.

6
 3. Unify discrete and continuous: The single axiomatic framework handles both coin
tosses (discrete) and Brownian motion (continuous) without changing the theory.
4. Model information: The sigma-algebra F represents available information. Restricting
to sub-sigma-algebras G ⊆ F models “partial information” - essential for conditional
expectation and martingales.

⋆ Note: Midsem Q6 asked exactly this. The diagram shows Ω → X (random variable) → R,
with F ⊇ σ(X) below Ω, and B(R) below R. P lives on (Ω, F), µX lives on (R, B(R)).
Sigma-Algebra - Key Examples

• Trivial σ-algebra: F = {∅, Ω}. Represents NO information (you know nothing beyond
“something happens”).
• Discrete σ-algebra: F = 2Ω (all subsets). Appropriate for finite/countable Ω – repre-
sents FULL information.
• Borel σ-algebra B(R): Generated by all open intervals in R. Contains all open sets,
closed sets, all sets you can write down explicitly. The “right” σ-algebra for real-valued
random variables.
• Generated sigma-algebra σ(C): The smallest σ-algebra containing a collection C of
sets.

Closure properties of a sigma-algebra: If A, B ∈ F , then A ∩ B = (Ac ∪ B c )c ∈ F (closed under

finite intersections by De Morgan and axioms). Also closed under countable intersections.
Borel Sigma-Algebra B(R) - Detailed
B(R): The smallest σ-algebra on R containing all open intervals (a, b) for a < b in R. Equiva-
lently generated by: all open sets, all closed sets, all half-open intervals (−∞, x] for x ∈ R. B(R)
contains: all open sets, all closed sets, all countable unions of intervals, all countable intersec-
tions (Gδ sets), all points {x} (since {x} = ∩n (x − 1/n, x + 1/n)), all intervals [a, b], (a, b], [a, b),
all of R and ∅. B(R) does NOT contain Vitali sets (these require Axiom of Choice to construct
and are non-measurable).
⋆ Note: For the purpose of this course: B(R) contains every set you can explicitly describe or
construct.
Uniform Lebesgue Measure on [0,1]
Lebesgue Measure λ: On (R, B(R)): λ((a, b)) = b − a for any interval. The natural extension of
“length” to Borel sets. Extends uniquely from intervals to all of B(R) by Carathéodory’s Exten-
sion Theorem. The Uniform distribution on [0, 1]: The probability space is (Ω = [0, 1], B([0, 1]), λ).
For any Borel set B ⊆ [0, 1], P(X ∈ B) = λ(B) = length of B. In particular, P(a ≤ X ≤ b) =
b − a for 0 ≤ a ≤ b ≤ 1. Properties of Lebesgue measure: (1) λ({x}) = 0 for any point (count-
able sets have measure zero); (2) σ-finiteness: R = ∪n [−n, n] with λ([−n, n]) = 2n < ∞; (3)
Translation invariant: λ(A + c) = λ(A).
Random Variable Formal Definition
Random Variable: A function X : Ω → R (or Rn ) such that X is F /B(R)-measurable: for
every Borel set B ∈ B(R), the pre-image X −1 (B) = {ω ∈ Ω : X(ω) ∈ B} ∈ F. Measurability
ensures P(X ∈ B) is well-defined. Intuition: X is a random variable if knowing “which event
in F occurred” is sufficient to determine X(ω). Non-measurable X would mean P(X ∈ B) asks
for the probability of a non-event.

• Discrete RV: Takes countably many values x1 , x2 , . . . . Characterized by PMF pi =

P(X = xi ). Examples: Bernoulli(p), Binomial(n,p), Poisson(λ), Geometric(p).

7
 • Continuous RV: Has a density fX (x) ≥ 0 with fX (x)dx = 1, such that P(a ≤ X ≤
R
Rb
b) = a fX (x)dx. Examples: Uniform, Normal/Gaussian, Exponential.

 2

Key: N (µ, σ 2 ) has density f (x) = √ 1 exp − (x−µ)
2σ 2 . The standard normal N (0, 1) has
2πσ 2
CDF Φ(x) = P(Z ≤ x).
Distribution Measure of a Random Variable
The distribution (or law) of X is the pushforward measure µX on (R, B(R)) defined by:

µX (B) = P(X −1 (B)) = P(X ∈ B) for B ∈ B(R)

Characterized equivalently by the CDF: FX (x) = P(X ≤ x) = µX ((−∞, x]). Two RVs can have
the same distribution even on different probability spaces (e.g., −Bt and Bt both have N (0, t)
distribution – both are Brownian motions, as seen in Midsem Q1(a)).
Filtration - The Mathematics of Information Flow
Filtration: An increasing family of σ-algebras {Fn }n≥0 (discrete) or {Ft }t≥0 (continuous) with
Fs ⊆ Ft for all s ≤ t. Intuitively: information only accumulates, never disappears. Examples:

• Natural filtration of {Xn }: Fn = σ(X0 , X1 , . . . , Xn ) – the minimum information to

determine X0 through Xn .

• Natural filtration of Brownian motion: Ft = σ(Bs : 0 ≤ s ≤ t) - information revealed

by the BM path up to time t.

• Public filtration (Midsem Q4): Ft = σ(all publicly observed actions up to time t). Per-
(i)
sonal filtration of player Pi : Gt = σ(Ci , publicly observed actions up to time t). Note
(i)
Gt ⊇ Ft since player Pi knows their own card Ci in addition to public info.

Sigma-Algebra Generated by a Random Variable

σ(X): The σ-algebra generated by X: σ(X) = {X −1 (B) : B ∈ B(R)} = the smallest σ-
algebra making X measurable. σ(X) represents exactly the information contained in knowing
the value of X - no more, no less. If Y = f (X) for measurable f , then σ(Y ) ⊆ σ(X). For
naturalPfiltration: Fn = σ(ξ1 , . . . , ξn ) = σ(Xn , Xn−1 , . . . , X1 , X0 ) for a symmetric random walk
Xn = ni=1 ξi . Both generate the same information because knowing all past steps is equivalent
to knowing all past positions.
G-Measurable Random Variable
G-measurable: A random variable Y is G-measurable (for σ-algebra G ⊆ F) if σ(Y ) ⊆ G.
Equivalently, for every B ∈ B(R), {Y ∈ B} ∈ G. Intuition: Y is G-measurable ⇔ the value of
Y can be completely determined from the information in G. Example: Xn (the n-th position of
SRW) is Fn -measurable but NOT Fn−1 -measurable (you need the n-th step to know the n-th
position).
Stochastic Process
Stochastic Process: A collection {Xt }t∈T of random variables on a common probability space
(Ω, F, P), indexed by time t. For each t, Xt : Ω → R is a random variable. For each ω, the
map t 7→ Xt (ω) is a sample path (or trajectory) of the process. Types by time index: Discrete-
time (T = {0, 1, 2, . . . }) or continuous-time (T = [0, ∞)). Examples: Symmetric random
walk (discrete), Brownian motion (continuous), stock price process (continuous), portfolio value
process.
Adapted Stochastic Process
Adapted Process: A stochastic process {Xn }n≥0 is adapted to filtration {Fn }n≥0 if for each

8
n ≥ 0, Xn is Fn -measurable. No “peeking into the future.” The process cannot anticipate
events that have not yet occurred. Intuition: At each time n, Xn is known given the information
available at time n. Financial trading strategies MUST be adapted (you can only trade based
on current and past information, not future prices). Theorem: The symmetric random walk
Xn = ξ1 + · · · + ξn is adapted to Fn = σ(ξ1 , . . . , ξn ), since Xn = f (ξ1 , . . . , ξn ) is clearly Fn -
measurable.
RT
⋆ Note: The Itô integral 0 ∆(t)dBt requires the integrand ∆(t) to be adapted. This is the
“non-anticipating” condition – we use the LEFT endpoint value ∆(tj ) in the Riemann sum, not
the right.
Pictorial Illustration of Axiomatic Probability (Midsem Q6 Diagram)
The diagram shows: On the left is Ω (the sample space), containing outcome ω. An arrow
labeled X points right to R (the real line), where X(ω) lives. Below Ω is the layer F (sigma-
algebra on Ω), which contains σ(X) (the sub-sigma-algebra generated by X). Pre-images of
Borel sets B from B(R) land in σ(X) ⊆ F. The probability measure P : F → [0, 1] sits on the
left. The distribution measure µX : B(R) → [0, 1] sits on the right, with µX (B) = P(X −1 (B)).
This illustrates: P measures events in Ω; µX measures events in R; they are connected by X
through the pre-image. X is measurable ⇒ pre-images of Borel sets are events ⇒ probabilities
like P(X ≤ x) are well-defined.

MODULE 3: ASSET PRICING MODELS

3.1 Binomial Asset Pricing Models

General Concept - What Is Asset Pricing?

Asset pricing answers: “What is the fair price today of a financial contract (derivative) that pays
off in the future?” The binomial model is the simplest model with all the key features: discrete
time, two possible outcomes per step, risk-neutral pricing. As the number of steps n → ∞ and
step size → 0, the binomial model converges to the Black-Scholes continuous model.
One-Period Binomial Asset Pricing Model
Setup: Stock price today S0 . Risk-free rate r per period (so $1 today → $(1 + r) tomorrow).
At time 1, stock price is either:

• S1 = uS0 (up-move, probability p under real-world measure P)

• S1 = dS0 (down-move, probability 1 − p under real-world measure P)

Where u > d > 0 are the up-factor and down-factor.

No-Arbitrage Condition: d < 1 + r < u.

• If 1 + r ≤ d: borrow cash at r, buy stock → always profit (stock always beats cash).
Arbitrage!

• If 1 + r ≥ u: short stock, invest cash at r → always profit (cash always beats stock).
Arbitrage!

