Risk & Position Sizing

Phase 9 · Week 70 · ~20 hours

Where it sits. Two investors can buy the identical twenty stocks over the same decade and end it in different financial universes, purely because they sized them differently. This is the first module of the final analytical phase, and the hinge of the whole program. For sixty-four weeks you have been building the machine that answers "is this a good business, run by good people, and is it cheap?"; Phase 8 certified that you can do it fast and unaided. This module changes the question. It stops asking what do I think of this company? and starts asking how much of my capital should ride on that opinion, given that I might be wrong? That is a different and, for most people, harder skill, and it is the one that decides whether decades of good analysis actually turn into wealth or get handed back in a single oversized mistake. The order matters: you learn to size only after you can analyze, because sizing without analysis is just gambling with a spreadsheet, and analysis without sizing is a research report that never had to survive a bad year. M9.02 will take the single-position sizing you build here and assemble it into a whole portfolio (hurdles, cash, factors, the active/passive debate); M9.03 will teach you when to un-size (selling); M9.04 welds all of it into one written process. This module supplies the atom they are built from: one position, correctly sized against its real downside.

Suggested hour budget. Core content (what risk really is; the impairment taxonomy; Kelly operationalized; conviction × downside; hidden correlation; the anti-ruin overlay): 9h · worked examples (a full Kelly-to-size workup; a hidden-correlation decomposition; a conviction × downside portfolio): 3h · practice set (P×magnitude, Kelly both forms, factor decomposition, "size this given the downside," a ruin audit): 4.5h · applied mini-project (size a real live position end-to-end): 2h · flashcards + mastery check: 1.5h.

App link. This module opens the Portfolio & Risk branch and lights up the sizing tools you will use forever. The Kelly sizing sandbox (shared plumbing with M6.03's decision-science tools) now runs in investment mode: enter your upside, downside, and probability, and it returns raw Kelly, your chosen fractional-Kelly size, the hard cap, and the geometric-growth curve so you can see the overbetting cliff. The conviction × downside sizer turns the matrix in §4.8 into a live grid. The factor-exposure decomposer takes your holdings, maps each to its dominant factors, and shows your effective number of bets, which is the hidden-correlation lesson made mechanical (competency C9). The ruin-check overlay flags any position or correlated cluster sized to threaten survival. The module gate, a sizing problem set at ≥85%, is delivered as a mixed numeric-and-reasoning set in the assessment engine, and it is the first gate the app scores partly on restraint: the "right" answer is frequently "smaller than you wanted."

Learning objectives

By the end you can:

1. State the two definitions of risk and judge when each applies. Give the practitioner's definition (risk = the probability and magnitude of a permanent loss of capital) and the academic one (risk = volatility / beta), explain why the difference is not semantic, and argue fairly for the case where volatility genuinely is risk (when you might be a forced seller).
2. Formalize risk as probability × magnitude of impairment and explain Howard Marks's central point: that risk is an ex ante property of a decision that is largely invisible, cannot be read off a track record, and is not the same as the volatility you can measure after the fact.
3. Use the margin of safety as your primary risk-control tool and connect it directly to sizing: a wider margin of safety shrinks the downside d, which mathematically justifies (but never mandates) a larger position.
4. Name and detect the four ways capital is permanently impaired (business deterioration, overpaying for a good business, leverage/forced selling, and fraud) and map each to the diagnostics you already own from Phases 1–5.
5. Size a position with Kelly and fractional Kelly. Re-derive f* from the growth-of-wealth objective, compute it for capped-downside investments (f* = p/d − q/u), read the geometric-growth curve, and justify half- or quarter-Kelly from estimation error, drawdown tolerance, and correlation, showing numerically that double Kelly earns zero growth.
6. Apply the conviction × downside heuristic, bigger where conviction is high and downside small, smaller or pass where either fails, using a sizing matrix that embeds an implicit "no single mistake costs more than ~2.5% of the portfolio" rule.
7. Decompose apparent diversification into factor exposure. Show why 30 names loaded on one factor (Indian NBFCs, US regional banks, oil) is one bet, compute the correlation floor σ√ρ and the effective number of independent bets n/[1+(n−1)ρ], and resize to a factor cap.
8. Apply the anti-ruin overlay. Set single-name and factor caps, refuse leverage that can force selling, size smaller under Knightian uncertainty than under quantifiable risk, and explain why survival is a mathematical (ergodic), not a merely prudent, first-order goal.

Prerequisites & connections

Builds on. M6.03 (Mental Models & Decision Science) is the direct parent: it derived the intuition of Kelly (f* = (b·p − q)/b, read as edge/odds), the investing generalization (f* = p/d − q/u), the risk-vs-uncertainty distinction (Knight), and ergodicity/ruin-avoidance, all as theory. This module is the practice: the same tools re-derived a little deeper and turned into a live sizing policy you can run on a real position tomorrow. M6.01 (Graham to Buffett) gave you the margin of safety as a valuation idea; here it becomes a sizing idea. M6.02 §4.9 staged the concentration-vs-diversification debate and Pabrai's Dhandho asymmetry ("heads I win, tails I don't lose much"); this module supplies the mathematics that adjudicates both. M3.02 (Risk, Return & the ERP) is where you met the academic definition of risk (variance, beta, CAPM, the efficient frontier), which we now put in honest dialogue with the practitioner's definition. M2.03 (Returns on Capital) and M2.07 (fraud screens) supply the impairment diagnostics; M5.01–M5.02 (banks, NBFCs) supply the canonical hidden-correlation factor (Indian credit); M3.05 (reverse DCF) is how you quantify "what's priced in," which is the raw material for estimating the downside d.

Feeds forward. M9.02 (Portfolio Construction) takes single-position sizing and assembles the book (opportunity-cost hurdles, cash as a position, the factor menu, active vs passive) using the factor-exposure lens built here. M9.03 (Selling & Mistakes) adds "position grown beyond tolerance" as one of the five legitimate sell triggers, which is this module's caps enforced through time. M9.04 (The Process) wires sizing into the checklist → journal → calibration loop. Phase 10's capstones require a sized recommendation, not just a valuation: a thesis with no position size attached is graded incomplete. And the app's portfolio simulator scores you, from here on, on whether your sizes were survivable and process-consistent, never on whether they happened to pay.

The one-sentence version of this module. Risk is the probability and size of a permanent loss, an unknowable, forward-looking thing you can never read off a price chart or a track record, so you control it not by predicting it but by sizing: buy with a margin of safety, put more weight where conviction is high and the downside is genuinely capped, remember that thirty names exposed to one factor is one bet, and never let any single position or correlated cluster be large enough that being wrong could end the game, because you only get to compound if you survive.

Core teaching content

4.4 The margin of safety as the primary risk tool: now applied to sizing

You met the margin of safety in M6.01 as Graham's central idea: buy a security for meaningfully less than your conservative estimate of its intrinsic value, so that the gap, the margin, absorbs bad luck, analytical error, and the future's refusal to match your forecast. The Intelligent Investor's chapter 20 calls it the three words that capture the secret of sound investment: margin of safety. In M6.01 it was a buy rule. Here it becomes a sizing rule, and the connection is exact and worth making mechanical.

The margin of safety controls both terms of risk = probability × magnitude:

• It lowers the probability of permanent impairment, because a price already well below conservative value has less distance to fall to reach a genuinely impaired level, and more of the bad news is already in the price.
• It lowers the magnitude of impairment if you are wrong, because you began closer to the floor. If your conservative ("bear") value is ₹640 and you paid ₹800, your estimated downside is roughly −20%; if you had paid ₹1,000 for the same business, your downside is −36% — same company, wider hole, purely from the entry price.

And here is the link to sizing that the rest of the module runs on. In the Kelly formula you will operationalize in §4.6, the downside fraction d is precisely your permanent-loss estimate if the thesis is wrong, and the margin of safety is what sets d. A bigger margin of safety means a smaller d, and a smaller d (as f* = p/d − q/u makes explicit) raises the size the mathematics will justify. The margin of safety and the position size are therefore the same decision viewed twice. A deeply-discounted, well-protected position both deserves and permits more weight; a fully-priced one, however wonderful the business, both permits and deserves less. This is why "how much?" can never be answered from the quality of the business alone. It depends on the price you paid, because the price is what sets your downside. We make this quantitative in §4.6 and §4.8; for now, hold the principle: the margin of safety is your primary risk tool, and it feeds directly into how much you are allowed to own.

4.8 Conviction × downside sizing: the practical heuristic

Full Kelly needs inputs you don't have to a precision you can't achieve; you are not going to compute f* = p/d − q/u to three decimals for every idea and bet it. What you will do, and what most excellent concentrated investors actually do, is use a heuristic that captures Kelly's DNA (size ∝ edge, size ∝ 1/downside) without the false precision. It has two axes:

• Conviction. How confident are you that the thesis is broadly right? This is Kelly's p, but held honestly: it folds in the strength of your evidence, whether the business is inside your circle of competence, and how much of your case is fact versus hope. Calibrate it against base rates (M6.03 §4.6): "high conviction" should be rare.
• Downside. If you are wrong, how much do you permanently lose? This is Kelly's d, set by the margin of safety (§4.4): the gap between price and your conservative/bear-case value, plus the asset backing beneath it. A capped, asset-backed downside is small d; an expensive, story-driven stock with nothing underneath is large d.

Cross them into a sizing matrix. The numbers below are a sane default for a concentrated-but-survivable book (adjust the scale to your own risk tolerance and portfolio size); the structure is the lesson:

Read the matrix and notice what it encodes. The biggest positions live in the top-left, where high conviction meets a small downside, never on high conviction alone. A brilliant idea you're sure about, but with a large downside if you're wrong, gets a small size (top-right), because the asymmetry isn't there. And a small downside you're only lukewarm about gets a modest size (bottom-left). This is Pabrai's Dhandho made operational: you swing hard only when heads-you-win-big is paired with tails-you-barely-lose. The matrix also embeds a hidden discipline: check the product of size and downside in each cell. The worst case is roughly High/Medium at 8% × 30% ≈ 2.4% of the portfolio, and High/Small at 12% × 15% ≈ 1.8%. Every cell keeps a single position's permanent-loss contribution under about 2.5% of the whole portfolio. That is not an accident; it is the anti-ruin overlay (§4.12) baked into the grid, so that even a total surprise on your largest position costs a couple of percent, not a couple of decades. When you build your own matrix, design it around that constraint: no single mistake should cost more than you can shrug off.

The heuristic's honesty is its virtue. It does not pretend you know p = 0.73; it asks you to sort into three conviction bands and three downside bands you can defend, and it returns a size that has Kelly's shape without Kelly's false precision. Use it as your default day-to-day sizer, and reserve the explicit Kelly computation for the rare, high-stakes position where the extra rigor earns its keep.

4.10 Hidden correlation: thirty names, one factor, is one bet

Now the most important and most under-appreciated idea in the module, the one that quietly destroys portfolios that look diversified. Diversification counts names, but risk lives in factors. A portfolio of thirty names all exposed to the same underlying factor is, in the way that matters, a single large bet, not thirty small ones.

