You know how orders travel from strategy to matching engine. Today we ask the harder question: how does anyone know what price to put on that order? The answer — all of it — is probability.
Probability foundations → distributions → random walks → options pricing → Greeks → vol surface → Kelly criterion → the live desk
Count equally-likely outcomes.
Fair die: P(6) = 1/6. Symmetric, clean. Breaks immediately when outcomes are not equally likely — i.e. in all real markets.
Long-run relative frequency in infinite trials.
NIFTY moved >2% on 47 of 1,000 days → P ≈ 4.7%. Foundation of backtesting. Fails when regimes change.
Degree of belief, updated by evidence.
Prior belief × new evidence → updated belief. How traders react to news. How markets move on surprises.
Classical to build payoff models. Frequentist to calibrate volatility from history. Bayesian to update positions when the RBI speaks, when earnings land, when geopolitics shift. The maths is unified — the interpretation depends on context.
When implied volatility jumps after a surprise central bank decision, which type of probability just updated — and who is doing the updating?
Your desk has studied 50 past RBI meetings.
Prior belief: P(rate hike) = 20%
Then at 8:05am: 10-year bond yields spike +15bps in 5 minutes. Bonds are pricing something in. Should you update?
| What happened at RBI | Bond yield spiked? |
|---|---|
| Hiked rates (10 times) | 9 out of 10 → 90% |
| Held rates (40 times) | 12 out of 40 → 30% |
Bond traders react to inside information. A yield spike is strong evidence of a hike, but not certainty — it happens 30% of the time even when rates are held.
Prior: 20% → Posterior: 43% — probability more than doubled from one data point.
Immediate actions: buy near-dated puts on rate-sensitive stocks, trim short-vega positions before 2pm, widen spreads on index options. All driven by this one Bayes update.
P(hold | spike) = 1 − 42.9% = 57.1%. Cross-check: (0.30 × 0.80) / 0.42 = 0.24 / 0.42 = 57.1% ✓. All posterior probabilities must sum to exactly 100%.
A call with strike K pays max(S−K, 0) at expiry. The option price is:
Expected payoff, discounted to today. Black-Scholes solves this integral in closed form. This is literally what an option price is.
Game A: Win ₹10 for certain. EV = ₹10.
Game B: Win ₹1,000 at 1% probability, else −₹0.10. EV ≈ ₹9.90.
Same EV. Completely different risk. What's missing? Variance. The options desk manages both, always.
A market maker quotes a spread of ₹0.50 and does 10,000 trades per day. Their edge is ₹0.25 per side. Why can they still lose money over a week?
# 5 days: +1.2%, −0.8%, +2.1%, −1.5%, +0.4% μ = (1.2 − 0.8 + 2.1 − 1.5 + 0.4) / 5 = 0.28% deviations² = 0.846, 1.166, 3.312, 3.168, 0.014 Var = mean(deviations²) = 1.701 σ_daily = √1.701 = 1.30% # Annualise → multiply by √252 σ_annual = 1.30% × √252 = 1.30% × 15.87 = 20.6%
σ (sigma) IS the price of uncertainty. An option on a stock with σ=30% costs roughly twice as much as the same option on a stock with σ=15%. Volatility is literally what options traders buy and sell.
Variance adds linearly for independent variables. So why does σ scale with √T and not T? This matters enormously for options pricing.
σ_annual = 18% → what's the 25-day range?
An ATM NIFTY option with 25 days to expiry must price in a 1-standard-deviation move of ±5.65%. That determines the time value.
4× the time → only 2× the uncertainty. Long-dated options are not linearly more expensive than short-dated ones. This √T scaling is why option pricing is non-linear.
A 1-month option has σ = 5.65%. A 3-month option has σ = 9.78% (×√3). Why isn't it 16.95% (×3)? What stops uncertainty from tripling?
| Range | Probability | NIFTY (σ=1.13%/day) |
|---|---|---|
| ±1σ | 68.3% | Move <1.13% → ~172 days/yr |
| ±2σ | 95.4% | Move <2.26% → ~12 days outside |
| ±3σ | 99.7% | Move <3.39% → ~1 day/yr (theory) |
| ±4σ | 99.994% | Should almost never happen |
Add many independent random variables → their sum approaches normal, regardless of individual distributions. Daily returns = sum of thousands of individual trade impacts. CLT says the aggregate is approximately bell-shaped. This is the statistical foundation of the normal model.
The normal distribution is symmetric. Real market crashes happen more often than rallies of the same size. What does this asymmetry tell you about the shape of the true returns distribution?
Normal distribution allows S < 0. A share cannot be negative. Solution: model log-returns as normal → prices are log-normally distributed and always positive.