Risk-Neutral Probability: Unique q ∈ (0, 1) such that the discounted stock price is a mar-
tingale:  
Q S1 q · uS0 + (1 − q) · dS0
S0 = E =
1+r 1+r

9
 1+r−d
⇒q=
u−d
Note: q is determined solely by r, u, d - NOT by the real-world probability p. This is the key
non-intuitive insight: derivative prices do not depend on real-world beliefs about the stock.
Pricing a derivative with payoff V1 (payoff at time 1):
 
q · V1 (up) + (1 − q) · V1 (down) V1
V0 = = EQ
1+r 1+r

Replicating Portfolio - The Arbitrage-Free Approach

Find ∆ shares and B in the risk-free bond such that the portfolio replicates the derivative
payoff:

∆ · uS0 + B(1 + r) = V1 (up) . . . (1) (1)

∆ · dS0 + B(1 + r) = V1 (down) . . . (2) (2)

Subtract (2) from (1): ∆ = V1 (up)−V1 (down)

(u−d)S0 [the “delta” = hedge ratio]. Then: B = V1 (down)−∆·dS
1+r
0
.
No-arbitrage price: V0 = ∆·S0 +B. This must equal the risk-neutral formula (both are correct).
Captain Bubbles Problem (Midsem Q3) - Full Solution Method
Setup: Stock i has up-factor ui = 1 + pi , down-factor di = 1 − pi where pi = Ri /(Ri + Wi ).
Risk-free rate r = 10%. Given risk-neutral probabilities qA = 0.60, qB = 0.625, qC = 0.75.

• Part (a): pi = Ri /(Ri + Wi ) (probability of drawing red ball = probability of up-move

in real world).
(1+r)−(1−p) r+p
• Part (b): From q = 1+r−d
u−d = 2p = 2p :

r + pi r
qi = ⇒ pi =
2pi 2qi − 1
0.10 0.10
pA = 2(0.60)−1 = 0.20 = 0.50
0.10 0.10
pB = 2(0.625)−1 = 0.25 = 0.40
0.10 0.10
pC = 2(0.75)−1 = 0.50 = 0.20

• Part (c): Smallest positive integers Ri , Wi with pi = Ri /(Ri + Wi ): pA = 0.5 = 1/2 ⇒

(R, W ) = (1, 1). pB = 0.4 = 2/5 ⇒ (R, W ) = (2, 3). pC = 0.2 = 1/5 ⇒ (R, W ) = (1, 4).

• Part (d): ui = 1 + pi , di = 1 − pi : uA = 1.5, dA = 0.5; uB = 1.4, dB = 0.6; uC = 1.2, dC =

0.8. Verify no-arbitrage: di < 1.1 < ui for all three stocks. ✓
(k)
• Part (e): Terminal stock prices S3 = S0 · uki · d3−k
i for k = 0, 1, 2, 3.

Multi-Period Binomial Asset Pricing Model

Extend to n periods. At each node (time j, state k = number of up-moves), the stock can go
up by u or down by d. The stock price at time n after k up-moves and (n − k) down-moves:

Sn(k) = S0 · uk · dn−k

Risk-neutral pricing by BACKWARD INDUCTION (working backwards from terminal payoff):

Vn(k) = h(Sn(k) ) [terminal payoff, e.g., (Sn(k) − K)+ for a call]

(k+1) (k)
(k) q · Vj+1 + (1 − q) · Vj+1
Vj = [j = n − 1, n − 2, . . . , 0]
1+r

10
Direct risk-neutral pricing formula (n-period):
n  
1 1 X n k
V0 = n
Q
E [Vn ] = n
q (1 − q)n−k Vn(k)
(1 + r) (1 + r) k
k=0

Independence in the Binomial Model

The up/down moves ξ1 , ξ2 , . . . , ξn are i.i.d. under BOTH P and Q. Under Q: PQ (ξj = u) = q,
PQ (ξj = d) = 1 − q for each j, independently. The independence of increments in the binomial
model parallels the independence of BM increments.
Conditional Expectation - Definition and Intuition
Conditional Expectation E[X | G]: Given R R G ⊆ F, E[X | G] is the unique G-
σ-algebra
measurable random variable Y satisfying A Y dP = A XdP for all A ∈ G. It is the “best
prediction” of X given only the information in G (in the L2 sense,
P it minimizes E[(X − Z)2 ]
over all G-measurable Z). For discrete
R RV: E[X | Y = y] = x x · P(X = x | Y = y). For
continuous RV: E[X | Y = y] = x · fX|Y (x | y)dx. Existence and uniqueness: Guaranteed by
the Radon-Nikodym Theorem.
Properties of Conditional Expectation - Complete List

1. Linearity: E[aX + bY | G] = aE[X | G] + bE[Y | G]

2. Tower Property (Law of Iterated Expectations): If H ⊆ G ⊆ F, then E[E[X | G] |

H] = E[X | H]. “Outer (smaller) σ-algebra wins.”

3. Taking Out Known Quantities: If Z is G-measurable, then E[ZX | G] = Z · E[X | G].

4. Independence: If X is independent of G (i.e., independent of every A ∈ G), then E[X |

G] = E[X].

5. Positivity: If X ≥ 0 a.s., then E[X | G] ≥ 0 a.s.

6. Total Expectation: E[E[X | G]] = E[X]. (Take G = {∅, Ω} in the Tower Property.)

7. Consistency: If X is already G-measurable, then E[X | G] = X.

8. Contractivity: E[|E[X | G]|] ≤ E[|X|].

⋆ Note: The Tower Property is the single most important property for martingale proofs. For
SRW: E[Xn+1 | Fn ] = E[Xn + ξn+1 | Fn ] = Xn + E[ξn+1 | Fn ] = Xn + E[ξn+1 ] = Xn (used
independence since ξn+1 ⊥ Fn ).
Jensen’s Inequality - Statement, Proof, and Applications
Jensen’s Inequality: If φ : R → R is convex and X is an integrable random variable, then
E[φ(X)] ≥ φ(E[X]). A function φ is convex if for all x, y and λ ∈ [0, 1]: φ(λx + (1 − λ)y) ≤
λφ(x) + (1 − λ)φ(y). Equivalently (for differentiable φ): φ′′ (x) ≥ 0.
Proof (using supporting hyperplane): By convexity, for any point µ, there exists a slope c
(the subgradient at µ) such that φ(x) ≥ φ(µ) + c(x − µ) for all x. Setting µ = E[X] and taking
expectations:
E[φ(X)] ≥ φ(E[X]) + c · E[X − E[X]] = φ(E[X]) ✓
Conditional version: If φ is convex, E[φ(X) | G] ≥ φ(E[X | G]) a.s.
Key applications in this course:

• φ(x) = x2 : E[X 2 ] ≥ (E[X])2 ⇒ Var(X) ≥ 0.

11
 • φ(x) = ex : E[ex ] ≥ eE[X] - moment generating function inequality.

• φ(x) = |x| : E[|X|] ≥ |E[X]|.

• φ(x) = (x − K)+ : E[(X − K)+ ] ≥ (E[X] − K)+ – call option price lower bound.

⋆ Note: Midsem Q1(c): If (Xn , Fn ) is a martingale and φ is convex with φ(Xn ) integrable,
then (φ(Xn ), Fn ) is a SUBMARTINGALE. Proof of Q1(c): E[φ(Xn+1 ) | Fn ] ≥ φ(E[Xn+1 |
Fn ]) = φ(Xn ). Hence φ(Xn ) is a submartingale. ✓
Martingales and their Properties
Martingale: A stochastic process (Mn , Fn )n≥0 is a martingale if: (i) Mn is Fn -measurable
(adapted); (ii) E[|Mn |] < ∞ for all n; (iii) E[Mn+1 | Fn ] = Mn a.s. for all n ≥ 0.

• Supermartingale: E[Mn+1 | Fn ] ≤ Mn a.s. → tends to DECREASE on average. Think:

a losing gambler’s fortune.

• Submartingale: E[Mn+1 | Fn ] ≥ Mn a.s. → tends to INCREASE on average. Example:

φ(martingale) for convex φ.

Equivalent martingale characterization: For all s ≤ t : E[Mt | Fs ] = Ms .

Key properties:

• Constant expectation: E[Mn ] = E[M0 ] for all n. (Set G = {∅, Ω} in the martingale
property.)

• Optional Stopping Theorem (OST): Under mild conditions, E[Mτ ] = E[M0 ] for a
stopping time τ . This is used to solve hitting time problems.

• Doob’s Martingale Inequalities: Bounds on the maximum of a martingale in terms

of its terminal value.

Important Martingale Examples - with Proofs

1. Symmetric Random Walk Xn : E[Xn+1 | Fn ] = Xn + E[ξn+1 ] = Xn . ✓ Martingale.

2. Xn2 −n: E[Xn+1

2 −(n+1) | Fn ] = E[(Xn +ξn+1 )2 | Fn ]−(n+1) = Xn2 +2Xn ·0+1−n−1 =
2
Xn − n. ✓ Martingale.

3. Xn3 − 3nXn : (see Module 4 / Quiz 1 Q1 for full proof).

 n
2
4. Sn = eσXn eσ +e −σ . This is the stochastic exponential of SRW. To verify: E[Sn+1 |
    σ −σ
2 2 e +e
Fn ] = Sn · eσ +e −σ E[eσξn+1 ] = Sn · eσ +e −σ 2 = Sn . ✓

 n
2
⋆ Note: The Midsem Q2(c) process Sn = eσXn eσ +e−σ
is a martingale. Proof: E[Sn+1 |
eσ +e−σ
Fn ] = Sn × factor × 2 = Sn (the factor cancels the moment).
Risk-Neutral Measure - Formal Definition
Risk-Neutral Measure Q: An equivalent probability measure Q ∼ P (same null sets: P(A) =
0 ⇔ Q(A) = 0) under which the discounted asset price S̃n = (1 + r)−n Sn is a Q-martingale. In
the n-period binomial model: risk-neutral probabilities q = (1 + r − d)/(u − d) for the up-move

12
at each step. Key property: ANY derivative price equals EQ [discounted payoff]. This is the
Fundamental Theorem.
 
Q Vn
V0 = E (risk-neutral pricing formula)
(1 + r)n
Under Q, stocks grow at the risk-free rate r (not their true rate µ). This makes math tractable
but does NOT mean all investors are risk-neutral.
First Fundamental Theorem of Asset Pricing (FFTAP) - Discrete
FFTAP (Discrete): A financial market model is ARBITRAGE-FREE if and only if there
EXISTS at least one equivalent martingale measure (EMM/risk-neutral measure) Q equivalent
to P under which all discounted asset prices are martingales. Formal arbitrage definition: A
trading strategy θ (a sequence of portfolio positions) is an arbitrage if: X0 (θ) = 0 (zero initial
cost), Xn (θ) ≥ 0 a.s. (non-negative payoff), P(Xn (θ) > 0) > 0 (positive probability of gain).
No-arbitrage ⇔ EMM ⇔ the risk-neutral probabilities q ∈ (0, 1).