Here is the trap. You hold thirty stocks, no position over 4%, and you feel diversified: no single company can hurt you much. But suppose twenty of them are Indian lenders: banks, NBFCs, and housing-finance companies. They have different names, tickers, and management teams, but they are all driven by the same underlying forces, namely the domestic credit cycle, the availability and cost of wholesale funding, RBI policy, and the health of the borrower. When one of those factors turns, whether a large lender defaults and the funding market freezes (IL&FS, September 2018; the DHFL cascade that followed) or a nationwide shock spikes expected credit losses (COVID, March 2020), all twenty move together, and their pairwise correlations, which looked moderate in calm times, snap toward 1.0 exactly when you needed them to be low. You did not own twenty independent lender bets; you owned one leveraged bet on Indian credit, wearing twenty costumes. The same structure recurs everywhere: thirty US regional banks in March 2023 (all exposed to deposit flight, held-to-maturity duration losses, and commercial real estate, and the SVB cascade took the whole group down at once); thirty oil & gas E&P names (all one bet on the crude price); thirty IT-services exporters (all one bet on US discretionary tech spending and the USD/INR rate).

Different tickers, one factor, one bet.

Make it quantitative, because the numbers are shocking. Take an equal-weighted portfolio of n names, each with volatility σ, and suppose every pair has the same correlation ρ. The portfolio variance works out to:

Var(portfolio) = σ² · [ 1/n + ρ · (n − 1)/n ]

Two limits tell the whole story. If the names were truly independent (ρ = 0), the bracket becomes 1/n and portfolio volatility falls as σ/√n, which is diversification's famous benefit, vanishing risk as you add names. But as n grows with ρ > 0, the first term shrinks to nothing and the variance approaches a floor:

Var(portfolio) → σ² · ρ, so volatility floor = σ · √ρ

This is the punchline: no amount of adding correlated names gets you below σ·√ρ. Suppose each lender has σ = 40% volatility and the group's pairwise correlation is ρ = 0.7. The floor is 40% × √0.7 = 40% × 0.837 = 33.5%. Twenty such names bring you essentially to that floor: a portfolio volatility around 34%, versus a single name's 40%. You added nineteen names and cut your risk by about 15%. Had those twenty names been genuinely independent, the same twenty would have cut volatility from 40% to 40%/√20 = 8.9%, a 78% reduction. The gap between a 15% reduction and a 78% reduction is the entire cost of hidden correlation, and you cannot see it in the list of holdings; you can only see it by asking what factors each name loads on.

The effective number of bets. A clean way to summarize the damage is to convert your n correlated names into the number of independent bets they are equivalent to:

N_eff = n / [ 1 + (n − 1)·ρ ]

For our twenty lenders at ρ = 0.7: N_eff = 20 / [1 + 19×0.7] = 20 / 14.3 = 1.4. Twenty names, one-and-a-bit bets. Sanity-check the formula at the extremes: ρ = 0 gives N_eff = n (fully independent, so every name is its own bet); ρ = 1 gives N_eff = 1 (perfectly correlated, one bet no matter how many names). Your feeling of diversification is calibrated to n = 20; your actual diversification is N_eff = 1.4. That factor-of-fourteen gap between perceived and real diversification is how portfolios that "did everything right" on position limits still lose 35% in a sector crisis.

What to do about it: size by factor, not by name. The correction is to stop sizing only at the name level and start sizing at the factor level:

• For every position, write down its two or three dominant factors, the macro/sector/style forces that actually drive it (e.g., "Indian credit cycle + wholesale funding"; "crude price"; "US enterprise IT spend + USD/INR"; "long-duration rates"; "China industrial demand").
• Set factor caps the same way you set single-name caps, e.g., no more than ~20–25% of the portfolio exposed to any one factor, and enforce them before the position limits, because the factor cap is the one that actually protects you.
• Treat a correlated cluster as one position for sizing purposes: twenty lenders at 4% each is a 80% bet on Indian credit, which no sane risk policy would allow; recognizing it as one 80% bet is the whole point.
• Ask, on every portfolio, the question that hidden correlation is designed to force: "How many bets do I actually have?" That is the N_eff question. If the honest answer is "three," then own three ideas at defensible sizes and stop pretending the thirty tickers are thirty bets.

This is the single most valuable habit in portfolio risk, and it is almost never taught to retail investors, who are told "own 20–30 stocks and you're diversified" with no mention of the factor structure that decides whether that is true or a dangerous illusion.

4.11 Sizing under uncertainty vs sizing under risk

Return to Knight's distinction from M6.03 §4.6, because it governs how much to trust your own sizing math. Risk is a situation where you can put reliable probabilities on outcomes, where you know the distribution (a roulette wheel, an actuarial pool, a stable consumer-staples business whose demand you can forecast within a tight band). Uncertainty is a situation where you cannot reliably assign probabilities because the distribution itself is unknown or unknowable: a first-of-its-kind business, a company facing a brand-new regulatory regime, a technology whose adoption curve has no precedent.

Kelly and the conviction × downside heuristic both assume you are in the world of risk: they take p, u, d (or μ, σ) as known and return an optimal size. Most consequential investing, though, happens under uncertainty: your p is a genuine guess, and your d might be built on assumptions that a novel event could void. Feeding guessed inputs into a precise formula produces a precise-looking size that is false precision (the LTCM error: dressing uncertainty in the math of risk). The response is not to abandon the tools, which still rank ideas correctly, but to discount their output in proportion to how uncertain the inputs are, along two levers you already have:

• Demand a wider margin of safety. The less you can trust your probabilities, the bigger the buffer you require between price and conservative value, because the margin of safety is precisely a protection against not knowing, and you need more of it when you know less. (Klarman's Margin of Safety is an entire book on this: in a world of genuine uncertainty, the discount is the risk control.)
• Size smaller: shrink the Kelly fraction. Move from half-Kelly toward quarter-Kelly or a token size as uncertainty rises. The formula's number is a ceiling that assumes you know the odds; the more that assumption fails, the further below the ceiling you belong.

State the rule as a single sentence, because it is one of the most useful in the module: your position size should track your certainty about the distribution of outcomes, not merely the outcomes themselves. Two ideas with the identical modeled EV and identical modeled downside deserve very different sizes if you trust one model far more than the other. The well-understood staple gets a half-Kelly size; the unprecedented business, however exciting its expected value, gets a quarter or a toe-hold, because a large size is a claim to knowledge you do not have. Sizing is, in the end, an expression of epistemic humility made quantitative: you bet big only where you both have edge and know that you have it.

Common mistakes & how experts think differently

1. Confusing volatility with risk, in both directions. The amateur equates a bouncy price with danger and a calm one with safety, and so sells sound businesses in drawdowns (crystallizing loss) while feeling safe in quietly-rising frauds. The over-corrected value investor makes the opposite error: he chants "volatility isn't risk" and forgets the forced-seller exception, using leverage or investing money he'll need soon, so that volatility does become his permanent loss. The expert holds both: volatility is not risk for a patient owner bought with a margin of safety, and it is real risk the moment financing, clients, or a calendar can force a sale. So he removes the forcing mechanisms first, then ignores the quotes.

2. Reading risk off outcomes ("resulting" on risk). "It went up, so it wasn't risky" and "she's made 20% a year, so she's a great risk manager" are the same error: inferring the invisible ex-ante risk from the one visible ex-post outcome. It is how oversized bets that happened to pay get praised and repeated until they don't. The expert judges risk from the decision, the range of things that could have happened and the size that was at stake, not from the single draw that materialized; he knows a low-volatility track record can be the most dangerous thing in the room (the steamroller hasn't arrived yet).

3. Sizing by upside, at the moment of maximum enthusiasm. The natural instinct is to size big where the upside is thrilling, and to decide the size in the flush of first conviction. But Kelly and the matrix both say size is driven by the downside and by conviction, not by the size of the dream. Enthusiasm is exactly when your p is most inflated. The expert asks "what's my downside if I'm wrong?" before "what's my upside if I'm right?", and computes the size as a written, deliberate step, not a feeling formed while in love with the thesis.

4. Betting raw Kelly, or mistaking the Kelly number for a size. Raw Kelly on a favorable, capped-downside idea routinely reads 150%, 250%, more; taken literally it is a ruin machine, and even full Kelly's drawdowns are unbearable. The expert treats raw Kelly as a ceiling and a ranking signal, a statement of which ideas the math wants bigger, then bets a fraction of it (half or less) and lets a hard cap bind below that. The overbetting cliff (double-Kelly earns zero growth) means erring small is nearly free and erring large is fatal.

5. Believing name-count is diversification. "I own thirty stocks, I'm diversified" is true only if the thirty load on different factors. Thirty lenders, or thirty oil names, or thirty long-duration compounders, is one bet wearing thirty costumes, and it will reveal itself as one bet in the crisis, when correlations snap to 1. The expert sizes by factor, not by name. He writes each holding's dominant factors, caps exposure to any one, treats a correlated cluster as a single position, and asks "how many bets do I actually have?" (N_eff), not "how many tickers?"

6. Refusing to express conviction (the opposite error). Terrified of concentration, some investors equal-weight everything at 2–3%, diluting their best insight into their fortieth-best idea and guaranteeing mediocrity. That is as much a sizing failure as over-concentration, because it throws away the edge that selection earned. The expert concentrates in proportion to sized-for edge: meaningfully in the few high-conviction, capped-downside ideas, minimally or not at all in the marginal ones (the punch-card discipline), while never letting a correlated cluster breach the caps.

7. Using leverage to "enhance" a safe, high-conviction idea. The most seductive path to ruin: the idea is so good and so safe that a little borrowing will juice the return. Leverage is precisely what converts every other mistake (a missed fraud, an eroding moat, a temporary drawdown) from survivable into fatal, by removing your ability to wait. The expert refuses any leverage that can force a sale, full stop, and regards "but this one is safe" as the exact thought that precedes most blow-ups.

8. Sizing under uncertainty as if it were risk. Feeding guessed probabilities into f* = p/d − q/u yields a precise size that is false precision: the LTCM error of dressing the unknowable in the math of the known. The expert discounts the formula's output by his confidence in the inputs: a well-understood staple gets half-Kelly; an unprecedented business with the same modeled EV gets a quarter or a toe-hold, and a wider margin of safety, because a large size is a claim to knowledge, and he sizes to what he actually knows.

9. Anchoring the downside on the recent price instead of on value. "It already fell from ₹100 to ₹60, so the downside is small" treats the last price as the floor. The floor is conservative intrinsic value or asset backing, which may be well below ₹60 (value traps and frauds fall much further). The expert sets d from a bear-case valuation and verifies the floor is real (genuine assets, no hidden leverage, no fraud), because the entire justification for a larger size is a downside that is genuinely capped; an illusory floor turns a "Kelly-justified" big position into a route to zero.

Worked examples

WE1: From a thesis to a size: Kelly, fractional Kelly, and the overbetting cliff

A US industrial distributor (illustrative, modeled on a mid-cap specialty-distribution business at a post-scare discount; pull the actual filing before acting). After a demand scare, the stock trades meaningfully below your conservative value. You judge: over ~2 years, probability 0.60 it re-rates and compounds to +80% (u = 0.80), probability 0.40 the scare proves real and it falls to a well-supported floor, a −30% permanent loss (d = 0.30, set by asset backing and a bear-case DCF, a genuine floor you have verified). How much do you buy?