Real crashes are far more frequent than normal predicts. The 1987 crash (−22.6% in one day) was a 20σ event under normality — predicted to occur once in 1089 years. It happened on a Tuesday.
Equity markets are negatively skewed: crashes are sharp and sudden; rallies grind higher. This asymmetry is encoded in the volatility smile — OTM puts trade at higher implied vol than OTM calls. Traders actively buy and sell this skew.
| Event | Normal says | Reality |
|---|---|---|
| Daily >4% move | 0.2% of days | ~1.5% of days |
| Daily >6% move | 0.0004% | ~0.3% of days |
| Crash (>10%) | ≈ never | Happens |
If fat tails are well-known, why does Black-Scholes still assume log-normal (thin tails)? What do practitioners do to compensate for this known flaw?
μS dt = drift — deterministic expected return. Slow. Predictable.
σS dW_t = diffusion — random shock each instant. Dominates in the short run.
Over 1 year: drift moves price ~8%. Over 1 millisecond: drift is essentially zero. HFT strategies ignore drift completely — they only see noise.
| Property | What it means |
|---|---|
| Independent increments | Tomorrow's move doesn't depend on yesterday's. Markets have no memory (in this model). |
| Scales with √t | Uncertainty grows with square root of time — not linearly. |
| Continuous paths | No instantaneous jumps. (Reality adds Poisson jump processes.) |
| Martingale | Under risk-neutral measure: best guess for tomorrow's price is today's price. No free money. |
GBM says: price moves are unpredictable in direction but predictable in scale. We can't say which way NIFTY moves tomorrow. But we can say that the distribution of moves has σ = 1.13% per day. And option prices are all about pricing that distribution.
If markets have no memory (independent increments), can technical analysis — chart patterns, support levels — ever work? What assumption would need to be violated for it to be valid?
Pay premium C upfront. At expiry:
If S < K: option expires worthless. Loss = C.
If S > K: payoff = S − K. Profit = S − K − C.
Break-even = K + C.
Maximum loss: capped at C. Upside: unlimited. This asymmetry is what you pay for.
Every option price splits into two components:
At expiry: time value = 0. Before expiry: you pay for the possibility of further favourable moves. This time value decays every day — that decay is theta.
A call with K=100 costs ₹5. Stock is at ₹98. Intrinsic value = zero. You paid ₹5 for pure time value. What must happen for you to break even?
N(d₂) = P(call expires in the money) under risk-neutral measure
N(d₁) = Delta — shares needed to hedge one call
Formula says: "Value = stock × hedge ratio − discounted strike × probability of paying it." It's a weighted expected value.
| Symbol | Meaning | Where it comes from |
|---|---|---|
| S | Current stock/index price | Live market feed |
| K | Strike price | Contract specification |
| T | Time to expiry (in years) | Calendar |
| r | Risk-free interest rate | Central bank / Tbills |
| σ | Volatility — the key unknown | You estimate this |
σ is the only input you cannot observe directly. Name three ways to estimate it. What's the fundamental problem with each approach?
| S | ₹100 (current price) |
| K | ₹100 (strike = ATM) |
| T | 30 ÷ 252 = 0.119 years |
| r | 6% = 0.06 |
| σ | 20% = 0.20 |
This is the 1-sigma log-return range over the 30-day life of the option — the core "uncertainty" quantity.
N(x) = area under the normal curve to the left of x
N(d₂) = 0.527 = 53% = probability call expires in the money.
= 100 × 0.20 × √(0.119 ÷ 6.283)
= 100 × 0.20 × 0.138 = ₹2.75
Off by ~12%. Good enough for a mental sanity check on a trading floor, not for actual pricing.
For an ATM option, ln(S/K)=0. So d₁ depends only on drift and time. Why is N(d₂)=53% and not exactly 50%? The drift term (r + σ²/2)T pushes d₁ and d₂ above zero, lifting both probabilities slightly.
① Hedge ratio: Δ=0.60 → buy 60 shares to hedge 1 call contract.
② Price sensitivity: stock moves ₹1 → option moves ₹Δ.
③ Probability: ATM call Δ ≈ 0.50 ≈ 50% chance of expiring ITM.
As stock moves, Δ moves. A hedge that was correct at ₹100 is wrong at ₹110. The desk must continuously re-hedge. The cost of this re-hedging is what you buy when you pay time value on an option.
Buy a call with Δ=0.30 and sell a put with Δ=−0.70 at the same strike and expiry. What is your net delta — and what does it mean to hold that position?