MODULE 4: STOCHASTIC CALCULUS

4.1 Symmetric Random Walk (SRW) Detailed

Definition
Let ξ1 , ξ2 , . . . be i.i.d. with P(ξj = +1) = P(ξj = −1) = 1/2. Define X0 = 0 and:
Xn = ξ1 + ξ2 + · · · + ξn , n≥1
Natural filtration: Fn = σ(ξ1 , . . . , ξn ). Note ξn+1 ⊥ Fn (independence of future increments
from past).
SRW as a Martingale - Proof
Claim: (Xn , Fn ) is a martingale. Proof:

1. Adapted: Xn = ξ1 + · · · + ξn is a function of (ξ1 , . . . , ξn ) hence Fn -measurable. ✓

2. Integrable: E[|Xn |] ≤ E[|ξ1 | + · · · + |ξn |] = n < ∞. ✓
3. Martingale: E[Xn+1 | Fn ] = E[Xn + ξn+1 | Fn ] = Xn + E[ξn+1 | Fn ] = Xn + E[ξn+1 ] =
Xn + 0 = Xn . ✓ (Used: Xn is Fn -measurable; ξn+1 ⊥ Fn so its conditional expectation
equals its unconditional expectation = 0)

Expectation and Variance of SRW

E[Xn ] = 0 (by linearity; each ξi has mean 0).
Var(Xn ) = E[Xn2 ] = n (since i.i.d. with E[ξi2 ] =P1 and cross Pterms vanish by independence).
Variance proof: E[Xn2 ] = E[(ξ1 + · · · + ξn )2 ] = i E[ξi2 ] + 2 i<j E[ξi ]E[ξj ] = n · 1 + 0 = n.
Key Martingale: Xn2 − n - Full Proof
Claim: Mn = Xn2 − n is a martingale. Proof:
2
E[Mn+1 | Fn ] = E[Xn+1 − (n + 1) | Fn ]
= E[(Xn + ξn+1 )2 | Fn ] − (n + 1)
= E[Xn2 + 2Xn ξn+1 + ξn+1
2
| Fn ] − (n + 1)
= Xn2 + 2Xn · E[ξn+1 | Fn ] + E[ξn+1
2
| Fn ] − (n + 1)
= Xn2 + 2Xn · 0 + 1 − (n + 1) 2
(since ξn+1 ⊥ Fn , E[ξn+1 2
| Fn ] = E[ξn+1 ] = 1)
= Xn2 − n = Mn ✓

13
Quiz 1 Q1: Xn3 − 3nXn is a Martingale - Full Proof
Problem: Find an such that Zn = Xn3 + an Xn is a martingale with a0 = 0.

• Step 1: Write Xn+1 = Xn + ξn+1 . Expand the cube:

3
Xn+1 = (Xn + ξn+1 )3 = Xn3 + 3Xn2 ξn+1 + 3Xn ξn+1
2 3
+ ξn+1

• Step 2: Take E[· | Fn ]. Use ξn+1 ⊥ Fn so E[f (ξn+1 ) | Fn ] = E[f (ξn+1 )]:
3
E[Xn+1 | Fn ] = Xn3 + 3Xn2 · 0 + 3Xn · 1 + 0 = Xn3 + 3Xn
2 ] = 1, E[ξ 3 ] = 0)
(Used: E[ξn+1 ] = 0, E[ξn+1 n+1

• Step 3: E[Zn+1 | Fn ] = E[Xn+1

3 | Fn ] + an+1 E[Xn+1 | Fn ] = (Xn3 + 3Xn ) + an+1 · Xn =
3
Xn + (3 + an+1 )Xn .

• Step 4: For martingale, need E[Zn+1 | Fn ] = Zn = Xn3 + an Xn so:

3 + an+1 = an ⇒ an+1 = an − 3

With a0 = 0: a1 = −3, a2 = −6, . . . , an = −3n.

Answer: Zn = Xn3 − 3nXn is an (Fn )-martingale.

4.2 Scaled Symmetric Random Walk (SSRW)

Definition and Motivation

√
The SSRW is constructed by compressing both space and time. For step size 1/ n (in space)
and time unit 1/n:
1
W (n) (t) = √ X⌊nt⌋ for t ∈ [0, T ]
n
√
This is the standard “n-step” scaling: each step covers time 1/n and distance ±1/ n. The
√
scaling factor 1/ n comes from the CLT: for i.i.d. mean-zero, unit-variance RVs, the sum of n
√ √
terms has standard deviation n, so dividing by n gives variance n · (1/n) = 1.
Properties of SSRW

• E[W (n) (t)] = 0 (since E[Xn ] = 0)

• Var(W (n) (t)) = (1/n) · ⌊nt⌋ → t as n → ∞ (since Var(X⌊nt⌋ ) = ⌊nt⌋)

The SSRW is also a martingale (it is a scaled version of the SRW martingale; scaling does not
destroy the martingale property).
Quadratic Variation of SSRW
Quadratic Variation: For a process X over [0, T ] with partition Π = {0 = t0 < t1 < · · · <
tm = T }: X
[X, X]T = lim (Xti+1 − Xti )2
∥Π∥→0
i
√
For the SSRW: each increment W (n) (tk+1 ) − W (n) (tk ) has magnitude |1/ n| (always!), so
squared increment = 1/n. Over nT steps total (for t in [0, T ]):

[W (n) , W (n) ]T = (number of steps) × (step size)2 = nT × (1/n) = T

14
The quadratic variation of the SSRW equals T - the SAME as Brownian motion! This is
why SSRW converges to BM: they share the same quadratic variation (non-trivial QV = t
distinguishes BM from all smooth paths).
⋆ Note: For a differentiable function f : [f, f ]T = 0 (squared increments
√ of order (∆t)2 → 0).
BM is different: its QV = T > 0 arising from increments of order ∆t.
Convergence of SSRW to BM
By Donsker’s Theorem (Functional CLT): W (n) (t) → Bt in distribution on C([0, T ]) as n → ∞.
√
For each fixed t, by the ordinary CLT: W (n) (t) = (1/ n)X⌊nt⌋ → N (0, t) in distribution.

4.3 Brownian Motion (BM) - Complete Theory

Definition (Standard Brownian Motion)

Standard Brownian Motion: A stochastic process B = {Bt }t≥0 on (Ω, F, P) satisfying:

1. B0 = 0 almost surely

2. Independent increments: for 0 ≤ t0 < t1 < · · · < tn , increments Bt1 − Bt0 , Bt2 −
Bt1 , . . . , Btn − Btn−1 are mutually independent

3. Stationary increments: Bt − Bs ∼ N (0, t − s) for all 0 ≤ s ≤ t

4. Continuous sample paths: t 7→ Bt (ω) is continuous for a.e. ω ∈ Ω

Basic BM Properties

• E[Bt ] = 0, Var(Bt ) = t, E[Bt2 ] = t

• Cov(Bs , Bt ) = E[Bs Bt ] = min(s, t) for s, t ≥ 0

2 t/2
• Bt ∼ N (0, t) ⇒ E[eθBt ] = eθ (moment generating function)

Proof of E[Bs Bt ] = min(s, t): WLOG s ≤ t. E[Bs Bt ] = E[Bs (Bs + (Bt − Bs ))] = E[Bs2 ] +
E[Bs (Bt − Bs )] = s + E[Bs ] · E[Bt − Bs ] = s + 0 · 0 = s = min(s, t). (Used independence of Bs
and Bt − Bs .)
BM as a Martingale - Proof
Claim: Bt is an (FtB )-martingale. Proof: For s ≤ t: E[Bt | Fs ] = E[Bs + (Bt − Bs ) | Fs ] =
Bs + E[Bt − Bs | Fs ] = Bs + 0 = Bs . ✓ (Used: Bt − Bs ⊥ Fs by independent increments;
E[Bt − Bs ] = 0 by stationary increments.)
Midsem Q2(b): Is Bt2 a Martingale? - Full Working
Compute E[Bt2 | Fs ] for s ≤ t. Write Bt = Bs + (Bt − Bs ) where (Bt − Bs ) ⊥ Fs .

E[Bt2 | Fs ] = E[(Bs + (Bt − Bs ))2 | Fs ]

= E[Bs2 | Fs ] + 2E[Bs (Bt − Bs ) | Fs ] + E[(Bt − Bs )2 | Fs ]
= Bs2 + 2Bs · E[Bt − Bs | Fs ] + E[(Bt − Bs )2 ]
= Bs2 + 2Bs · 0 + (t − s) = Bs2 + (t − s) ̸= Bs2

So Bt2 is NOT a martingale (it’s a submartingale: E[Bt2 | Fs ] = Bs2 + (t − s) ≥ Bs2 ). ✓

However, Mt = Bt2 − t IS a martingale:

E[Bt2 − t | Fs ] = Bs2 + (t − s) − t = Bs2 − s. ✓

15
Midsem Q2(a): Is Bt3 a Martingale?
Compute E[Bt3 | Fs ]. Use Bt = Bs + Z where Z = Bt − Bs ∼ N (0, t − s), Z ⊥ Fs .

E[Bt3 | Fs ] = E[(Bs + Z)3 | Fs ] = Bs3 + 3Bs2 · E[Z] + 3Bs · E[Z 2 ] + E[Z 3 ]

= Bs3 + 0 + 3Bs (t − s) + 0 = Bs3 + 3Bs (t − s) ̸= Bs3

So Bt3 is NOT a martingale. But Bt3 − 3tBt IS:

E[Bt3 − 3tBt | Fs ] = Bs3 + 3Bs (t − s) − 3t · E[Bt | Fs ] = Bs3 + 3Bs (t − s) − 3tBs = Bs3 − 3sBs

Midsem Q1(a): Is −Bt a Brownian Motion?