Step 1. Raw investing Kelly. f* = p/d − q/u = 0.60/0.30 − 0.40/0.80 = 2.000 − 0.500 = 1.500 → raw full Kelly = 150% of capital. The number is above 100% because the downside is capped at −30% while the upside is +80% with better-than-even odds. That is the Dhandho asymmetry the formula rewards. As a literal instruction ("borrow to put 150% in one name") it is absurd, which is the point: raw Kelly is a ceiling and a ranking signal, not a size.

Step 2. Read the geometric-growth curve. Compute the true long-run growth g(f) = 0.60·ln(1 + 0.80f) + 0.40·ln(1 − 0.30f) at several fractions of Kelly (f* = 1.5):

Step 3. Read the lesson in the table. Half-Kelly earns 77% of full-Kelly's growth (matching the 2c − c² rule's 75%) while cutting the size, and the drawdown pain, in half. Double-Kelly does not merely stall growth; on these real, large outcomes it drives growth sharply negative (−18.7%), a worse penalty than the quadratic approximation's "zero," because big skewed outcomes punish overbetting even harder than the textbook curve. Overbetting a positive-edge idea is how you lose money being right.

Step 4. Fractionalize, then cap. You would not bet even half-Kelly's 75% on one name. Apply the workflow: raw Kelly 150% → half-Kelly ~75% (still absurd for one position) → hard single-name cap ~10–12%. The cap binds far below the Kelly number, which is exactly right: Kelly told you this idea is a big-relative one (high conviction, capped downside, so put it near your single-name limit), and the cap kept you alive. Final size: ~10%, contingent on one verification: that the −30% floor is genuine (real assets, no hidden leverage, no accounting surprise), since the whole case for a near-cap size is the capped downside. If that floor is a hope rather than a fact, d is really large, Kelly collapses, and the position shrinks to a toe-hold.

WE2: Decomposing "diversification" into a single factor bet

An Indian portfolio (illustrative). An investor holds 24 equal-weighted names (4.17% each) and feels well diversified: no position over ~4%, spread across what he calls six sectors. Assume each name has annual volatility σ = 40%. On inspection, 15 of the 24 are lenders, private banks, NBFCs and housing-finance companies making up 62.5% of the book, all driven by one factor: the Indian credit cycle and wholesale-funding conditions. The other 9 names (37.5%) are genuinely spread across unrelated factors.

Step 1. What he thinks he owns: 24 independent bets. If the 24 names were mutually independent (ρ = 0), portfolio volatility would be σ/√n = 40%/√24 = 40%/4.90 = 8.2%. That is the diversification he feels he has.

Step 2. What the lender cluster actually is. The 15 lenders share a factor; take their pairwise correlation as ρ = 0.75 in normal times. Their effective number of independent bets: N_eff = 15 / [1 + (15−1)·0.75] = 15 / [1 + 10.5] = 15 / 11.5 = 1.30. Fifteen lender names ≈ 1.3 independent bets. The 62.5% of the portfolio sitting in lenders is, in risk terms, one leveraged bet on Indian credit, not fifteen bets, and certainly not diversification.

Step 3. The whole portfolio's effective diversification. Even if you (charitably) treated all 24 names as sharing a modest ρ = 0.5 (they don't; the lenders are far more correlated with each other), N_eff = 24/[1 + 23×0.5] = 24/12.5 = 1.9. Realistically the book is one big credit bet (~62.5%) plus a scattering of smaller ones: perhaps 2–3 effective bets, not 24. The volatility floor for the correlated lender sleeve is σ·√ρ = 40%·√0.75 = 40%×0.866 = 34.6%, and adding lender names past a handful buys essentially nothing.

Step 4. What happens in the crisis. When the credit factor turns (IL&FS, September 2018; the DHFL cascade; COVID, March 2020), the lenders' correlations snap toward 0.9+ and they fall together. Say the lenders drop 45% and the 9 others drop 15%. Portfolio loss = 0.625×45% + 0.375×15% = 28.1% + 5.6% = −33.7%. The investor who believed he held 24 uncorrelated 4% positions takes a −34% hit from what was, all along, a single 62.5% bet on Indian credit. (The identical structure sank portfolios of US regional banks in March 2023 and recurs in any book heavy with one sector: oil E&P, IT-services exporters, long-duration compounders.)

Step 5. The fix: size by factor. Impose a factor cap of ~25% on any single factor. Financials must fall from 62.5% to ≤25%, i.e., sell ~37.5% of the portfolio out of lenders and redeploy into genuinely uncorrelated factors (or hold cash). Now the credit factor can cost the portfolio at most ~11% in a 45% lender crash instead of ~28%, and the N_eff of the book rises from ~2 toward a real handful. Nothing about the names changed the risk; recognizing the factor did.

WE3: Building a sized sleeve with the conviction × downside matrix

Four candidates (illustrative; two Indian, two US) that have cleared your analysis; now size them. You run each through conviction (Kelly's p, calibrated to base rates) and downside (Kelly's d, set by the margin of safety), read the §4.8 matrix, then apply the overlay (single-name cap 12%; financials factor cap 25%; no leverage).

Read the sizing. A and B earn near-meaningful sizes because high conviction pairs with a capped, quality-backed downside. They do not earn the 12% top of the range, because the downside is medium, not small (neither is deeply discounted). C, despite a plausible thesis, gets only 2%: a cyclical with a ~55% downside cannot carry weight no matter how attractive the upside, because the asymmetry isn't there. This is the matrix refusing to let the dream size the position. D is a pass: low conviction and a large downside is the bottom-right of the matrix, the "not at any meaningful size, so not at all" cell; forcing a 1% "starter" here would just be a punch wasted on a lottery ticket. Notice the discipline: two of four candidates that passed the analysis still get a small size or none, because sizing is a second, independent gate.

Apply the factor overlay. A (Indian credit, 7%) is the only financial here, well under the 25% cap, so it is fine. But suppose you already hold 20% in Indian financials elsewhere; then A's 7% would push the factor to 27%, breaching the cap, and you would trim A to 5% (or trim the existing sleeve) to keep total Indian-credit exposure ≤25%. The factor cap, not the single-name cap, is what binds. That is exactly the hidden-correlation lesson: A is not an independent 7% bet, it is 7% more of your existing credit bet, and it must be sized against the whole factor, not in isolation.

The finished sleeve (assuming no prior financials): A 7% + B 7% + C 2% = 16% deployed across 3 genuine, low-mutual-correlation bets (Indian credit, US consumer, global steel), with D passed and the balance available for other uncorrelated ideas or cash. Three real bets, each sized to its edge and its downside, none able to cost more than ~2–2.5% of the portfolio if it goes to its worst case. That is a portfolio built to survive being wrong, which is the only kind worth building.

WE4: Drawdown, recovery, and the arithmetic of ruin (India)

Ananya (synthetic) holds a ₹50,00,000 (₹50 lakh) portfolio. One bad outcome, two sizing policies, and why in compounding the third loss is the one that ends you. All figures synthetic and Python-verified.

Step 1. The recovery asymmetry. A permanent loss of fraction L leaves you 1 − L of your capital, so the gain on what remains needed to climb back to even is L ÷ (1 − L), a convex function that runs away from you as the loss deepens:

A −20% dip asks a routine +25%; a −50% loss demands a double just to break even; a −90% loss needs a ten-bagger. Permanent loss is the enemy sizing exists to contain precisely because the hole is not linear in its own depth.

Step 2. The same −70% impairment, sized two ways. A credit shock permanently impairs NBFC-X (synthetic) by 70%. Under a concentrated policy Ananya held it at 40% of the book; under a sized policy, at 8%.

• Concentrated: portfolio hit = 0.40 × 0.70 = 28% → ₹14,00,000 gone, ₹36,00,000 left, needing 0.28 ÷ 0.72 = 38.9% on the rest to recover, a lost couple of years.
• Sized: portfolio hit = 0.08 × 0.70 = 5.6% → ₹2,80,000 gone, ₹47,20,000 left, needing 0.056 ÷ 0.944 = 5.9%, a scratch. Identical company outcome; the damage to Ananya is set entirely by the size.

Step 3. The ruin path (sequence, not average, decides survival). Losses compound multiplicatively, so a string of bad outcomes matters more than any single average. Call "ruin" a cumulative permanent drawdown to about half of capital, a level from which recovery inside a working lifetime is unrealistic. Suppose each policy keeps taking its own worst-case hit (the concentrated 40% bet repeatedly, because it is Ananya's one big conviction and credit is a factor she reloads):

The concentrated policy is nearly finished after two hits and gone after three; the sized policy needs 13 straight worst-cases (0.944^k first falls to ½ at k ≈ 12.0) to reach the same ruin line.

Step 4. The lesson. Analysis lowers the odds of any one of these hits; it does nothing about the sequence, which is the domain of sizing alone. Because recovery is convex (Step 1) and ruin is an absorbing barrier reached in very few multiplicative steps (Step 3), the number that decides whether decades of good work survive is not your hit-rate but the loss you let a single position, or a single factor, inflict on the whole. Size sets how many times you are allowed to be wrong before you are out; keep that number large.

WE5: The gambling Kelly vs the investing Kelly: a binary catalyst, correctly sized (US)

Riverbend Logistics (synthetic), a US small-cap, trades at $12 into a make-or-break refinancing that resolves in about a year. It is a genuinely binary situation, the case where the classic gambling-form Kelly, not the investing form, is the honest model. All figures synthetic and Python-verified.

Step 1. When the downside is total, use the gambling form. If the refinancing succeeds (p = 0.55) the equity roughly doubles, giving net odds b = 1, i.e. +100%; if it fails (q = 0.45) the equity is wiped and you lose the whole stake (d ≈ 1). This is the one case where f* = (b·p − q)/b applies literally:

f* = (1 × 0.55 − 0.45) ÷ 1 = 0.10 (equivalently p − q/b = 0.55 − 0.45 = 0.10) → full Kelly = 10%.

Even on a favorable double, full Kelly stakes only a tenth of capital, because a total-loss downside is ruinous to the very thing Kelly protects: the compounding path.

Step 2. A favorable double-or-nothing barely compounds. Read the true growth g(f) = 0.55·ln(1 + f) + 0.45·ln(1 − f):

• at full Kelly f = 0.10: 0.55·ln 1.10 + 0.45·ln 0.90 = 0.0524 − 0.0474 = +0.50%/yr, a positive-edge bet whose optimal growth is a rounding error.
• at double Kelly f = 0.20: 0.55·ln 1.20 + 0.45·ln 0.80 ≈ 0.00%: the overbetting cliff, confirmed on real numbers.

The total-loss structure crushes the growth rate even when sized perfectly, and any overbetting drives it straight to zero. This is why the module distrusts binary, wipe-out payoffs: the math barely pays you to take them and punishes oversizing instantly.

Step 3. Find a real floor and the size transforms. Now suppose patient work uncovers hard asset backing (owned terminals plus collectible receivables) worth about 35% of today's price, so "failure" means a −65% permanent loss, not −100%: d = 0.65, upside unchanged at u = 1.0. Switch to the investing form:

f* = p/d − q/u = 0.55 ÷ 0.65 − 0.45 ÷ 1.0 = 0.846 − 0.450 = 0.396 → ~40%.

The Kelly size quadrupled (10% → ~40%) with the odds and the upside untouched, purely because the downside fell from total to −65%. Here is the module's central claim made vivid: size is bought with a smaller downside, not a bigger dream.