10 NIFTY call contracts
Strike K = ₹10,000 · Spot S = ₹10,000 (ATM)
σ = 18% · T = 30 days · Δ = 0.52
NIFTY lot size = 50 units per contract
Equivalent to holding 260 units of NIFTY outright. If NIFTY rises ₹1, you gain ₹260.
Delta-neutral: small moves in NIFTY no longer affect your P&L.
| Position | Calculation | P&L |
|---|---|---|
| 10 calls (Δ=0.52) | 0.52 × ₹200 × 50 × 10 | +₹52,000 |
| Short 260 futures | −₹200 × 260 | −₹52,000 |
| Net | ≈ ₹0 ✓ |
Hedge worked. The ₹200 move was neutralised.
Every re-hedge you sell a little higher or buy a little lower. The slippage cost of all these re-hedges over the life of the option is exactly theta — the daily time decay you pay as the option buyer.
NIFTY then falls ₹300 from ₹10,200 back to ₹9,900. You now need to buy back some of those futures. At what price are you buying vs where you sold? Who wins from this re-hedging — the buyer or the seller?
Long gamma (buy options): profit from large moves in either direction. Delta grows when stock rallies — good. Delta shrinks when it falls — also good (less exposure).
Short gamma (sell options): you collect premium but lose on large moves. Delta grows against you as stock moves.
Long gamma → costs you theta (you pay daily time decay).
Short gamma → earns you theta (you collect daily time decay).
Every morning: "Are we long or short gamma? Is the theta income worth the gamma risk we're carrying?"
You are short gamma. NIFTY drops 3% in 10 minutes. Describe exactly what happens to your delta hedge and why you might lose money even if you re-hedge immediately.
The Black-Scholes PDE rearranged. Theta income ≈ cost of gamma hedging. Short gamma earns theta. Long gamma pays theta. They are two sides of the same trade.
Sell an ATM call for ₹10 with 30 days to expiry. With stable vol, you earn roughly ₹0.33/day in time value (but non-linearly — much more in the final week).
Friday 4pm. You're long a call expiring Monday. You don't hedge over the weekend. What happens to your option value over Saturday and Sunday — and can you do anything about it?
Reverse the B-S formula: observe the option price → solve for σ that produces it. That σ is implied volatility — the market's consensus on future vol.
IV > expected RV → option is overpriced → sell vol
IV < expected RV → option is cheap → buy vol
The desk is not just trading stocks. It is trading volatility itself.
| Event | IV moves | Effect |
|---|---|---|
| Earnings due | ↑ spike | All options expensive |
| Market crash begins | ↑ explodes | Puts very expensive |
| Quiet August | ↓ vol crush | Sellers profit |
| Post central bank | ↓ fast | "Sell the news" |
| RBI surprise hike | ↑ spike | All options reprice |
Historically: IV > RV on average by 2–5 vol points. Selling options earns this "volatility risk premium" over time. But it comes with gamma risk. The desk tracks IV − RV daily. When IV is unusually high, they sell. When unusually low, they buy.
NIFTY VIX jumps from 14 to 22 overnight after political news. Your delta is near zero. Your vega is +₹50,000 per vol point. What happened to your P&L — and why did it happen even though you had no directional bet?
OTM puts have higher implied vol than ATM or OTM calls. The market assigns higher probability to large crashes than log-normal predicts. This "smirk" encodes the true fat-left-tail distribution of market returns.
If Black-Scholes were correct, all options on the same underlying would have the same implied volatility. They don't. The surface is the market's honest confession that B-S is wrong.
Short-dated options often spike before events (earnings, RBI, elections) then crash after ("vol crush"). Long-dated options are more stable. The term structure captures the market's time-varying uncertainty forecast.
① Mark all positions at live vol surface (not flat vol)
② Find mispricings between strikes → volatility arbitrage
③ Hedge vega across the surface — not just total vega
④ Maintain a smooth, arbitrage-free surface model (SVI, SABR)
Put-call parity is model-free and always holds. So if put IV ≠ call IV, something else must explain the difference — and it does: it's the skew of the true probability distribution, not an arbitrage.
The vol surface gives different IVs to calls and puts at the same strike. But they must satisfy put-call parity. How can the surface be internally consistent?
# 60% coin, even odds (b=1) f* = (1×0.60 − 0.40) / 1 = 0.20 / 1 = 20% of capital per flip # For portfolio management: f* = μ/σ² = Sharpe ratio / σ
Kelly maximises E[log(wealth)] — the geometric growth rate. Bet more than Kelly: variance destroys you. Bet less: you're suboptimal.
Most desks use ½ Kelly: accept 25% less growth in exchange for dramatically lower drawdown and ruin risk.