Check all four BM axioms for −Bt :

1. −B0 = −0 = 0 a.s. ✓

2. Increments of −B: −Bt − (−Bs ) = −(Bt − Bs ), independent across disjoint intervals

(same independence as B) ✓

3. −(Bt − Bs ) ∼ N (0, t − s) (negating N (0, σ 2 ) gives N (0, σ 2 )) ✓

4. Continuous paths: −Bt is continuous iff Bt is continuous ✓

CONCLUSION: −Bt IS a standard Brownian motion. ✓

Midsem Q1(b): Is cBt/c2 a Brownian Motion?
Check stationary increments: cBt/2 − cBs/2 = c(Bt/2 − Bs/2 ) ∼ N (0, c2 (t − s)/2). For this to
√
be N (0, t − s), need c2 /2 = 1, i.e., c = 2.
More generally, for cBt/n (scaling of space by c and time by 1/n): The variance of the increment
is c2 (t − s)/n. For this to equal t − s, need c2 = n. This is the self-similarity property of BM:
c · Bt/c2 is also a BM.
√
CONCLUSION: cBt/2 is a BM if and only if c = 2.
Non-Differentiability of BM - Mathematical Argument
BM paths are continuous (by definition) but nowhere differentiable with probability 1. Intuitive
argument:
√ √
• ∆B ∼ h (since Bt+h − Bt ∼ N (0, h), std dev = h)
√
• ∆B
∆t ≈ h
h
= √1
h
→ ∞ as h → 0

The “derivative” dB/dt would have to be infinite everywhere. This is why we cannot use
ordinary calculus for stochastic integrals – we need the Itô integral. Despite non-differentiability,
BM has finite quadratic variation:
X
[B, B]T = lim (Bti+1 − Bti )2 = T a.s.
∥Π∥→0

Heuristic rule (Itô calculus): (dBt )2 = dt. This is the KEY formula driving the Itô correction
term.
Lévy’s Characterization Theorem (Statement Only)
Lévy’s Theorem: A continuous local martingale Mt starting at M0 = 0 with quadratic variation
[M, M ]t = t for all t ≥ 0 a.s. is a standard Brownian motion. ⋆ Note: Used in the proof of
Girsanov’s Theorem: after showing B̃t is a Q-martingale with QV = t, Lévy’s theorem implies
B̃t is a Q-BM.

16
4.4 Itô Integral

Construction and Properties

Motivation: Why Ordinary
RT Calculus Fails
We want to define 0 ∆(t)dBt (integral against BM). Problems with classical Riemann-Stieltjes:

• BM has infinite total variation:

P
|Bti+1 − Bti | → ∞. Riemann-Stieltjes requires finite
variation in the integrator.

• BM is nowhere differentiable, so we cannot write dBt = Bt′ dt (there is no such B ′ ).

Solution: Define the Itô integral as an L2 -limit of Riemann sums using LEFT endpoints only
(non-anticipating).
Construction of the Itô Integral - Step by Step
Step 1: Simple (elementary) processes:
∆(t) is a step function: ∆(t) = ∆(tj ) for t ∈ [tj , tj+1 ), where ∆(tj ) is Ftj -measurable (knows
the past, not the future). For such ∆:
Z T X
∆(t)dBt := ∆(tj )(Btj+1 − Btj )
0

The LEFT endpoint evaluation is critical: ∆(tj ) is based on information BEFORE the BM
increment Btj+1 − Btj . This makes the integral adapted (non-anticipating).
Step 2: Extension by density: RT 2 2
The space of adapted processes with E[ 0 ∆(t) dt] R< ∞ is an L space. Simple processes
are dense in this space. Define ∆dB = limn→∞ ∆n dB in L2 (Ω), where ∆n are simple
R

approximations to ∆.
Itô Isometry (Proof
 )
RT 2  hR i
T 2 dt
Itô Isometry: E 0 ∆(t)dB t = E 0 ∆(t)
P
Proof for simple processes: Let I = j ∆(tj )∆Bj where ∆Bj = Btj+1 − Btj .
X 2  XX
2
E[I ] = E ∆(tj )∆Bj = E[∆(ti )∆Bi ∆(tj )∆Bj ]
i j

For i ̸= j (say i < j): ∆(ti ), ∆Bi , ∆(tj ) are all Ftj -measurable or earlier, while ∆Bj ∼
N (0, tj+1 − tj ) is independent of Ftj . Hence E[∆(ti )∆Bi ∆(tj )∆Bj ] = E[∆(ti )∆Bi ∆(tj )] ·
E[∆Bj ] = 0.
For i = j: E[∆(ti )2 ∆Bi2 ] = E[E[∆(ti )2 ∆Bi2 | Fti ]] = E[∆(ti )2 E[∆Bi2 | Fti ]] = E[∆(ti )2 (ti+1 −ti )]
(taking out ∆(ti ) which is Fti -measurable; E[∆Bi2 | Fti ] = E[∆Bi2 ] = ti+1 −ti by independence).
hP i hR i
T
Summing: E[I 2 ] = j E[∆(tj )2 (tj+1 − tj )] = E 2 (t 2 dt . ✓
P
j ∆(tj ) j+1 − t j ) → E 0 ∆(t)
Itô Integral as a Martingale
RT Rt
If ∆(t) is adapted and E[ 0 ∆(t)2 dt] < ∞, then It = 0 ∆(s)dBs is a martingale w.r.t. {Ft }.
In particular:

• E[It ] = 0 for all t (since I0 = 0 and martingales have constant expectation)

Rt
• E[It2 ] = E[ 0 ∆(s)2 ds] (by Itô isometry, setting T = t)

17
 Rt
• For s ≤ t : E[It | Fs ] = Is (Proof: E[ s ∆(u)dBu | Fs ] = 0 since it’s a zero-mean
Rt
martingale increment; then It = Is + s ∆dB, so E[It | Fs ] = Is .)

Linearity
RT of Itô IntegralR
T RT
0 (α∆(t) + βΓ(t))dBt = α 0 ∆(t)dBt + β 0 Γ(t)dBt
This follows from linearity of Riemann sums and the L2 limit construction.
Itô Integral as an
R t Adapted Stochastic Process
The process It = 0 ∆(s)dBs is an adapted process: for each t, It is Ft -measurable (it depends
only on the BM path up to t and ∆(s) for s ≤ t, all of which are Ft -measurable). Furthermore,
It has continuous sample paths (a deeper result).

4.5 Itô’s Formula (Itô-Doeblin Formula) - Key Theorem

Taylor Series Motivation and Heuristic Derivation

For a smooth function f (t, x), the two-variable Taylor series gives:
1 1
f (t + ∆t, x + ∆x) = f (t, x) + ft ∆t + fx ∆x + fxx (∆x)2 + ftx ∆x∆t + ftt (∆t)2 + . . .
2 2
For an Itô process dXt = µt dt + σt dBt , the increments are:

(dXt )2 = σt2 dt [since (dBt )2 = dt, (dBt dt) = 0, (dt)2 = 0]

The multiplication table:

dBt dBt = dt, dBt dt = 0, dt · dt = 0

Only (dBt )2 = dt is non-negligible. All higher-order terms vanish. This is the KEY insight:
unlike classical calculus (where (dx)2 is negligible), stochastic calculus has a quadratic term
that survives!
Itô’s Formula - Statement (Two Versions)

• Version 1 (f of BM only): Let Bt be a BM and f (t, x) ∈ C 1,2 (R+ × R). Then:

1
df (t, Bt ) = ft (t, Bt )dt + fx (t, Bt )dBt + fxx (t, Bt )dt
2
Integral form:
Z t Z t Z t
1
f (t, Bt ) = f (0, 0) + ft (s, Bs )ds + fx (s, Bs )dBs + fxx (s, Bs )ds
0 0 2 0

• Version 2 (f of Itô process): If dXt = µt dt + σt dBt and f (t, x) ∈ C 1,2 :

1
df (t, Xt ) = ft dt + fx dXt + fxx (dXt )2
 2 
1 2
= ft + µt fx + σt fxx dt + σt fx dBt
2

⋆ Note: The “Itô correction term” is 12 σt2 fxx dt. It is ABSENT in classical calculus. This term
makes e.g. GBM differ from classical exponential growth.
Itô’s Formula - Proof Sketch (Quiz 2 Q2 / Midsem)
Given in the exam as: state and prove Itô’s Lemma. Follow these steps:

18
 1. Partition [0, t] into n subintervals with mesh h = t/n. Let tk = kh, k = 0, 1, . . . , n.

2. Apply 2-variable Taylor expansion to f on [tk , tk+1 ]:

1
f (tk+1 , Btk+1 ) − f (tk , Btk ) = ft ∆t + fx ∆Bk + fxx (∆Bk )2 + ftx ∆t∆Bk + . . .
2

3. Sum over k from 0 to n − 1. Identify each term in the limit:

•
P R
ft ∆t → ft dt (Riemann integral)
•
P R
fx ∆Bk → fx dBt (Itô integral by construction)
• 21 fxx (∆B 2 1 2
P R
Pk ) → 2 2fxx ds: R Since (∆Bk ) R ≈ tk+1 − tk = h by the LLN for QV.
Formally: fxx (∆Bk ) → fxx d[B, B]s = fxx ds.
• √
p
Terms ∆t∆Bk : |∆t∆B√ k | ≤ ∆t· 2h log(1/h) → 0 in L2 (BM increments are of order
h, so ∆t∆Bk ∼ h h → 0).
• (∆t)2 terms: O(h2 ) with n = t/h terms → O(h) → 0.

4. Taking the limit n → ∞ gives Itô’s formula. ✓

(∆Bk )2 → t (QV = t). This uses the L2 convergence

P
⋆ Note: The CRITICAL step is showing
of the quadratic variation of BM.
Itô Product Rule (Integration by Parts)
For two Itô processes Xt and Yt :

d(Xt Yt ) = Xt dYt + Yt dXt + dXt dYt

The extra term dXt dYt (quadratic covariation) is the Itô correction. Using the multiplication
table: if dXt = µX dt + σX dBt and dYt = µY dt + σY dBt , then dXt dYt = σX σY dt. Example:
d(Bt St ) = Bt dSt + St dBt + dBt dSt .
Applications of Itô’s RFormula - Worked Examples
t
Example 1: Compute 0 Bs dBs . Apply Itô to f (x) = x2 /2:
1
d(Bt2 /2) = Bt dBt + dt
2
Z t
⇒ Bt2 /2 = Bs dBs + t/2
0
Z t
⇒ Bs dBs = (Bt2 − t)/2
0
Rt
Compare to classical calculus: 0 xdx = t2 /2 ̸= (Bt2 − t)/2. The Itô result has the correction
term −t/2.
Example 2: Geometric Brownian Motion. Apply Itô to f (t, x) = ln(x):
1 1
d(ln St ) = dSt − (dSt )2
St 2St2
If dSt = µSt dt + σSt dBt :
1 1 2 2
d(ln St ) = (µSt dt + σSt dBt ) − σ St dt = (µ − σ 2 /2)dt + σdBt
St 2St2

⇒ ln(St /S0 ) = (µ − σ 2 /2)t + σBt

⇒ St = S0 exp (µ − σ 2 /2)t + σBt


[Geometric Brownian Motion]

19
4.6 Stochastic Differential Equations (SDEs)

Basic Formulation and Interpretation

An SDE is an equation of the form:

dXt = µ(t, Xt )dt + σ(t, Xt )dBt , X0 = x0

Interpretation: Over a small interval [t, t + dt]:

Xt+dt ≈ Xt + µ(t, Xt )dt + σ(t, Xt )(Bt+dt − Bt )

The drift µ(t, Xt )dt is the “deterministic” part of the change. The diffusion σ(t, Xt )dBt is the
“random” part. σ controls the volatility.
Geometric Brownian Motion (GBM) - The Stock Price Model
The Black-Scholes stock price model:

dSt = µSt dt + σSt dBt (µ = drift, σ = volatility, both constant)

Solution via Itô (apply Itô to ln St as above):

St = S0 exp (µ − σ 2 /2)t + σBt

Properties: St > 0 always (cannot go negative). Log returns ln(St /Ss ) ∼ N ((µ − σ 2 /2)(t −
s), σ 2 (t − s)). This is the log-normal distribution model. Under the real-world measure P:
E[St ] = S0 eµt (exponential growth at rate µ). Under the risk-neutral measure Q (after Girsanov
with θ = (µ − r)/σ): EQ [St ] = S0 ert (growth at risk-free rate r).
Quiz 2 Q3: NPS Liquidity Discount - SDE Problem
Setup: Ft = fundamental price (GBM), Lt = liquidity discount, St = Ft − Lt = observed price.
Given: dFt = µFt dt + σFt dBt (GBM); dLt = −λLt dt (decays exponentially); L0 = l > 0.