Step 4. Fractionalize hard, then cap, because the floor is contested. Even 40% is absurd for one name, and a distressed refinancing is Knightian (§4.11), so quarter-Kelly rather than half: 0.25 × 0.40 ≈ 10%, right where the single-name cap sits. But the whole case for any real size rests on that 35% floor being genuine in a distressed balance sheet, which is exactly where asset values are least trustworthy. So the anti-ruin rule (§4.12) bites: an unverified floor is a hope, and you revert to the Step-1 total-loss mindset. Final size: a ~2–3% toe-hold until the floor is proven; if it cannot be, this is a 10% gambling-Kelly bet on a coin you should not be flipping with real money. The gambling form and the investing form are the same formula. Which one is honest depends entirely on whether your downside floor is real.

Practice set

Problems 1–4 are guided (framework applied with scaffolding), 5–9 independent, 10–12 exam-condition (12 minutes each, closed book). Numeric answers to the stated tolerance; currencies and companies are illustrative unless named, and synthetic figures are labeled. The required set is a Kelly/fractional-Kelly computation, a conviction × downside problem, a hidden-correlation decomposition, and a "size this given the downside" case; those are covered across P3/P4 (Kelly), P5/P10 (conviction × downside and sizing), P7/P8 (hidden correlation), and P10/P11 (size-given-downside and ruin).

P1 (guided — risk as probability × magnitude). A ₹100 position has: 76% chance of no permanent impairment; 16% chance the business deteriorates and you permanently lose 35%; 8% chance of a severe impairment (leverage spiral or fraud) costing 80%. (a) Compute the expected impairment. (b) Which scenario contributes more, and what does that teach? (c) Why would a volatility-based risk measure potentially miss all of this?

Solution. (a) Expected impairment = 0.76×0 + 0.16×0.35 + 0.08×0.80 = 0 + 0.0560 + 0.0640 = 0.120 → 12.0%. (b) The low-probability 8% tail contributes 6.4 points versus the more-likely 16% scenario's 5.6. The rarer, larger catastrophe carries more weight, because probability and magnitude multiply. The lesson: never dismiss a scenario for being unlikely without asking "unlikely times what magnitude?" (c) A volatility measure summarizes the observed price wiggle symmetrically (counting upside bounces as "risk" too) and is blind to a permanent-loss tail that has not yet shown up in prices. A quietly-compounding fraud or a hidden leverage cliff reads as low volatility right until it detonates. Volatility measures the past chart; impairment risk lives in the unrealized future (Marks).

P2 (guided — recovery asymmetry drives sizing). A position permanently loses 50% of its value. (a) If it was sized at 10% of your portfolio, what is the portfolio hit and the gain-on-the-remainder needed to get back to even? (b) If it was sized at 40%? (c) State the sizing lesson in one sentence.

Solution. (a) Portfolio hit = 10% × 50% = 5%; portfolio falls to 95, needs 5/95 = +5.3% on the rest to recover, a shrug. (b) Hit = 40% × 50% = 20%; portfolio to 80, needs 20/80 = +25% to recover, a serious wound. (c) The permanent loss a company can inflict is fixed by the business; the permanent loss to you is (size × that loss), so sizing is the one lever that fully controls how a bad outcome maps onto your survival. Same −50% event; one version is a footnote, the other a lost year, and the difference is entirely the size.

P3 (guided — Kelly, both forms). (a) A bet wins +100% (net odds b = 1) with p = 0.60 and loses the whole stake with q = 0.40; compute full Kelly f* = (b·p − q)/b. (b) An investment wins +60% (u = 0.60) with p = 0.60 and loses −25% (d = 0.25) with q = 0.40; compute f* = p/d − q/u. (c) Why is (b) so much larger than (a), and what do you actually do with it?

Solution. (a) f* = (1×0.60 − 0.40)/1 = 0.20 → 20%. (b) f* = 0.60/0.25 − 0.40/0.60 = 2.400 − 0.667 = 1.733 → 173%. (c) The only difference is the downside: (a) loses 100% on a loss, (b) loses just 25%. Capping the downside from −100% to −25% raises raw Kelly roughly nine-fold (0.20 → 1.73). That is the mathematical engine of "heads I win, tails I don't lose much." What you do: treat 173% as a ceiling/ranking signal, apply half-Kelly (~87%, still absurd for one name), then bind the hard single-name cap (~10–12%), and verify the −25% floor is genuine, because the whole case for a big size is the capped downside.

P4 (guided — continuous Kelly and the overbetting cliff). A position has expected excess return over cash μ = 12%/yr and return volatility σ = 30%/yr. Using the continuous approximation: (a) full-Kelly fraction f* = μ/σ²; (b) maximum growth g_max = μ²/(2σ²); (c) growth at half-Kelly; (d) growth at double-Kelly, and the moral.

Solution. (a) f* = 0.12 / 0.30² = 0.12/0.09 = 1.333 → 133% (levered: a ceiling, not a literal size). (b) g_max = 0.12²/(2×0.09) = 0.0144/0.18 = 0.080 → +8.0%/yr. (c) Half-Kelly: growth = g_max·(2c − c²) with c = 0.5 = 0.080 × 0.75 = +6.0%/yr (75% of the max for half the size and volatility). (d) Double-Kelly: c = 2 → 2×2 − 2² = 0 → 0%/yr. The moral: overbetting by 2× converts an 8% compounder into a 0% one. The penalty for oversizing is not a haircut, it is the entire edge, so you deliberately sit left of the peak where estimation errors are cheap.

P5 (independent — conviction × downside sizing). Using the §4.8 matrix (single-name cap 12%, financials factor cap 25%), size these three, then check the caps. (i) A: high conviction, small downside (~12%), a non-financial. (ii) B: low conviction, large downside (~55%). (iii) C: high conviction, medium downside (~30%), an Indian NBFC — and you already hold 20% in Indian financials.

Solution. (i) A → High/Small = 10–12%; take 12% (top of range, since downside is genuinely small and conviction high), the archetypal "swing hard" position. (ii) B → Low/Large = pass; low conviction and a large downside is the discard cell, so no meaningful size is justified and you take none. (iii) C → matrix says High/Medium = 6–8%, but C is an Indian financial and you already hold 20% in that factor; adding even 6% breaches the 25% factor cap (20 + 6 = 26%). The factor cap binds: trim C to ≤5% (or reduce the existing sleeve to make room). C is not an independent 6% bet. It is 6% more of a bet you already largely own, and hidden correlation means it must be sized against the whole 25% factor limit, not on its own merits.

P6 (independent — margin of safety sets the downside, which sets the size). A stock trades at ₹800. Case 1: your conservative bear-case value is ₹640 (downside −20%), base case ₹1,200 (+50%), with p = 0.65 base / q = 0.35 bear. Case 2: same stock, same base case, but a thinner margin of safety puts the bear case at ₹480 (downside −40%). (a) Raw Kelly in each case. (b) What does the comparison prove about margin of safety and sizing?

Solution. (a) Case 1: u = 0.50, d = 0.20 → f* = 0.65/0.20 − 0.35/0.50 = 3.250 − 0.700 = 2.55 → 255%. Case 2: u = 0.50, d = 0.40 → f* = 0.65/0.40 − 0.35/0.50 = 1.625 − 0.700 = 0.925 → 92.5%. (b) Nothing changed but the downside, a wider margin of safety (bear at ₹640, d = 20%) against a thinner one (bear at ₹480, d = 40%), and it cut the Kelly-justified size by nearly two-thirds (255% → 92.5%). The margin of safety and the position size are the same decision: a deeper discount both lowers your risk and earns a bigger size, while a full price both raises risk and shrinks the size the math will bear. That is why "how much?" can never be answered from the business quality alone, only from the price you paid.

P7 (independent — hidden correlation, the volatility math). A portfolio holds 16 equal-weighted names, each with volatility σ = 35%. (a) Portfolio volatility if the names were independent (ρ = 0). (b) Portfolio volatility if all pairs share ρ = 0.6. (c) The effective number of independent bets in (b), and the one-sentence interpretation.

Solution. (a) Independent: σ/√n = 35%/√16 = 35%/4 = **8.75%.** (b) Equicorrelated: Var = σ²[1/n + ρ(n−1)/n] = 0.35²[1/16 + 0.6×15/16] = 0.1225[0.0625 + 0.5625] = 0.1225 × 0.625 = 0.0766, so volatility = √0.0766 = 27.7%. (c) N_eff = n/[1+(n−1)ρ] = 16/[1 + 15×0.6] = 16/10 = **1.6.** Interpretation: 16 correlated names are worth only ~1.6 independent bets. Diversification cut volatility from 35% only to 27.7% (a 21% reduction), versus the 75% reduction (to 8.75%) that 16 independent names would have delivered; the missing benefit is the cost of the hidden correlation.

P8 (independent — factor decomposition and resize). A "20-stock diversified" book turns out to hold 12 names (60% of the portfolio) that are all IT-services exporters — one factor: US enterprise tech spending plus the USD/INR rate — and 8 genuinely unrelated names. (a) Roughly how many bets does the manager actually have? (b) Name a single event that hits 60% of the book at once. (c) To respect a 25% factor cap, how much must be cut, and to where?

Solution. (a) The 12 correlated exporters collapse to ~1 bet (high mutual correlation), so the portfolio is really ≈ 1 large bet + up to 8 smaller ones ≈ 2–3 effective bets, not 20. (b) A US enterprise-tech-spending recession (or a sharp INR appreciation, which compresses rupee-reported revenue and margins) hits all twelve exporters simultaneously: 60% of the book moving as one. (c) To get IT-services ≤25%, cut it from 60% to 25%, which means sell ~35% of the portfolio out of IT-services exporters and redeploy into genuinely uncorrelated factors (or hold cash). The names were never the risk; the shared factor was, and only factor-level sizing controls it.

P9 (independent — sizing under uncertainty vs risk). Two ideas have the identical modeled expected value (+30%) and identical modeled downside (−25% at 30% probability), so Kelly returns the same number for both. Idea A is a stable Indian consumer-staples business you have modeled with tight confidence; Idea B is a first-of-its-kind business in a brand-new regulatory regime, where your probabilities are frank guesses. Should the sizes be equal? Explain using the risk-vs-uncertainty distinction.

Solution. No. Despite identical formula outputs, B should be sized far smaller. A sits in the world of Knightian risk: the distribution is knowable, the inputs defensible, so a half-Kelly-scale size is appropriate. B sits in the world of uncertainty: the distribution itself is a guess, so the precise Kelly number is false precision (the LTCM error of dressing the unknowable in the math of the known). The correct response to B is to (i) shrink the Kelly fraction hard, to quarter-Kelly or a toe-hold, and (ii) demand a wider margin of safety, because the discount is your only protection when the odds are unknowable. Size should track your certainty about the distribution, not just the point estimates. A large size is a claim to knowledge, and you don't have it for B.

P10 (exam — 12 min — size this given the downside). Illustrative India case. A well-capitalized private bank (HDFC Bank-like) trades at ~2.5× book. Your 3-year view: base case +55% (u = 0.55) with p = 0.70; bear case a −20% permanent loss (d = 0.20, floored by strong provisioning and asset quality) with q = 0.30. Your book already holds 18% in Indian financials; single-name cap 10%, financials factor cap 25%. (a) Raw Kelly. (b) Half-Kelly and the single-name cap. (c) The factor-cap check. (d) Final size and the one fact you must verify.