For any f > f*, probability of eventual ruin → 1 as bets → ∞. Regardless of your edge. This is why risk management exists: unconstrained overconfidence always ends the same way.
Kelly assumes you know p exactly. In practice you estimate it from data. What happens if you overestimate your edge by 10%? Design a rule to protect against this estimation error.
| Metric | Value |
|---|---|
| Total trades | 200 |
| Winning trades | 116 (58%) |
| Average profit on wins | ₹8,000 |
| Average loss on losses | ₹5,000 |
| Starting capital | ₹10,00,000 |
For every ₹1 you risk, you win ₹1.60 when right.
Positive EV confirmed. If this were negative, Kelly would output f* < 0 — meaning "don't trade at all."
Full Kelly says bet 31.75% of capital on each trade.
Risk ₹1,59,000 with avg loss of ₹5,000 per lot →
Max lots = 159,000 ÷ 5,000 = 31 lots
Round down, never up. The Kelly math is asymmetric — overbet by a small amount destroys more wealth than underbetting by the same amount.
These backtest stats assume the future looks like the past. If the true win rate is 52% (not 58%), what does Kelly output? Recalculate f* with p=0.52. Does the sign change?
Long ATM straddle (call + put). S=K=₹10,000. Delta ≈ 0 (delta-neutral). Over one day:
| Scenario | ΔS | Gamma P&L | Theta cost | Net P&L |
|---|---|---|---|---|
| Quiet day | ₹50 | +₹2,813 | −₹12,000 | −₹9,187 |
| Moderate move | ₹200 | +₹45,000 | −₹12,000 | +₹33,000 |
| Big crash | ₹500 | +₹281,250 | −₹12,000 | +₹269,250 |
Gamma P&L = ½ × 450 × (ΔS)². Break-even daily move ≈ √(2×12,000/450) ≈ ₹231 ≈ 2.3%
| Greek | Bet you're making | Win when |
|---|---|---|
| Long Δ | Market goes up | Stock rallies |
| Long Γ | Market moves big | High realized vol |
| Short Θ | Pay for time | You buy it cheaply |
| Long V | Vol will rise | IV jumps |
Risk-adjusted return. Combines EV and variance into one number. SR=1.0 is respectable. SR=2.0 is excellent. SR>3 is almost certainly data-mining or fraud.
Not maximum EV. Not minimum variance. Maximum risk-adjusted return subject to regulatory capital limits, drawdown limits, and position concentration limits. Kelly gives the theoretical optimum; regulatory reality constrains the practical answer.
A desk has positive EV on every individual trade. How can it still lose money over a month? Name three mathematically distinct mechanisms.
NIFTY futures down 0.8%. India VIX 14.2 → 17.1 (+20%). US flat overnight.
VIX +20% → IV up → long vega positions profitable overnight. Gamma shifted our deltas — need to re-hedge. Bayesian: update forward vol expectations.
NIFTY opens −1.2% on high volume. Spot vol spikes. Market makers widening spreads.
Net Δ: −42 → buy futures
Net Γ: +₹8,500/1% move
Net Θ: −₹12,000/day
Net V: +₹45,000/vol pt
RBI meeting at 14:00. Prior: 80% no change. Bond yields rising.
Yields rising → update to 65% no change. Trim near-dated short vega to reduce event risk before 14:00.
① What is our EV? For every position: is E[P&L] positive after all costs?
② What is our variance? How wrong could we be? Can we survive a 3σ adverse day?
③ Are we being paid for the risk? Is IV above RV? Is theta above gamma cost?
At 13:55 (5 minutes before RBI decision), should you buy or sell short-dated options? Walk through the EV, gamma, vega, and theta calculation for each choice.
Mathematical
Why does σ scale with √T? Prove it from first principles using the definition of variance of a sum of independent variables.
N(d₂) is the probability of expiring ITM. Why is delta N(d₁) not N(d₂)? What's the difference and why does it matter for hedging?
If you're delta-neutral, can you still lose money on a big market move? Walk through the exact mathematics of why.
Kelly assumes you know p exactly. What happens if you systematically overestimate your edge by 5%? Derive the result.
No easy answer
Is implied volatility a probability? It's extracted from option prices using a model (B-S) that is known to be wrong. Does a probability derived from a wrong model mean anything?
Markets are supposedly efficient — prices reflect all information. If that's true, how does the volatility risk premium (IV > RV on average) persist? Who is paying for it and why?
The 2008 financial crisis. Models said 5σ was impossible. It happened. Was the model wrong, the inputs wrong, or is "probability" itself the wrong concept for rare events?
An AI predicts the next price move with 51% accuracy. Walk through Kelly, EV, and what happens to the edge as every firm deploys the same model.