• Part (a): Solve dLt = −λLt dt (ODE, deterministic):

dL
= −λdt ⇒ ln(L/L0 ) = −λt ⇒ Lt = l · e−λt
L

• Part (b): Solve for St = Ft − Lt :

Ft = F0 exp (µ − σ 2 /2)t + σBt

where F0 = S0 + l

St = Ft − l · e−λt
To find dSt : dSt = dFt − dLt = (µFt dt + σFt dBt ) − (−λLt dt) = (µFt + λLt )dt + σFt dBt .
Substituting Ft = St + Lt : dSt = (µ(St + Lt ) + λLt )dt + σ(St + Lt )dBt . This SDE has
explicit Lt = le−λt substituted in:

dSt = (µSt + (µ + λ)le−λt )dt + σ(St + le−λt )dBt

4.7 Evolution of Portfolio & Black-Scholes-Merton

Discounted Stock Price

Risk-free rate r. Define discounted stock price S̃t = e−rt St . By Itô applied to e−rt St :

dS̃t = −re−rt St dt + e−rt dSt = e−rt (dSt − rSt dt)

= e−rt ((µ − r)St dt + σSt dBt ) = S̃t (µ − r)dt + σ S̃t dBt

20
Under Q (Girsanov with θ = (µ − r)/σ, so dB̃t = dBt + θdt is Q-BM):
dS̃t = S̃t σdB̃t (pure martingale under Q − no drift term)

Portfolio Value and SDE

Portfolio: ∆t shares of stock, remainder in risk-free bond. Total value Xt .
dXt = ∆t dSt + r(Xt − ∆t St )dt = rXt dt + ∆t (dSt − rSt dt)
= rXt dt + ∆t St (µ − r)dt + ∆t σSt dBt
Three components: rXt dt = risk-free return on entire portfolio; ∆t St (µ − r)dt = excess return
(equity risk premium × stock holding); ∆t σSt dBt = stochastic volatility term.
Discounted Portfolio
X̃t = e−rt Xt
dX̃t = ∆t dS̃t = ∆t S̃t (µ − r)dt + ∆t σ S̃t dBt
Under Q: dX̃t = ∆t σ S̃t dB̃t (pure martingale - no drift). This means the discounted portfolio
value is a Q-martingale, which is equivalent to no-arbitrage.
Stochastic Hedging Strategy - Key Idea
The key idea of BSM: form a self-financing hedging portfolio where ∆t = ∂V ∂S (number of shares
held). By choosing ∆t correctly, the portfolio value tracks the option price exactly. The residual
risk (exposure to dBt ) is cancelled. Remaining portfolio risk equals the risk-free rate r.
Black-Scholes-Merton PDE - Derivation
Consider a European option V (t, St ) with payoff h(ST ) at time T .

• Step 1: Apply Itô to V (t, St ):

1
dV = (Vt + µSVs + σ 2 S 2 Vss )dt + σSVs dBt
2
• Step 2: Form delta-hedged portfolio Π = V − ∆S (short option, long ∆ shares):
1
dΠ = dV − ∆dS = (Vt + µSVs + σ 2 S 2 Vss − ∆µS)dt + σS(Vs − ∆)dBt
2
• Step 3: Choose ∆ = Vs to eliminate the dBt term (delta-hedge):
 
1 2 2
dΠ = Vt + σ S Vss dt
2
• Step 4: No-arbitrage requires dΠ = rΠdt = r(V − Vs S)dt:
1
Vt + σ 2 S 2 Vss = rV − rSVs
2
BSM PDE: Vt + rS · Vs + 12 σ 2 S 2 · Vss − rV = 0
With terminal condition V (T, S) = h(S) = (S − K)+ for a European call.
BSM Formula for European Call
C = S0 N (d1 ) − Ke−rT N (d2 )
ln(S0 /K) + (r + σ 2 /2)T
d1 = √
σ T
√ ln(S0 /K) + (r − σ 2 /2)T
d2 = d1 − σ T = √
σ T
Where N (·) is the standard normal CDF. Interpretation: S0 N (d1 ) = expected payoff from
receiving stock (probability d1 of exercising × stock value); Ke−rT N (d2 ) = expected cost (PV
of strike × probability d2 of exercise under Q).

21
MODULE 5: RISK-NEUTRAL PRICING

5.1 Change of Measure - Concept

Two probability measures P and Q on (Ω, F) are equivalent (P ∼ Q) if they have the same
null sets: P(A) = 0 ⇔ Q(A) = 0. Equivalent measures agree on what is “impossible” but can
disagree on probabilities of non-null events.
Why change measure in finance? Under the real-world measure P, the stock drift is µ (market
growth rate). Under the risk-neutral measure Q, the drift becomes r (risk-free rate). This
switch makes the discounted stock price a martingale under Q, which makes derivative pricing
tractable via EQ [discounted payoff].
Intuition: Changing measure “adjusts probabilities” to account for risk aversion. Risk-neutral
investors price assets using expected returns equal to r regardless of actual drift µ. The change
of measure converts the “real world” to a “risk-neutral world.”

5.2 Radon-Nikodym Derivative - Discrete

Definition (Finite/Discrete Spaces)

Radon-Nikodym Derivative (Discrete): If P and Q are equivalent probability measures on a finite
dQ P
Ω,
P define Z = dP by Z(ω) = Q({ω})/P({ω}) for each ω ∈ Ω. Then Q(A) = ω∈A Q({ω}) =
ω∈A Z(ω)P({ω}) = E [Z · 1A ].
P

Properties (Discrete)

• Z = dQ/dP ≥ 0 a.s. under P (with Z > 0 a.s. since P ∼ Q)

• EP [Z] = ω∈Ω Z(ω)P({ω}) = ω∈Ω Q({ω}) = Q(Ω) = 1

P P

• Change of expectation: EQ [X] = EP [ZX] for any random variable X.

• Bayes formula: EQ [X | Ft ] = EP [ZX | Ft ]/EP [Z | Ft ]. Used in the proof of Girsanov.

Radon-Nikodym Derivative as a Martingale (Discrete)

Define the process Zn = EP [Z | Fn ] where Z = dQ/dP. Then:

• (Zn , Fn ) is a P-martingale (by Tower Property: EP [Zn+1 | Fn ] = EP [EP [Z | Fn+1 ] | Fn ] =

EP [Z | Fn ] = Zn )

• Z0 = EP [Z] = 1

• For finite horizon T : ZT = Z = dQ/dP a.s.

Qn  q I{ξi =+1}  1−q I{ξi =−1}

In the binomial model: Zn = i=1 p 1−p = product of likelihood ratios at
each step.

5.3 Radon-Nikodym Derivative - Continuous

Radon-Nikodym Theorem (Statement)

Radon-Nikodym Theorem: If Q is absolutely continuous w.r.t. P on (Ω, F) - written Q ≪ P,
meaning P(A) = 0 ⇒ Q(A) = 0 - then there exists a unique (up to P-a.s. equality) non-
negative F-measurable RV Z = dQ/dP (the Radon-Nikodym derivative or density) such that

22
 R
Q(A) = A ZdP for all A ∈ F. If P ∼ Q (both absolutely continuous w.r.t. each other), then
Z > 0 P-a.s. and dP/dQ = 1/Z.
Change of Measure for a Gaussian Random Variable
Setup: X ∼ N (µ, σ 2 ) under P. We want Q such that X ∼ N (0, σ 2 ) under Q (shift mean to 0).
The RN derivative is:
µ2
 
dQ µX
= exp − 2 + 2
dP σ 2σ
h  2
i
Verification: EP [ dQ P exp − µX + µ . Writing X = µ + σZ with Z ∼ N (0, 1) ⇒ X ∼
dP ] = E σ2 2σ 2
N (µ, σ 2 ) under P:
 2      2   2   
µ P µ(µ + σZ) µ µ µZ
exp E exp − = exp · exp − 2 E exp −
2σ 2 σ2 2σ 2 σ σ

µ2
   2 
µ
= exp − 2 · exp =1 ✓
2σ 2σ 2

More generally
 (Cameron-Martin):
 If X ∼ N (θ, σ 2 ) under P and Y ∼ N (0, σ 2 ) under Q, then
θ2
dQ/dP = exp − θXσ2
+ 2σ 2 . As n → ∞ with θn → 0, dQ/dP → 1 (measures merge – see Quiz

3 Q4).
Characterization of RN Derivative on Continuous Spaces as an Adapted Process
In continuous time (0 ≤ t ≤ T ), define Zt = EP [ZT | Ft ] where ZT = dQ/dP (the terminal RN
derivative). Then:

• Zt is a P-martingale (by Tower Property)

• For constant market price of risk θ: Zt = exp(−θBt − θ2 t/2) (the stochastic exponential)

• dZt = −θZt dBt (Itô formula: dZt = Zt (−θ2 /2 + 21 θ2 )dt + Zt (−θ)dBt = −θZt dBt ). No dt
term → Zt is a (local) martingale.
RT
Is Zt a TRUE martingale? Yes, under Novikov’s condition: EP [exp( 12 0 θt2 dt)] < ∞. For
constant θ: exp(θ2 T /2) < ∞✓

5.4 Risk-Neutral Probability - Summary

Risk-Neutral Measure Q: An equivalent probability measure Q ∼ P such that the discounted

price process S̃t = e−rt St is a Q-martingale.
Pricing formula: V0 = e−rT EQ [payoff at T ].
Key formula connecting P and Q: EQ [X] = EP [ZT X] where ZT = dQ/dP.