Solution. (a) f* = p/d − q/u = 0.70/0.20 − 0.30/0.55 = 3.500 − 0.545 = 2.955 → ~296% (huge, because the downside is tightly capped and conviction high: a ranking signal, not a size). (b) Half-Kelly ≈ 148%, still far above any sane single-name limit, so the 10% single-name cap would bind next. (c) But the factor cap binds first: financials are already 18%, so adding a full 10% → 28% > 25% cap. You may add at most 7% here (18 + 7 = 25). (d) Final size ≈ 7%, set by the factor cap, not by Kelly or the name cap. The one fact to verify: that the −20% floor is real (GNPA/NNPA, provision coverage, no hidden restructured or promoter-linked book, M5.01–M5.02), because the entire justification for a near-limit size is the capped downside; if the floor is a Yes Bank/DHFL-style illusion, d is really large, Kelly collapses, and the position must shrink to a toe-hold.

P11 (exam — 12 min — the ruin audit). A proposed portfolio: 25% in one high-conviction small-cap with high operating leverage and net debt/EBITDA of 4×; 20% in NBFC-X; 15% in NBFC-Y; funded partly by a 15% portfolio margin loan. (a) Identify the ruin paths. (b) Fix the portfolio so it survives its realistic worst case.

Solution. (a) Three converging ruin paths. (i) The 25% leveraged name: operating leverage plus 4× financial leverage means a downturn can take it down 70–80%; 25% × 75% = a ~19% portfolio hit from one position. (ii) The two NBFCs (35% combined) are one factor bet (Indian credit/funding, §4.10); a funding shock takes both down ~50% together = a ~17.5% hit, and it is correlated with (i) in any recession, so they arrive at once. (iii) The 15% margin loan can trigger a forced sale at the bottom, crystallizing the temporary decline into permanent loss, the fatal amplifier that removes your ability to wait. Combined bad state: a −40%+ drawdown with a margin call converting it to a realized wipeout. (b) Fixes: cut the leveraged name to the single-name cap (~8–10%), because its own leverage already makes it high-risk, so it earns less, not more, room; treat NBFC-X + NBFC-Y as one credit bet and cap the pair at the ≤25% factor limit (recognizing they are ~1 bet, not two); and eliminate the margin loan entirely, since no leverage that can force a sale is the one non-negotiable rule. Now no single name, no factor, and no forced-sale mechanism can threaten survival.

P12 (exam — 12 min — Marks, ex ante vs ex post). Two managers each delivered +18%/yr for five years. Manager X did it in quality compounders at no leverage; Manager Y did it by concentrating in one leveraged theme (and selling out-of-the-money puts) that simply did not break during the window. (a) Did they take the same risk? (b) What does the identical five-year record tell you about the risk each took? (c) State Marks's principle and its consequence for how you judge risk.

Solution. (a) No. Identical outcomes, very different risk. X's returns came from durable businesses that could survive a bad year; Y's came from collecting premiums and riding a factor that had not yet detonated ("picking up pennies in front of a steamroller"). (b) The track record alone tells you almost nothing about the risk taken. Y's return series looked exactly as good as X's, right up until the steamroller, because risk is the range of things that could have happened and the odds of the bad ones, most of which did not happen and so left no trace in the realized path. A smooth, high record can be the most dangerous thing in the room. (c) Marks: "risk exists only in the future, and the future is unknowable". You cannot read risk off returns, because good outcomes come from risky bets that got lucky and bad outcomes from prudent bets that didn't. The consequence: judge risk ex ante, from the process and the downside structure, never ex post from the P&L, which is the whole reason this program grades process over outcome (M6.03 §4.8).

P13 (guided — drawdown, recovery, and the ruin path). An investor holds a ₹50,00,000 portfolio. (a) A position suffers a 60% permanent loss; what gain on the remaining capital recovers it? (b) If that position was 25% of the portfolio, what is the portfolio-level hit and the recovery gain needed on the rest? (c) Separately, the whole portfolio takes three independent −20% drawdowns in a row; what is the cumulative wealth multiplier and total drawdown? Give the ruin lesson in one line.

Solution. (a) L ÷ (1 − L) = 0.60 ÷ 0.40 = 1.50 → +150%. A −60% hole needs a two-and-a-half-bagger on the remainder, not +60%. (b) Portfolio hit = 0.25 × 0.60 = 15%; the book falls to 85, so recovery = 0.15 ÷ 0.85 = 17.6%, painful but survivable. (c) 0.80³ = 0.512 → a −48.8% cumulative drawdown (not −60%: each blow lands on a smaller base), needing +95% to recover. What keeps you alive is not being right more often but never letting size × loss grow large enough that the required recovery becomes unreachable.

P14 (guided — the gambling-form Kelly and a total-loss downside). Use f* = (b·p − q)/b, where b is net odds and a loss forfeits the whole stake. (a) A binary event pays +200% (b = 2) with p = 0.45 and wipes out with q = 0.55; compute full Kelly. (b) A different event pays +100% (b = 1) with p = 0.58; compute full Kelly. (c) You judge (a) genuinely binary and hard to handicap — what do you actually stake, and why is the number small even though the odds are favorable?

Solution. (a) f* = (0.45 × 2 − 0.55) ÷ 2 = (0.90 − 0.55) ÷ 2 = 0.175 → 17.5%, positive despite a sub-even win rate, because 2:1 odds pay for the frequent losses. (b) f* = (0.58 − 0.42) ÷ 1 = 0.16 → 16%. (c) These are ceilings on a total-loss bet, the harshest kind. Apply quarter-Kelly (the situation is Knightian, §4.11): 0.25 × 0.175 ≈ 4.4%, then a hard cap → a ~3–4% position at most. The size is small because the downside is total: f* = (b·p − q)/b is exactly the d = 1 special case of f* = p/d − q/u, and a total loss caps every gambling-form size no matter how good the odds. Uncover a real floor (d < 1) and the same odds justify far more (WE5).

P15 (independent — continuous Kelly and choosing the fraction). A position has expected excess return over cash μ = 8%/yr and volatility σ = 20%/yr. (a) Full-Kelly f* = μ/σ². (b) Maximum growth g_max = μ²/(2σ²). (c) Growth at half- and quarter-Kelly, using g = g_max(2c − c²). (d) With μ a shaky estimate, which fraction do you run, and what did you sacrifice?

Solution. (a) f* = 0.08 ÷ 0.20² = 0.08 ÷ 0.04 = 2.0 → 200% (a levered ceiling, never literal). (b) g_max = 0.08² ÷ (2 × 0.04) = 0.0064 ÷ 0.08 = 0.08 → +8.0%/yr. (c) Half-Kelly (c = 0.5): 0.08 × (1.0 − 0.25) = 6.0%; quarter-Kelly (c = 0.25): 0.08 × (0.5 − 0.0625) = 3.5%. (d) Run quarter-Kelly or less: you keep 3.5 of the 8 growth points (44%) at one-eighth of the 200% ceiling, buying heavy insurance against an overestimated edge, and it is nearly free, since overbetting is punished far worse than underbetting. (Note g_max = μ²/(2σ²) is half the squared Sharpe ratio: Sharpe = 0.08 ÷ 0.20 = 0.4, so g_max = 0.4² ÷ 2 = 0.08.)

P16 (independent — correlation-adjusted concentration, US regional banks). A "diversified" US book holds 12 regional banks (synthetic) at 4% each (48% of the portfolio); each has volatility σ = 38% and the group's pairwise correlation is ρ = 0.65. (a) The effective number of independent bets in the cluster. (b) The volatility floor σ√ρ. (c) To respect a 25% factor cap, how much of the portfolio must leave regional banks?

Solution. (a) N_eff = 12 ÷ [1 + 11 × 0.65] = 12 ÷ 8.15 = 1.47: twelve names, ~1.5 bets. (b) σ√ρ = 38% × √0.65 = 38% × 0.806 = 30.6%, so adding regional-bank names past a handful cannot pull the sleeve's volatility below ~31%, barely under a single name's 38%. (c) Cut the factor from 48% to 25% → sell ~23% of the portfolio out of regional banks, into genuinely uncorrelated factors or cash. This is the March-2023 SVB cascade priced in advance: twelve tickers that were one bet on deposit flight, held-to-maturity duration losses, and commercial real estate. Only factor-level sizing controls it.

P17 (independent — risk of ruin across sizing policies). A position's realistic worst case is a 50% permanent loss. Define "ruin" as a cumulative drawdown to 40% of starting capital (a −60% level this investor cannot recover from). (a) Under Policy A (size 30%), give the per-event portfolio multiplier and the number of independent worst-cases that reach ruin. (b) Under Policy B (size 8%). (c) What does the gap prove?

Solution. (a) Per-event loss = 0.30 × 0.50 = 15%, multiplier 0.85; ruin when 0.85^k ≤ 0.40 → k = ln 0.40 ÷ ln 0.85 = 5.6 → 6 events (0.85⁶ = 0.377). (b) Per-event loss = 0.08 × 0.50 = 4%, multiplier 0.96; 0.96^k ≤ 0.40 → k = 22.4 → 23 events (0.96²³ = 0.391). (c) Same business risk, same run of bad luck, and the sized policy survives almost four times as many worst-cases. Position size, not win rate, sets how many times you may be wrong before you are removed from the game; survival is a property of sizing, and it is the precondition for the edge ever paying (ergodicity, §4.12).

P18 (exam — 12 min — full sizing to a survivable number). Illustrative US case. Cascade Chemical (synthetic), a US specialty-chemicals maker, trades below your conservative value. Your 2-year view: base case +70% (u = 0.70, p = 0.65); bear case a −30% permanent loss (d = 0.30, floored by mid-cycle earnings power and no net debt; q = 0.35). Your book already holds 18% in the "US cyclical industrials" factor; single-name cap 10%, factor cap 25%; a broker offers a 20% portfolio margin loan to "amplify a high-conviction idea." (a) Raw Kelly. (b) Half-Kelly and the single-name cap. (c) The factor-cap check. (d) The leverage decision, the final size, and the one fact to verify.

Solution. (a) f* = p/d − q/u = 0.65 ÷ 0.30 − 0.35 ÷ 0.70 = 2.167 − 0.500 = 1.667 → ~167%, a ranking/ceiling signal (high conviction, tightly capped downside), not a size. (b) Half-Kelly ≈ 83%, still far above any sane limit, so the 10% single-name cap binds next. (c) But the factor cap binds first: industrials are already 18%, and 18% + 10% = 28% > 25%, so you may add at most 7% (18 + 7 = 25). (d) Reject the margin loan outright. No leverage that can force a sale is the one non-negotiable rule (§4.12), and "but it's high-conviction" is precisely the thought that precedes most blow-ups. Final size ≈ 7%, set by the factor cap, not by Kelly and not by the name cap. The one fact to verify: that the −30% floor is real, that mid-cycle earnings and the debt-free balance sheet actually hold in a cyclical trough, because the entire case for a near-limit size is the capped downside; if the floor softens, d rises, Kelly collapses, and the position shrinks to a toe-hold.