5.5 Girsanov’s Theorem - Full Statement and Proof

Setup and Statement

Girsanov’s Theorem: Let Bt be a P-Brownian motion on (Ω, F, P) with filtration {Ft }. Let θt
RT
be an adapted process satisfying Novikov’s condition: EP [exp( 12 0 θt2 dt)] < ∞. Define:
 Z T Z T 
1
ZT = exp − θt dBt − θt2 dt [the stochastic exponential]
0 2 0

23
Set Q by dQ = ZT dP (i.e., ZT = dQ/dP). Then the process:
Z t
B̃t = Bt + θs ds
0

is a standard Brownian motion under Q. Equivalently, dBt = dB̃t − θt dt, so replacing dBt
everywhere with dB̃t − θt dt converts P-dynamics to Q-dynamics.
Proof of Girsanov’s Theorem
Strategy: Use Lévy’s Theorem – show B̃t is a continuous Q-local martingale starting at 0 with
QV = t.
Rt
1. Continuity: B̃t = Bt + 0 θs ds is continuous (sum of continuous processes). ✓

2. Initial value: B̃0 = B0 + 0 = 0. ✓

3. Q-martingale property: Must show EQ [B̃t | Fs ] = B̃s for s ≤ t. Using Bayes formula
(EQ [X | Fs ] = EP [ZT X | Fs ]/EP [ZT | Fs ] = EP [Zt X | Fs ]/Zs ), and the fact that Zt B̃t
is a P-martingale (can be verified via Itô product rule: d(Zt B̃t ) = B̃t dZt + Zt dB̃t + dZt ·
dB̃t = B̃t (−θZt dBt ) + Zt (dBt + θdt) + (−θZt dt) = Zt dBt (pure martingale)). Hence
EQ [B̃t | Fs ] = EP [Zt B̃t | Fs ]/Zs = Zs B̃s /Zs = B̃s . ✓
R R R
4. Quadratic variation: [B̃, B̃]t = [B + θds, B + θds]t . The term θds has bounded
variation (smooth), so it contributes zero to QV. Thus [B̃, B̃]t = [B, B]t = t. ✓

5. By Lévy’s Theorem: B̃t is a continuous Q-martingale with B̃0 = 0 and [B̃, B̃]t = t ⇒ B̃t
is a Q-Brownian motion. ✓

Application of Girsanov’s Theorem to Black-Scholes

Under P: dSt = µSt dt + σSt dBt . Market price of risk θ = (µ − r)/σ.
Define Q via Girsanov: ZT = exp(−θBT − θ2 T /2); dQ = ZT dP.
Then B̃t = Bt + θt is a Q-BM. Rewrite the SDE using dBt = dB̃t − θdt:

dSt = µSt dt + σSt (dB̃t − θdt) = µSt dt − σSt θdt + σSt dB̃t
 
µ−r
= µ−σ St dt + σSt dB̃t = rSt dt + σSt dB̃t
σ

Under Q: dSt = rSt dt + σSt dB̃t . The drift becomes r (risk-free rate). Hence S̃t = e−rt St is a
Q-martingale. ✓
Girsanov and Existence/Uniqueness of Risk-Neutral Measure

• Complete markets (unique Q): When there is ONE source of randomness (one BM B)
and assets span all risks, θ = (µ − r)/σ is uniquely determined ⇒ unique Q ⇒ complete
market.

• Incomplete markets (many Q): When there are more sources of uncertainty than traded
assets. For each choice of θ satisfying the martingale condition (any θ component in the
“unspanned” subspace is free to choose), we get a different Q. Infinitely many EMMs ⇒
incomplete market.

DiceChain (Quiz 1 Q3): 6 states, 1 risky asset ⇒ 4 degrees of freedom in choosing Q ⇒

incomplete market ⇒ infinitely many Q. (Complete market needs as many assets as states.)

24
5.6 Martingale Representation Theorem (MRT) - Statement Only

MRT: Let Bt be a BM on (Ω, F , P) and {Ft } = {FtB } (the natural filtration of B). If Mt is
any (Ft )-martingale with E[Mt2 ] < ∞, then there exists a unique adapted process φt satisfying
RT
E[ 0 φ2s dt] < ∞ such that:
Z t
Mt = M0 + φs dBs
0
In other words, every square-integrable martingale in the filtration of a BM is an Itô integral
(stochastic integral). There are no “extra” martingales.
MRT and Completeness / Uniqueness of Q
The MRT establishes market completeness in the BSM model:

1. Any contingent claim with square-integrable payoff h(ST ) defines a Q-martingale Mt =

EQ [e−rT h(ST ) | Ft ].
Rt
2. By MRT (under Q, B̃t is a Q-BM): Mt = M0 + 0 φs dB̃s for some φs .

3. Setting ∆t = φt /(σ S̃t ) gives a replicating portfolio: dX̃t = ∆t σ S̃t dB̃t = φt dB̃t = dMt .

4. Since X̃0 = M0 = e−rT EQ [h(ST )], we have X̃t = Mt for all t ⇒ Xt = Vt for all t.

Conclusion: EVERY derivative can be replicated ⇒ market is complete ⇒ Q is unique (SFTAP).

5.7 First & Second Fundamental Theorems of Asset Pricing

First Fundamental Theorem (Continuous Time)

FFTAP: The financial market is arbitrage-free if and only if there exists at least one equivalent
martingale measure (EMM) Q ∼ P under which all discounted asset prices are martingales.
Formal arbitrage in continuous time: A self-financing trading strategy θ is an arbitrage if
Xt (θ) ≥ 0 for all t ∈ [0, T ] a.s., P(XT (θ) > 0) > 0, and X0 (θ) = 0.
Second Fundamental Theorem
SFTAP: An arbitrage-free market is complete (every attainable contingent claim can be repli-
cated by a self-financing trading strategy) if and only if the equivalent martingale measure Q
is unique.
Summary:

• No EMM exists ⇒ Arbitrage opportunity exists

• Exactly one EMM (Q unique) ⇒ Market is arbitrage-free AND complete

• Multiple EMMs (Q not unique) ⇒ Market is arbitrage-free but incomplete

⋆ Note: Quiz 1 Q3 (DiceChain): 6 outcomes, 1 asset ⇒ 4 degrees of freedom in Q → infinitely

many EMMs → incomplete market. Quiz 3 Q3 (Continuous space RN derivative): Shows how
to compute dQ/dP explicitly for Gaussian shifts.
Quiz 3 Q4: Radon-Nikodym Derivative for Continuous Spaces - Worked Solution
Setup: X ∼ N (0, 1) under P. Define Yn = X + θn , Zn = X/θn where 0 < θ < 1.

• Part (a): Find RN derivative Vn = dPZn /dPYn . Under P: Yn = X + θn ∼ N (θn , 1)

and Zn = X/θn ∼ N (0, 1/θ2n ). These are different distributions – we need the RN

25
 derivative of Zn w.r.t. Yn . Equivalently, find dQ/dP where P corresponds to observations
from the Yn distribution and Q corresponds to observations from Zn distribution. By the
Cameron-Martin formula for Gaussian shifts:
θ2n
 
dPZn n
Vn = = exp −θ · X −
dPYn 2

• Part (b): As n → ∞ with 0 < θ < 1, θn → 0 so Vn → exp(0) = 1. This means the two
distributions PZn and PYn become asymptotically identical as n → ∞ - the shift θn → 0
makes them indistinguishable.

MODULE 6: FEYNMAN-KAC THEOREM

6.1 Feynman-Kac Theorem

Bridge Between SDEs and PDEs

The Feynman-Kac theorem is a deep result connecting two worlds:

• Probabilistic: Expected value of a functional of a diffusion process

• Analytical: Solution to a parabolic (heat-type) PDE

In finance, this bridge is crucial: the BSM PDE (analytic) and risk-neutral pricing formula
(probabilistic) are two sides of the same coin, connected by Feynman-Kac. Solving the PDE
gives the derivative price; alternatively, computing the Q-expectation gives the same answer.
Statement of Feynman-Kac Theorem
Feynman-Kac: Let Xt satisfy the SDE: dXt = µ(t, Xt )dt + σ(t, Xt )dBt . Define the function:
  Z T  
u(t, x) = E h(XT ) · exp − c(s, Xs )ds | Xt = x
t

where h is the terminal payoff function and c(s, Xs ) ≥ 0 is a discount/killing rate. Then u
satisfies the PDE:
1
ut + µ(t, x)ux + σ 2 (t, x)uxx − c(t, x)u = 0
2
with terminal condition u(T, x) = h(x).
⋆ Note: The BSM PDE arises by setting µ(t, x) = rx, σ(t, x) = σx, c(t, x) = r in Feynman-Kac.
Discounted Feynman-Kac
Special case with c(t, x) = r (constant discount rate = risk-free rate):
h i
u(t, x) = EQ e−r(T −t) h(XT ) | Xt = x

satisfies: ut +µux + 12 σ 2 uxx −ru = 0. This is exactly the no-arbitrage pricing formula: V (t, St ) =
e−r(T −t) EQ [h(ST ) | St = S].
Deriving BSM PDE from Feynman-Kac (Quiz 3 Q2)

• Step 1: Stock under Q: dSt = rSt dt + σSt dBt (GBM with drift r, volatility σ).

• Step 2: Option price V (t, S) = e−r(T −t) EQ [h(ST ) | St = S]. Define u(t, x) = er(T −t) V (t, x)
(undiscounted):
u(t, x) = EQ [h(ST ) | St = x]

26
 • Step 3: Apply Feynman-Kac with µ(t, x) = rx, σ(t, x) = σx, c = 0:
1
ut + rx · ux + σ 2 x2 · uxx = 0
2

• Step 4: Convert back to V = e−r(T −t) u. Using Vt = −r · e−r(T −t) u + e−r(T −t) ut etc., the
PDE becomes:
1
Vt + rSVS + σ 2 S 2 VSS − rV = 0 ← BSM PDE ✓
2

Alternatively, apply Feynman-Kac DIRECTLY to V (t, S) = e−r(T −t) EQ [h(ST ) | St = S] with

c = r:
1
Vt + rS · VS + σ 2 S 2 · VSS − r · V = 0
2
Proof Strategy forR t Feynman-Kac (Sketch)
Define Mt = exp(− 0 c(s, Xs )ds) · u(t, Xt ) where u solves the FK PDE. Apply Itô’s formula to
Mt :
 Z  
1 2
dMt = exp − c ut dt + ux dXt + uxx (dXt ) − c · udt
2
 Z    Z 
1
= exp − c ut + µux + σ 2 uxx − cu dt + exp − c · σux dBt
2

If u satisfies the FK PDE, the dt term vanishes ⇒ Mt is a local martingale. Under regularity,
it is a true martingale. Therefore:
h RT i Rt
E[MT | Ft ] = Mt ⇒ E e− 0 cds h(XT ) | Ft = u(t, Xt ) · e− 0 cds ✓

EXAM PREP: KEY PROOFS, WORKED SOLUTIONS & FOR-

MULA SHEETS

Complete Proof Bank - Must-Know for Endsem

Proof 1: Xn2 − n is a Martingale
2
E[Xn+1 − (n + 1) | Fn ] = E[(Xn + ξn+1 )2 | Fn ] − (n + 1) = Xn2 + 2Xn · 0 + 1 − n − 1 = Xn2 − n. ✓
Proof 2: exp(θBt − θ2 t/2) is a Q-Martingale (Stochastic Exponential)
Let Zt = exp(θBt − θ2 t/2). Apply Itô to f (t, x) = exp(θx − θ2 t/2):

ft = −(θ2 /2)Zt , fx = θZt , fxx = θ2 Zt

 2 
θ 1 2
dZt = − + θ Zt dt + θZt dBt = θZt dBt
2 2
Since dZt = θZt dBt (no dt term), Zt is a stochastic integral (Itô integral) → local martingale.
Under Novikov’s condition (here: E[exp(θ2 T /2)] < ∞), it is a true martingale. ✓
Proof 3: Jensen’s Inequality
For convex φ, there exists c (supporting slope at µ = E[X]) with φ(x) ≥ φ(µ) + c(x − µ) ∀x.
Take expectations: E[φ(X)] ≥ φ(µ) + c · E[X − µ] = φ(E[X]) + 0. ✓
Proof 4: Convex Function of Martingale is Submartingale (Midsem Q1c)
If (Xn ) is a martingale, φ convex: E[φ(Xn+1 ) | Fn ] ≥ φ(E[Xn+1 | Fn ]) = φ(Xn ). (Conditional
Jensen + martingale property.)

27
Proof 5: Itô Isometry
See Module 4.4 for complete proof.
Key: cross-terms vanish by independence of non-overlapping BM increments; diagonal terms
= E[∆Bi2 ](∆t).
Proof 6: Risk-Neutral Probability in 1-Period Binomial
S0 = EQ [S1 /(1 + r)] ⇒ S0 (1 + r) = quS0 + (1 − q)dS0 ⇒ 1 + r = qu + (1 − q)d ⇒ q =
(1 + r − d)/(u − d). ✓
Proof 7: Zn = Xn3 − 3nXn is a Martingale
Full proof in Module 4.1, Quiz 1 Q1 section.

Proof 8: Sn = eσX n (2/(eσ + e−σ ))n is a Martingale (Midsem Q2c)

   σ −σ
2 2
E[Sn+1 | Fn ] = Sn · eσ +e −σ · E[eσξn+1 ] = Sn · eσ +e −σ · e +e
2 = Sn . ✓
eσ +e−σ
(Used: E[eσξn+1 ] = 12 eσ + 21 e−σ = 2 = cosh(σ), the moment generating function of ξn+1 .)
Proof 9: SRW is an Adapted Stochastic Process
Xn = ξ1 + · · · + ξn is a function of (ξ1 , . . . , ξn ), hence σ(ξ1 , . . . , ξn ) = Fn -measurable. ✓
Proof 10: −Bt is a Brownian Motion
Full proof in Module 4.3.
All four BM axioms satisfied. ✓
Proof 11: Girsanov’s Theorem
Full proof in Module 5.5.
Strategy: Lévy’s theorem. Show B̃t is a Q-continuous martingale starting at 0 with QV = t. ✓
Proof 12: BSM PDE Derivation
Full proof in Module 4.7.
Strategy: Itô formula on V (t, S) + delta-hedging + no-arbitrage. ✓

FULL WORKED SOLUTIONS - All Assessments

Quiz 1 Q2: Expected Tosses for HTHTHT Pattern

Method: Markov chain on states 0-5 (length of matched prefix). E6 = 0 (done). Key recursions
(with KMP-style failure function for restarts):

E0 = 2 + E1
E1 = 2 + E2
1 1
E2 = 1 + E3 + E0 (T → stay 0, H → state 3)
2 2
1 1
E3 = 1 + E4 + E1 (T → state 4, H → state 1)
2 2
1 1
E4 = 1 + E5 + E0 (H → state 5, T → state 0)
2 2
1 1
E5 = 1 + · 0 + E1 = 1 + E1 /2 (T → done, H → state 1)
2 2
Solution: From E5 = 1 + E1 /2; E4 = 1 + 12 (1 + E1 /2) + 21 E0 = 3/2 + E1 /4 + 12 E0 ; substitute
up. Answer: E0 = 84.
⋆ Note: Why not restart at 0 on failure? KMP insight: HTHTHT has overlapping prefix-suffix.
When you fail at position k, you jump to the longest suffix of the failed string that matches a
prefix of HTHTHT. E.g., state 3 (matched HTH) on seeing H: string becomes HTHH - longest
prefix of HTHTHT matching a suffix is H (length 1) → go to state 1, not state 0.

28
Quiz 1 Q3 (DiceChain) - All Parts
1 P
(a) S0 = 1+r qi ui S0 follows directly from EQ [S̃P
1 ] = S̃0 , where
P S̃1 = S1 /(1 + r).
(b) 6 unknowns (q1 , . . . , q6 ), only 2 constraints ( qi = 1; qi ui = 1 + r) → 4 degrees of free-
dom ⇒ infinitely many Q.
(c) Set q3 = q4 = 1/2, rest = 0. Check: 21 (1.00) + 12 (1.10) = 1.05 = 1 + r. ✓ Q =
(0, 0, 1/2, 1/2, 0, 0).
Midsem Q1: BM and Martingale Questions
(a) −Bt is a BM: [see Proof
√ 10 above]
(b) cBt/c2 is a BM iff c = 2: Variance of increment = c2 (t − s)/2. Need = t − s ⇒ c2 = 2 ⇒
√
c = 2.
(c) φ(Xt ) is a submartingale if φ convex: [see Proof 4 above]
Midsem Q2: BM Martingale Checks and Itô’s Lemma
(a) Bt2 is NOT a martingale (E[Bt2 | Fs ] = Bs2 + (t − s) ̸= Bs2 ). Bt2 − t IS a martingale. [See
Module 4.3]
(b) Bt3 is NOT a martingale. Bt3 − 3tBt IS a martingale. [See Module 4.3]
(c) Sn = eσXn (2/(eσ + e−σ ))n IS a martingale. [See Proof 8 above]
State and Prove Itô’s Lemma - see Module 4.5 for complete statement and proof

Midsem Q3: NPS Liquidity Discount

Full solution in Module 4.6, Quiz 2 Q3 section

Key formulas: Lt = le−λt , Ft = F0 exp((µ − σ 2 /2)t + σBt ), St = Ft − Lt : dSt = (µFt +

λLt )dt + σFt dBt .
Midsem Q4: Card Game Filtrations - Complete Answer
Time indices: t = 0 (before cards), t = 1 (after dealing), t = 2 (after P1 acts), t = 3 (after P2
acts), t = 4 (after P3 acts).

• Public filtration (a): F0 = {∅, Ω}, F1 = {∅, Ω}, F2 = σ(A1 ), F3 = σ(A1 , A2 ), F4 =

σ(A1 , A2 , A3 ).
Why F1 = F0 ? Cards are dealt PRIVATELY. No public information is revealed at t = 1.
The public only sees actions.
(1) (1) (1) (1)
• Personal filtration of P1 (b): G0 = {∅, Ω}, G1 = σ(C1 ), G2 = σ(C1 , A1 ), G3 =
(1)
σ(C1 , A1 , A2 ), G4 = σ(C1 , A1 , A2 , A3 ).
(2) (2) (2) (2)
• Personal filtration of P2 (c): G0 = {∅, Ω}, G1 = σ(C2 ), G2 = σ(C2 , A1 ), G3 =
(2)
σ(C2 , A1 , A2 ), G4 = σ(C2 , A1 , A2 , A3 ).
(3) (3) (3) (3)
• Personal filtration of P3 (d): G0 = {∅, Ω}, G1 = σ(C3 ), G2 = σ(C3 , A1 ), G3 =
(3)
σ(C3 , A1 , A2 ), G4 = σ(C3 , A1 , A2 , A3 ).

• Part (e): Prove each collection is a σ-algebra on Ω. For any of these (e.g., F2 = σ(A1 )):

1. Ω ∈ F2 : “A1 takes any value” is certain, so Ω ∈ F2 . ✓

2. Closed under complement: If {A1 ∈ B} ∈ F2 for Borel B, then {A1 ∈
/ B} = {A1 ∈
B c } ∈ F2 . ✓
3. Closed under countable unions: σ(A1 ) is defined as a σ-algebra (by construction as
the smallest σ-algebra making A1 measurable), so closed under countable unions by
definition. ✓

29
Midsem Q5: Abnormal Returns Event Study - Complete Calculations
Market model: Ri,t = αi + βi Rm,t + ϵi,t . Parameters from Table 1.
Part (b) Expected returns on Day 0 (Rm,0 = 0.60%):

• NVDA: 0.02 + 1.35(0.60) = 0.02 + 0.81 = 0.83%

• AVGO: 0.01 + 1.20(0.60) = 0.01 + 0.72 = 0.73%

• DAL: 0.03 + 1.10(0.60) = 0.03 + 0.66 = 0.69%

• XOM: 0.01 + 0.85(0.60) = 0.01 + 0.51 = 0.52%

Part (c) Abnormal returns: ARi,t = Ri,t − E[Ri,t ]:

• Day 0: NVDA: 3.80 − 0.83 = 2.97%. AVGO: 5.10 − 0.73 = 4.37%. DAL: −1.20 − 0.69 =
−1.89%. XOM: 0.90 − 0.52 = 0.38%.