P19 (independent — name the risk, then pick the method). Five exposures on one book. (i) A ₹6 crore position in a listed name whose median daily traded value is ₹4 crore. (ii) A US holding whose thesis is entirely rupee-unhedged and whose earnings are in dollars. (iii) A Kelly size produced from a p you assigned by judgement in ninety seconds. (iv) An unlisted subsidiary of a holding you own, whose accounts you cannot get. (v) The office server that holds your model and your journal, backed up nowhere. For each, name the risk from §4.13's taxonomy and pick one of the five methods, with the reason in a clause.

Solution. (i) Market liquidity risk. Your position is 1.5 days of median volume, so the exit price is not the last price. Prevent and mitigate: cut the size until the exit is a fraction of one day, because liquidity risk is one of the few that a smaller position genuinely removes rather than merely shrinks. (ii) Currency risk, a subdivision of market risk. Either shift it with a forward, or retain it deliberately and say so in the memo, having decided the dollar exposure is part of the thesis rather than an accident. What is not available is leaving it unnamed. (iii) Model risk. The formula is precise and the input is a guess (§4.11). Prevent and mitigate: shrink the Kelly fraction and widen the margin of safety, both of which are what "I do not trust this input" looks like in a number. (iv) Credit and operational risk, and an information gap that is neither. Avoid, or pay the MCA21 fee and read the AOC-4 (E11.04); an exposure whose accounts you cannot obtain is not an exposure you have sized, it is one you have guessed at. (v) Operational risk. Prevent and mitigate by backing it up, and note that this is the only row where the fix is cheap, boring and absolute, which is precisely why it is the row most often left undone.

P20 (independent — the same tail, two prices). (numeric; tolerance ±0.01 per 100 of exposure on each premium.) A ₹25 crore single-name position, one year, annualised volatility 24%, dividend yield 1.0%, one-year risk-free rate 6.0%. You want the loss floored at a 15% fall, so the protection strikes at 85 against a spot of 100. The traded one-year put on those inputs is worth 2.193787 per 100. An indemnity writer offering the identical payoff prices from its own expected loss: assume a real-world expected total return of 10.5%, so a price drift of 9.5%. (a) The pure premium, meaning the present value of the expected payout. (b) The quote at a 1.35 loading, in rupees. (c) The traded put's own markup over the same pure premium. (d) Which route is cheaper here, and what the answer does and does not generalise to.

Solution. (a) Under a 9.5% drift the probability of finishing below 85 is 17.0296%, the expected payout is discounted back to a pure premium of 1.573909 per 100. (b) 1.573909 × 1.35 = 2.124778 per 100, so ₹53,11,944 on ₹25 crore. (c) 2.193787 ÷ 1.573909 = 1.393846, a 39.3846% markup inside the traded price. (d) The insurance quote is ₹1,72,524 cheaper than the traded put, because the writer's 35% loading is smaller than the 39.38% risk premium the option market is charging, and that ranking does not generalise at all. Which of the two is cheaper is a comparison of two markups on the same expected loss, and both markups move: the option's with implied volatility, the insurer's with its own capital cost and claims experience. What does generalise is the shape of the answer. Both routes charge more than the expected loss, neither is free, and the analyst's job is to price the expected loss independently so that a quote can be recognised as expensive.

P21 (exam — 12 min — fat tails and the size you thought was safe). (numeric; tolerance ±0.02 percentage points on each loss cut, ±0.5 percentage points on the shrink.) Daily returns on a position come from a two-regime mixture. On 85% of days the standard deviation is 0.80%, on the other 15% it is 2.40%, and both regimes are mean-zero normal. (a) The mixture's daily standard deviation and its excess kurtosis. (b) The 99% one-day loss cut computed the normal way from that standard deviation. (c) The mixture's true 99% cut. (d) By how much the position must shrink to keep the same rupee loss at that level, and how many 99% breaches a normal-sized position actually takes in 250 trading days.

Solution. (a) Variance = 0.85(0.0080²) + 0.15(0.0240²) = 0.00014080, so the daily standard deviation is 1.186592% and the annualised figure at 252 days is 18.83%. The fourth moment is 3[0.85(0.0080⁴) + 0.15(0.0240⁴)], giving an excess kurtosis of 5.057851 against a normal's zero. (b) 2.3263479 × 1.186592% = 2.760425%. (c) Solving the mixture's own tail to 1% gives 3.602958%, which is 1.305218 times the normal cut. (d) To hold the same rupee loss at the true cut the position must shrink by 1 − 2.760425 ÷ 3.602958 = 23.3845%, close to a quarter. And the normal-sized position is not merely a little wrong: its own 2.760425% cut is breached with probability 1.899310%, so across 250 trading days it takes 4.75 breaches where it promised 2.5. The single standard deviation is identical in both models. Everything that differs is in the fourth moment, which is why a downside floor d estimated from a volatility number, and then fed to Kelly, is a floor with a hole in it. QM1.04 owns the moment arithmetic; §4.11's response, a wider margin of safety and a smaller fraction, is the sizing answer.

Applied mini-project: size one real, live position end-to-end

This project produces a durable artifact, a one-page sizing memo, that plugs straight into your decision journal (M0.04) and your written process (M9.04). You will take a real company you can get good data on and carry it all the way from analysis to a defensible position size, which is the step most investors skip. Pick a name inside your circle of competence, Indian or US, either from the starter bank (India: HDFC Bank, Bajaj Finance, Asian Paints, Titan, Divi's; US: Costco, Visa, Moody's, Copart, Deere) or another you have already studied. Use real filings for the numbers: SEC EDGAR (10-K, latest 10-Q) for a US name; the company's annual report + investor presentation and screener.in for an Indian one. Label every estimate as illustrative and show your work.

Produce, in order:

1. The downside d, from a bear-case value and not the recent price. Build a conservative intrinsic value: a bear-case reverse-DCF or a simple asset/earnings-power floor (M3.05, M3.10). State the permanent-loss fraction from today's price to that floor, and then, critically, name what the floor rests on (asset backing, through-cycle earnings, net cash) and one thing that could void it. This is the single most important number in the memo.
2. The upside u and conviction p. State your base-case value and the implied upside, and your honest probability the thesis is broadly right, calibrated against a base rate (M6.03 §4.6), not your enthusiasm. Justify p in two sentences of evidence, and note whether this is a risk estimate (trustworthy distribution) or an uncertainty estimate (a guess), per §4.11.
3. Raw Kelly, fractional Kelly, and the matrix. Compute f* = p/d − q/u. Independently, place the idea in the conviction × downside matrix (§4.8). The two should roughly agree on direction; reconcile them and state a proposed size.
4. The factor and ruin overlay. Name the position's two or three dominant factors. State your existing exposure to each and check the proposed size against your single-name cap and your ~25% factor cap. Run the ruin pre-mortem: if this and everything correlated with it hit their realistic worst case at once, do I survive?
5. The final size and the one verification. State the final position size (often smaller than steps 3 suggested, because a cap binds), and the single fact you must verify before buying: usually that the downside floor is real.

Deliverable: a one-page memo with the five items above, every number sourced or labeled illustrative, ending in a single sentence: "I will own X% of [company], because my downside is d, my conviction is p, and the factor/ruin caps allow no more."

Scoring rubric (0/1/2 per criterion; 12 possible; ≥9 to pass):

Case Lab: run case CL-9.1 (Six Positions, One Hidden Factor) in the app's Cases tab for this module: six holdings that look like six independent bets until you decompose their factors, compute the book's true N_eff, and resize to a survivable set, defending each final size against the single-name and factor caps.

Reading & resources

• Howard Marks, The Most Important Thing (Illuminated ed., 2013). Chapters 5 ("Understanding Risk"), 6 ("Recognizing Risk"), and 7 ("Controlling Risk") are the single best treatment of the practitioner's definition of risk anywhere; read them slowly. Pair with the Oaktree memos (oaktreecapital.com/insights, free): "Risk" (2006), "Risk Revisited" (2014), "Risk Revisited Again" (2015), and "You Bet!" (2020, on decision quality vs luck). The founding texts for everything here. [Book Paid / Memos Free] [Beginner→Intermediate]
• Edward O. Thorp, A Man for All Markets (2017) and his paper "The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market" (2006, free PDF): the practitioner who actually ran Kelly sizing to a 20-year, near-lossless record; the paper is the rigorous source for fractional Kelly and its drawdown behavior. [Book Paid / Paper Free] [Intermediate→Advanced]
• William Poundstone, Fortune's Formula (2005). The narrative history and intuition of the Kelly criterion (Kelly, Shannon, Thorp), and why full Kelly is too much; the most enjoyable way to internalize sizing by edge. [Paid] [Beginner→Intermediate]
• Warren Buffett & Charlie Munger, the Berkshire Hathaway letters (berkshirehathaway.com, free) and Lawrence Cunningham, The Essays of Warren Buffett: for the "twenty punches," "diversification is protection against ignorance," "Rule No. 1: never lose money," and the risk-as-permanent-loss framing in the master's own words. Read the 1993 letter (on risk and diversification) and the 1997/2008 letters. [Letters Free / Essays Paid] [Beginner→Intermediate]
• Mohnish Pabrai, The Dhandho Investor (2007): chapters on "few bets, big bets, infrequent bets" and "heads I win, tails I don't lose much"; the clearest popular statement of the capped-downside asymmetry that Kelly formalizes. [Paid] [Beginner]
• Seth Klarman, Margin of Safety (1991). Out of print but findable as a PDF; the definitive book-length case that in a world of genuine uncertainty the discount is the risk control, and that sizing and margin of safety are one discipline. [Paid/Scarce] [Advanced]
• Benjamin Graham, The Intelligent Investor (Zweig ed.): chapter 20 ("Margin of Safety as the Central Concept of Investment") and chapter 8 (Mr. Market); the origin of the margin-of-safety idea you apply here to sizing. [Paid] [Beginner→Intermediate]
• Lee Freeman-Shor, The Art of Execution (2015). Data from real portfolios on how position management (adding, holding, cutting) decided returns even when the picks were the same; the empirical case that sizing and execution are half the game. Read now for the sizing lens; revisited in M9.03 for selling. [Paid] [Beginner]
• Nassim Taleb, Skin in the Game (2018), chapter on ergodicity ("The Logic of Risk Taking"): the sharpest statement of why avoiding ruin is a first-order, non-negotiable goal and why the ensemble average misleads the individual investor. Pair with Ole Peters, "The ergodicity problem in economics" (Nature Physics, 2019, free) for the rigorous version. [Taleb Paid / Peters Free] [Intermediate→Advanced]
• Peter Bernstein, Against the Gods: The Remarkable Story of Risk (1996). The intellectual history of risk and probability from the Greeks to modern portfolio theory; the best single book for understanding why the academic (volatility) and practitioner (loss) definitions diverged. [Paid] [Intermediate]
• CFA Program Level III, Portfolio Management readings (risk management, position sizing, behavioral portfolio construction): the institutional, standards-based treatment; useful for the vocabulary and the risk-budgeting framing. [Paid] [Advanced]
• India ties. Sanjay Bakshi's BFBV lecture notes and Fundoo Professor blog (free) are outstanding on margin of safety, position sizing, and the Indian promoter/governance risks that make downside floors illusory (the Yes Bank/DHFL/IL&FS lessons). Marcellus Investment Managers' newsletters (free) cover concentration, quality, and hidden correlation in Indian financials specifically. [Free] [Intermediate]
• Michael Mauboussin, "Size Matters" and the Counterpoint Global reports on position sizing and the Kelly criterion (free PDFs): a clean, rigorous bridge between the theory here and real portfolio construction. [Free] [Intermediate→Advanced]
• Going deeper: Victor Haghani & James White, The Missing Billionaires (2023), with the free Elm Wealth essays and the "$300 coin-flip" experiment (elmwealth.com). The sharpest modern demonstration that sizing, not selection, is where investors quietly bleed, and a hands-on treatment of fractional Kelly, expected-utility sizing, and surviving drawdowns; Haghani co-founded LTCM, which lends the chapters on leverage and ruin unusual authority. [Essays Free / Book Paid] [Intermediate]
• Going deeper: Attilio Meucci, "Managing Diversification" (Risk, 2009; free on SSRN). The rigorous portfolio-math generalization of §4.10's N_eff: it derives the effective number of bets from the entropy of principal-component variances, turning "how many bets do I really have?" into a computable statistic for a whole book. [Free] [Advanced]

The Modern Analyst's Addendum

Everything above teaches this skill from first principles, by hand — that is how you learn it, and the mastery check still tests it that way. This addendum shows how a working analyst amplifies the same skill today. It adds; it never replaces. (R1/R10)

AI-Augment this skill

{
  "skill": "Risk as probability x magnitude of permanent impairment, the Kelly criterion in both forms, fractional Kelly and the overbetting cliff, conviction x downside sizing, hidden correlation and N_eff, and the anti-ruin overlay",
  "use": "Sizing is arithmetic on top of two judgements, and AI is useful on exactly one of them. (1) DOWNSIDE HUNTING, which is where the whole size comes from: give it the 10-K or the annual report and ask it to inventory what would actually be left in a bear case — tangible asset backing, net cash, off-balance-sheet and contingent liabilities, related-party exposure, pledged promoter shares, debt maturity walls — and to list, per item, the disclosure that would falsify the floor. Section 4.12 says an unverified floor is a hope; this is the fastest way to turn it into a checklist. (2) FACTOR ELICITATION: paste your holdings and ask for each name's two or three dominant macro/sector/style drivers and which holdings share them. It is good at spotting that a specialty chemicals name and an auto ancillary are both a China-demand bet; you then confirm it against realised correlations rather than believing it. (3) PRE-MORTEM at the portfolio level: 'name the single event that takes down the largest fraction of this book at once.' (4) Adversarial review of your own sizing memo: ask it to argue that your conviction p is inflated and your downside d is optimistic, and make it cite the specific sentence in your memo it is attacking.",
  "tools": ["Chat assistants — Claude, ChatGPT for downside inventories, factor elicitation and the pre-mortem", "Document/PDF AI — for pulling contingent liabilities, pledge disclosures and maturity walls out of a 200-page annual report", "Coding assistants — for the Kelly engine, the estimation-error simulation and the drawdown paths below"],
  "prompt": "Here is the FY annual report / 10-K for <COMPANY>: <ATTACH>. Build a bear-case downside inventory: (a) what tangible value plausibly remains per share in a severe but non-fraud downside, itemised, with the page or note each item comes from; (b) every contingent liability, guarantee, related-party exposure, pledged-promoter-share disclosure and debt maturity inside 24 months; (c) for each item in (a), the single disclosure that would prove it is NOT there. Do not give me a target price, a probability, or a position size — I am computing those. Separately, for this holdings list <PASTE>, name each position's two or three dominant factors and group the ones that share a factor.",
  "verify": "The floor is the number the size is bought with, so it is the number to distrust. Every item in the AI's downside inventory has to be traced to the filing yourself — page, note, and the number as reported — because a fabricated asset line is a fabricated position size, and the failure mode here is silent: a floor that is 10 points too generous does not look wrong, it just makes the position bigger. Do not let it estimate p, u or d; it will produce a fluent, confident, unsourced probability and you will anchor on it. The factor groupings are a hypothesis, not a measurement: confirm them against realised pairwise correlations on daily returns over both a calm window and a stress window (Mar 2020, Sep 2018 for Indian credit, Mar 2023 for US regionals), because the correlation that matters is the stressed one and it is not the one the calm sample reports. And re-derive the arithmetic yourself: language models mis-handle f* = p/d - q/u often enough that a Kelly number that is not hand-checked is not a Kelly number.",
  "diy": "The gate is unaided: you compute both Kelly forms on paper, read the growth curve g(c) = g_max(2c - c^2) off memory, place an idea in the conviction x downside matrix, derive the correlation floor sigma*sqrt(rho) and N_eff = n/[1+(n-1)rho], and run the ruin arithmetic L/(1-L). No tool is in that room, and none of this callout is admissible in it."
}

Modern Data Analysis

By hand first. You derived Kelly from the log-growth objective, computed both forms, read the overbetting cliff off g(c) = g_max(2c − c²), decomposed a 24-name book into 1.3 effective bets, and worked the ruin arithmetic to the number of consecutive worst cases each sizing policy survives. Those stay in your hands; a size you cannot compute on paper is a size you will not defend under pressure. What code adds is the part the paper method structurally cannot reach: paper computes the optimal fraction given the inputs, and code computes what happens when the inputs are estimates. That is the situation you are always in, and the situation in which most real damage is done.

Today's workflow. Write the growth rate as a function, not a table, and then run three things through it that a table cannot hold.

• The exact log-growth curve rather than its quadratic approximation. g(c·f*) = g_max(2c − c²) is a second-order expansion, and §4.6 is explicit that it is one. On real, large, skewed outcomes the true curve is not symmetric about the peak, and the difference is entirely on the dangerous side.
• A sampling distribution on the inputs. Your p is an estimate with a standard error. Draw p̂, size on f*(p̂), and score the result against the true p. This converts "fractional Kelly is prudent" from a maxim into a computed multiplier.
• Paths, not periods. Ergodicity (§4.12) is a statement about a sequence, and a sequence is what a simulation produces. The ensemble mean and the median path of the same bet are computable side by side, and their divergence is the entire argument.

Tools & sources (IN + US). numpy for the vectorised paths, scipy.optimize.brentq for inverting the growth curve, pandas for the return panel that gives you σ and the correlation matrix, matplotlib for the growth curve and the drawdown fan. India. NSE and BSE daily bhavcopy (or the Nifty sectoral index series: Nifty Bank, Nifty Financial Services, Nifty IT, Nifty Metal) for the return histories behind σ and ρ; the consolidated annual report and screener.in for the asset backing, pledged-promoter-share disclosure and debt maturity wall that decide whether the downside floor d is real; RBI's Database on Indian Economy (DBIE) and the half-yearly Financial Stability Report for the credit-cycle and NBFC-funding conditions that are the shared factor in §4.10's lender example; SEBI's disclosure of promoter pledges for the specifically Indian way a floor evaporates. US. EDGAR XBRL companyfacts for the balance-sheet items behind the floor (Assets, Cash…, LongTermDebtCurrent); FRED for DGS10, the risk-free rate the excess return μ is measured over, plus BAMLH0A0HYM2 (ICE BofA US High Yield OAS) and NFCI (Chicago Fed National Financial Conditions Index) for the stress windows in which you should re-measure correlations; the Ken French Data Library (Dartmouth, free) for factor return series, which is how you test whether your names share a factor instead of asserting it; Mauboussin's "Size Matters" (Counterpoint Global, free) for the institutional treatment of Kelly-based sizing.

# Kelly with ESTIMATED inputs — the case the paper method cannot reach. R3, recomputed in-session.
import numpy as np
rng = np.random.default_rng(9101)
p, u, d = 0.60, 0.80, 0.30                       # WE1's outcomes
fstar = p/d - (1-p)/u                            # 1.500 — matches WE1 exactly
g = lambda f, pp=p: pp*np.log(1+f*u) + (1-pp)*np.log(1-f*d)

# 1. the exact curve is NOT symmetric about the peak
for c in (0.50, 1.50, 1.75, 2.00):
    print(f"c={c}: exact {g(c*fstar)/g(fstar):+.4f} of max, quadratic {2*c-c*c:+.4f}")
# 0.50: +0.7696 / +0.7500   1.50: +0.7190 / +0.7500
# 1.75: +0.2536 / +0.4375   2.00: -0.7984 / +0.0000   <- growth already negative

# 2. size on an ESTIMATE: p_hat from n comparable past decisions
n  = 30; se = np.sqrt(p*(1-p)/n)                 # 0.0894 -> SE(f*) = 0.410 = 27.3% of f*
ph = np.clip(rng.normal(p, se, 400_000), .01, .99)
fh = np.clip(ph/d - (1-ph)/u, 0, None)
for c in (0.25, 0.50, 0.75, 0.91, 1.00):
    f = c*fh; ok = f*d < 1
    e = np.where(ok, p*np.log(1+np.where(ok,f,0)*u) + (1-p)*np.log(1-np.where(ok,f,0)*d),
                 -np.inf).mean()
    print(f"c={c}: E[true g] = {e:.5f}")         # peaks at c ~ 0.91, not at 1.00

# 3. ergodicity: the ensemble mean and the life you live, on the same bet
f = 2*fstar                                      # double Kelly
print(p*(1+f*u) + (1-p)*(1-f*d) - 1, np.exp(g(f))-1)   # +1.0800 vs -0.1704
w = np.ones(200_000)
for _ in range(100):
    w *= np.where(rng.random(200_000) < p, 1+f*u, 1-f*d)
print(f"{w.mean():.3e} {np.median(w):.3e} {(w<1).mean():.4f}")  # 6.3e+18  7.7e-09  0.8707

Verify. Prove the engine against the module's own worked figures before it goes near a real position. WE1: f* = p/d − q/u = 0.60/0.30 − 0.40/0.80 = 1.500 exactly, with per-period growth of +23.394% at full Kelly, +18.005% at half (77.0% of full, against the 2c − c² rule's 75%), +10.968% at quarter, and −18.677% at double. Every figure in WE1's table is reproduced. WE5: the gambling form gives f* = 0.55 − 0.45 = 0.100 with growth of +0.501%/yr, and −0.014% at double Kelly, i.e. the module's "≈ 0.00%" confirmed to the fourth decimal; uncovering a 35% floor moves the investing form to 0.55/0.65 − 0.45/1.00 = 0.3962, the module's "~40%". P4 gives f* = 1.333 and g_max = 8.0%; P15 gives f* = 2.0, g_max = 8.0%, half-Kelly 6.0% and quarter-Kelly 3.5%. All match. That agreement is what licenses the results below, which the module does not compute.

Quantitative lens

Everything here is a property of the module's own WE1 outcome set (p = 0.60, u = +80%, d = −30%) under the stated draws. It is a statement about that bet and that engine, not a measurement of any market.