• Day +1 (Rm,+1 = −0.20%): Expected: NVDA: 0.02 + 1.35(−0.20) = −0.25%. AVGO:

−0.23%. DAL: −0.19%. XOM: −0.16%.
NVDA: 0.40−(−0.25) = 0.65%. AVGO: 1.00−(−0.23) = 1.23%. DAL: −0.80−(−0.19) =
−0.61%. XOM: 0.70 − (−0.16) = 0.86%.

Part (d) CAR(0, +1) = AR0 + AR+1 :

• NVDA: 2.97 + 0.65 = 3.62%

• AVGO: 4.37 + 1.23 = 5.60%

• DAL: −1.89 + (−0.61) = −2.50%

• XOM: 0.38 + 0.86 = 1.24%

Part (e) Volume Surprise V Si = ln(Avg event volume/Median pre-event volume):

• NVDA: ln(82/48) = ln(1.708) = 0.536

• AVGO: ln(12/6) = ln(2) = 0.693

• DAL: ln(15/11) = ln(1.364) = 0.310

• XOM: ln(24/18) = ln(1.333) = 0.288

Part (f ) Analysis: AVGO has strongest positive reaction (CAR = 5.60%, high VS, AI exposure
0.70 - Broadcom is a key AI chip supplier). NVDA also very strong (CAR = 3.62%, AI exposure
= 0.85). DAL has most adverse reaction (CAR = −2.50% AI investment does not directly help
airlines, but rising electricity/energy costs could increase operational costs). The headline helps
tech/infrastructure but hurts sectors dependent on energy.
Midsem Q6: Axiomatic Probability Full Answer
Full definitions in Module 2.2. For the diagram question, describe each component and explain
why axiomatic construction is needed – see Module 2.2, “Why Do We Need an Axiomatic
Construction?” section.

30
QUICK REFERENCE - FORMULAS AND CHECKLISTS

Martingale Verification Checklist

1. Adapted: Is Xn Fn -measurable? (Can we determine it from information up to time t?)

2. Integrable: Is E[|Xn |] < ∞?

3. Martingale condition: Compute E[Xn+1 | Fn ]. Does this equal Xn ?

Standard techniques: (1) Write Xn+1 = Xn + (something with ξn+1 ). (2) Use E[ξn+1 | Fn ] = 0,
2
E[ξn+1 3
| Fn ] = 1, E[ξn+1 | Fn ] = 0. (3) Check if dt term vanishes in dXt (continuous case).
Itô Formula Application Checklist

1. Identify f (t, x) and the process Xt (what SDE does Xt satisfy?)

2. Compute partials: ft , fx , fxx (and fxxx if needed)

3. Write df = ft dt + fx dXt + 21 fxx (dXt )2

4. Substitute (dXt )2 = σt2 dt; (dBt )2 = dt; dBt dt = 0; (dt)2 = 0

5. Collect dt terms (drift) and dBt terms (diffusion)

6. Is the resulting expression a martingale? (No dt term = local martingale)

Risk-Neutral Pricing Checklist

1. Find market price of risk θ = (µ − r)/σ

2. Apply Girsanov: dB̃t = dBt + θdt is Q-BM; dQ/dP = exp(−θBT − θ2 T /2)

3. Rewrite SDE under Q using dBt = dB̃t − θdt (drift becomes r for GBM)

4. Price = e−rT EQ [payoff at T | Ft ]

5. For PDE: apply Feynman-Kac. BSM PDE: Vt + rSVS + 21 σ 2 S 2 VSS − rV = 0

Key Gaussian/BM Formulas

• E[Bt ] = 0, Var(Bt ) = t, E[Bt2 ] = t, E[Bs Bt ] = min(s, t)

2 t/2
• E[eθBt ] = eθ (moment generating function of N (0, t))

• Bt − Bs ∼ N (0, t − s) independent of Fs

• For Z ∼ N (0, 1): E[Z 2n ] = (2n − 1)!! = (2n)!/(2n n!); E[Z 2n+1 ] = 0

Quadratic Variation Summary

[B, B]t = t, [B, f (t)]t = 0 (BV functions have zero QV with anything), [f (t), g(t)]t = 0
Multiplication table: dBt · dBt = dt, dBt · dt = 0, dt · dt = 0
If dXt = µdt + σdBt and dYt = νdt + ρdBt , d[X, Y ]t = σρ dt.
Binomial Model Summary
q = (1 + r − d)/(u − d); S0 = EQ [S1 ]/(1 + r), V0 = EQ [Vn ]/(1 + r)n

31
No-arbitrage: d < 1 + r < u ⇔ q ∈ (0, 1)
FFTAP: No-arbitrage ⇔ ∃ EMM Q. SFTAP: Complete ⇔ unique Q.
SDE/GBM Summary
dSt = µSt dt + σSt dBt ⇒ St = S0 exp((µ − σ 2 /2)t + σBt )
Under Q (θ = (µ − r)/σ): dSt = rSt dt + σSt dB̃t ⇒ St = S0 exp((r − σ 2 /2)t + σ B̃t )
EQ [St ] = S0 ert , VarQ [ln St ] = σ 2 t
BSM Formula Summary
C = S0 N (d1 ) − Ke−rT N (d2 ) √ √
d1 = [ln(S0 /K) + (r + σ 2 /2)T ]/(σ T ); d2 = d1 − σ T
Put price P = Ke−rT N (−d2 ) − S0 N (−d1 ) [by put-call parity]
Change of Measure Summary
EQ [X] = EP [ZX] where Z = dQ/dP; EP [Z] = 1; Z > 0 a.s.
Bayes: EQ [X | Ft ] = EP [ZX | Ft ]/EP [Z | Ft ] = EP [ZX | Ft ]/Zt
RT RT Rt
Girsanov: ZT = exp(− 0 θt dBt − 12 0 θt2 dt); B̃t = Bt + 0 θs ds is Q-BM

APPENDIX: COMPLETE DEFINITIONS QUICK REFERENCE

• σ-algebra (F): Collection of subsets of Ω closed under (i) complement, (ii) countable
unions; containing Ω. Elements are “events.”

• Probability Measure (P): Function F → [0, 1] with P(∅) = 0, P(Ω) = 1, countably

additive for disjoint events.

• Probability Space: Triple (Ω, F, P).

• Borel σ-algebra B(R): Smallest σ-algebra containing all open intervals – the “right”
σ-algebra for real-valued RVs.

• Lebesgue Measure: λ((a, b)) = b − a. Extends to all Borel sets uniquely.

• Random Variable (X): Measurable function X : Ω → R, i.e., X −1 (B) ∈ F for all

B ∈ B(R).

• Distribution of X: Measure µX on (R, B(R)) with µX (B) = P(X ∈ B).

• Filtration: Increasing family of σ-algebras Fs ⊆ Ft for s ≤ t (information only grows).

• Natural Filtration of Xn : Fn = σ(X0 , X1 , . . . , Xn ).

• Adapted Process: Xn is Fn -measurable for each n (no peeking into future).

• Martingale: Adapted, integrable, E[Mn+1 | Fn ] = Mn a.s.

• Submartingale: E[Mn+1 | Fn ] ≥ Mn (tends up). Supermartingale: E[Mn+1 | Fn ] ≤

Mn (tends down).

• Jensen’s Inequality: φ convex ⇒ E[φ(X)] ≥ φ(E[X]). Conditional: E[φ(X) | G] ≥

φ(E[X | G]).

• Conditional Expectation E[X | G]: Unique G-measurable Y with A Y dP = A XdP

R R

for all A ∈ G.

• Tower Property: E[E[X | G] | H] = E[X | H] for H ⊆ G.

32
• Brownian Motion: Continuous process, B0 = 0, independent N (0, t − s) increments,
continuous paths.

• QV of BM: [B, B]t = t a.s. Key heuristic: (dBt )2 = dt.

• Lévy’s Theorem: Continuous local martingale M with M0 = 0, [M, M ]t = t ⇒ M is a

BM.
RT
• Itô Integral: 0 ∆dBt = L2 limit of ∆(tj )(Btj+1 −Btj ) (left endpoints, non-anticipating).
P

 2  hR i
RT T
• Itô Isometry: E 0 ∆dBt = E 0 ∆ 2 dt .

• Itô’s Formula: df (t, Bt ) = ft dt + fx dBt + 21 fxx dt. For Itô process dXt = µdt + σdBt :
df = (ft + µfx + 12 σ 2 fxx )dt + σfx dBt .

• Itô Product Rule: d(Xt Yt ) = Xt dYt + Yt dXt + dXt · dYt .

• GBM: St = S0 exp((µ − σ 2 /2)t + σBt ) solves dSt = µSt dt + σSt dBt .

• BSM PDE: Vt + rS · VS + 12 σ 2 S 2 · VSS − rV = 0 (terminal condition V (T, S) = payoff).

• BSM Call: C = S0 N (d1 ) − Ke−rT N (d2 ).

• Risk-Neutral Measure Q: EMM equivalent to P; discounted prices are Q-martingales;

V0 = e−rT EQ [payoff].

• FFTAP: No-arbitrage ⇔ ∃ EMM.

• SFTAP: Complete market ⇔ unique EMM.

• Radon-Nikodym Derivative: Z = dQ/dP ≥ 0 with EP [Z] = 1; EQ [X] = EP [ZX].

• Girsanov’s Theorem: ZT = exp(− θdBt − 21 θ2 dt); B̃t = Bt + θds is Q-BM.

R R R

Rt
• MRT: Every square-integrable (FtB )-martingale = M0 + 0 φs dBs (unique φ).

• Feynman-Kac: u(t, x) = E[h(XT ) exp(− cds) | Xt = x] solves ut +µux + 21 σ 2 uxx −cu =

R

0.

• Discounted FK: u(t, x) = EQ [e−r(T −t) h(XT ) | Xt = x] solves BSM PDE.

• Market Price of Risk: θ = (µ − r)/σ. Unique in complete market; determines the

change of measure.
RT
• Novikov Condition: EP [exp( 21 0 θt2 dt)] < ∞ ⇒ stochastic exponential is a true mar-
tingale.

• Self-Financing Portfolio: Portfolio with no external inflows/outflows; dXt = ∆t dSt +

rBt dt.

33