The penalty function is asymmetric, and the module's own approximation hides it. §4.6's table is built on g(c·f*) = g_max(2c − c²), which is symmetric about the peak by construction: it says quarter-Kelly and 1.75×-Kelly both keep 43.75% of the growth. On WE1's actual outcomes they do not. Half-Kelly keeps 76.96% where the approximation promises 75.00%; 1.5× Kelly keeps 71.90% where it also promises 75.00%; 1.75× Kelly keeps 25.36% against a promised 43.75%, an 18.4-point overstatement; and double Kelly does not merely stall at zero, it delivers −79.84% of maximum growth, which is to say a per-period compounding rate of −18.68% on a bet with a genuine positive edge. Growth actually crosses zero at c = 1.834, not 2.000, and at c = 2.222 the size is f = 1/d = 3.33, where a single loss takes the whole account, which is not a poor outcome but an absorbing one. Measure the asymmetry directly: a sizing error of ±50% of f* costs 1.22× as much growth on the high side as on the low; at ±75% it costs 1.41×. The approximation is a good local guide and a bad global one, and it errs in the flattering direction exactly where the error is expensive. §4.6 says the curve "falls symmetrically after" the peak; that is true of the parabola and false of the bet.

Ergodicity is not a philosophical aside, it is the difference between two numbers on the same line of code. At double Kelly (f = 3.00) the ensemble average (the mean outcome across many parallel lives, the number an expected-value calculation returns) is +108.00% per period. The time-average growth rate, the one the single sequence you actually live through delivers, is −17.04% per period. Simulate 200,000 paths for 100 periods and both are visible at once: mean terminal wealth 6.30 × 10¹⁸, median terminal wealth 7.74 × 10⁻⁹, with 87.07% of paths finishing below their starting capital and 76.17% finishing below one per cent of it. The mean is real; it is carried by a vanishing minority of paths and describes nobody's life. Run the same 100 periods at full Kelly and the median finishes at 1.45 × 10¹⁰ with 0.05% of paths below their start; at half Kelly the median is 6.60 × 10⁷ with no path below its start. This is why the module insists that survival is first-order: an ensemble average can be enormous and positive while the modal life is ruin, and the arithmetic mean is silent about which of the two you are in.

Sizing on an estimated edge, and an honest number for how far it justifies fractionalising. Suppose your p = 0.60 rests on thirty comparable past decisions. The binomial standard error is 8.94 percentage points. Because df*/dp = 1/d + 1/u = 4.583, that becomes a standard error on the Kelly fraction of 0.410 (27.3% of the fraction itself), so f* is 1.50 with a 10th-to-90th-percentile range of 0.98 to 2.03. Now size at c × f*(p̂) and score against the true p: expected growth peaks at c ≈ 0.91, not at 1.00. Full Kelly on an estimate earns 21.61%/period against 21.83% at the optimum and 23.39% under perfect information. Push the estimate to ten comparables (SE 15.49pp) and the optimum falls to c ≈ 0.79; at a hundred (SE 4.90pp) it rises to c ≈ 0.97. Here is the honest part, and it cuts against the easy reading: estimation error in p alone does not get you to half-Kelly. It gets you to roughly 0.8–0.9. The rest of §4.7's case has to be carried by its other three reasons: drawdown tolerance, correlation across positions, and Knightian inputs where no standard error exists. It is carried. At these outcomes the median deepest drawdown over ten periods is 69.8% at full Kelly against 39.9% at half (a ratio of 0.57, so §4.7's "roughly half as deep" is right at a career-length horizon), and over 100 periods it is 95.0% against 72.0%, the ratio drifting to 0.76 as the horizon lengthens. Half-Kelly is a rational response to parameter uncertainty plus a human being's capacity to hold a position; the first alone buys about half of it.

And the input that most deserves your uncertainty budget is d, not p. At WE1's point, df*/dd = −p/d² = −6.667 against df*/dp = 4.583 and df*/du = q/u² = 0.625. A five-point error in the downside floor moves the Kelly fraction by 0.333; the same five-point error in conviction moves it by 0.229; in the upside, by 0.031. Walk d alone: 0.25 → f* = 1.900, 0.30 → 1.500, 0.35 → 1.214, 0.40 → 1.000, 0.50 → 0.700. The upside is nearly irrelevant to the size and the floor is nearly everything. That is §4.4, §4.8 and WE5's "size is bought with a smaller downside, not a bigger dream" as a derivative rather than a slogan, and it tells you where to spend the analysis: on proving the floor, not on refining the target price.

Honest limits, and they are not decoration. Every figure above is computed from a two-outcome bet with fixed, known-shape outcomes; a real position has a continuous, skewed, fat-tailed distribution whose left tail is precisely the region a two-point model cannot represent, and the direction of that omission is always flattering. The estimation-error result treats p̂ as unbiased and normally distributed around the truth, which is the kind case: the failure §4.7 actually names is systematic overconfidence, a biased estimator, and this simulation says nothing about it. Feed it a true p of 0.55 while you believe 0.60 and full Kelly earns 16.46% against half-Kelly's 14.38%, while at a true p of 0.50 half-Kelly finally wins outright (10.76% against 9.53%). The c* ≈ 0.91 figure is specific to these outcomes and to an SE derived from thirty comparables; it is a method to re-run with your own numbers, not a constant to quote. Nothing here models correlation across positions, so it cannot be used to size a book. §4.10's factor caps still bind and are still the thing that matters most. And the drawdown percentiles are properties of an i.i.d. sequence of this one bet; real drawdowns arrive with the correlated cluster attached, which is worse than anything above.

The paragraph above admits that a two-point bet cannot represent a fat left tail, and the fourth moment is what that admission is worth in numbers. Excess kurtosis measures how much of a distribution's weight sits in its peak and its tails rather than in its shoulders, and a normal scores exactly zero; QM1.04 owns the estimator and the sampling problems that come with it. Start with the bar a normal sets, because it is stricter than almost anyone guesses. A four-standard-deviation down day has a one-sided probability of 3.167124 × 10⁻⁵, which is one such day every 31,574 trading days, or once in 125.3 years. A five-sigma day is one every 3,488,556 trading days, or once in 13,843.5 years. Those are the numbers a risk report is implicitly asserting whenever it computes a loss cut by multiplying a standard deviation by a z-score.

Run the estimator on a series and the contrast becomes concrete. The synthetic broad-market index files shipped with this course are seeded geometric random walks of 499 daily returns each, and they behave exactly as a normal-tailed generator must. The India file returns an annualised standard deviation of 19.6657%, a sample skewness of +0.107583 and a sample excess kurtosis of +0.002347; the US file returns 17.3143%, +0.172226 and +0.102733. Both contain one absolute move past three sigma, against a normal expectation of 1.347198 in 499 days, and neither contains a single four-sigma or five-sigma day. That is not a defect in the files. It is the point: a random walk has no fat tails, so the shipped series is the control case, and every four-sigma day you have ever read about in a real index is evidence that the generator producing market returns is not the one producing these files. Do not treat the kurtosis figure above as a measurement of any market. Pull a long daily series for an index you actually hold, run the same four lines, and record what you get.

What a fat-tailed generator does to the same standard deviation is easiest to see on a mixture, because its moments can be written down exactly. Take the two-regime model QM1.04 works: on 90% of days the daily standard deviation is 0.70%, on the other 10% it is 2.60%, both regimes mean-zero normal. The blend's daily standard deviation is 1.056882%, an unremarkable 16.7775% annualised, and its excess kurtosis is 8.507315. Now count the tails against a normal with the identical standard deviation. Four-sigma days arrive 164.12 times more often, one every 96.2 trading days against the normal's one every 15,787. Five-sigma days arrive 7,344.55 times more often, one every 237.5 trading days against one every 1,744,278. The volatility number is honest and the tail count built from it is off by three orders of magnitude, which is how an institution ends up explaining a "twenty-five-sigma event" instead of admitting its distribution was wrong.

Then the sizing consequence, which is this module's own business rather than QM1.04's. Under the normal the one-day 99.9% loss cut on that mixture's standard deviation is 3.266012%. Under the mixture itself it is 6.048504%, or 1.851954 times as far out. Two things follow directly. A position sized so that its worst 99.9% day costs a fixed number of rupees must shrink by 46.0030%, close to half, once the true tail is used. And a position left at the normal-tailed size is not slightly optimistic about its own limit: its 3.266012% cut is breached with probability 1.045431%, so over 250 trading days it delivers 2.61 breaches where it promised 0.25. The Kelly arithmetic above takes d, the downside floor, as an input. Estimate that floor from a volatility number on a fat-tailed series and d is understated, f* is overstated, and the overbetting cliff of §4.6 is closer than the table says. This is the fourth moment's whole message to a sizer: it does not change the average outcome, it changes how far the bad one goes, and sizing is the only tool that controls that.

Do it in code: write g(f) once, check it against WE1's table, then put a sampling distribution on p and find your growth-maximising multiplier rather than quoting a rule of thumb; simulate the paths and report the median alongside the mean; and run the df*/dd sensitivity so you can see how many hours of the downside work are worth doing. Then compute the sample excess kurtosis and the three-, four- and five-sigma day counts on a real index series, before you trust any loss cut built from a standard deviation. See G1/QM1.01, G1/QM1.03, G1/QM1.04 and G2/QD2.02.

Where this goes next: galaxy cross-links

• G1 · QM1.01 Probability, Distributions & Estimation covers where the standard error on p̂ comes from, why a derived quantity like f* inherits it amplified by df*/dp, and the difference between an unbiased estimate and a confident one.
• G1 · QM1.03 Time Series, Monte Carlo & ML: path simulation, the mean-versus-median divergence of a multiplicative process, and how many paths you need before a tail statistic like "76% finish below 1% of start" is stable.
• G2 · QD2.02 Risk Systems, VaR & Expected Shortfall handles drawdown, tail loss and the institutional machinery for measuring what §4.12 asks you to survive, including why VaR is the wrong statistic for an absorbing barrier.
• G2 · QD1.04 Portfolio Optimization in Code: Kelly generalised from one position to a book with a covariance matrix, which is what §4.10's factor caps are approximating by hand.
• G1 · AA1.02 Black-Litterman & Risk Budgeting expresses the factor cap as a formal risk budget, so "no more than 25% in one factor" becomes a constraint the optimiser cannot violate rather than a rule you remember.

Teach it back & journal

Feynman prompt. Write a one-page explainer for a smart 15-year-old titled "Why owning thirty stocks can still be one bet — and why you must never bet the farm." Convey, without jargon: that "diversified" is about owning things that don't move together, so thirty companies that all rise and fall with the same one thing (say, whether banks in the country are healthy) are really just one big bet in disguise, and it will show up as one bet on the worst possible day when they all fall at once; and that you must never put so much on any single thing that being wrong could wipe you out, because losing everything is the one mistake you can't recover from. If you lose half your money you need to double what's left just to get even, and if you lose all of it, no amount of being clever afterward brings it back. If your page needs the words "correlation," "Kelly," or "ergodicity," you don't own it yet. Bonus: explain in one sentence why a person who made great returns for five years might have been taking terrible risks the whole time.

Journal prompt. Three entries, written before you look anything up. (1) Size a real position. Take one company you own or are seriously considering and write its sizing memo in five lines: your bear-case downside d (and what that floor rests on), your conviction p (calibrated against a base rate, not your hopes), its two or three dominant factors, and the size the conviction × downside matrix and your caps allow. Was it bigger or smaller than your gut wanted? (2) Count your real bets. List everything you own (or a paper portfolio) and tag each holding with its dominant factor. How many factors dominate? What is your honest N_eff: three bets? Five? If any single factor is above ~25% of the book, name it and the change you'd make. (3) A sizing autopsy. Recall one time you sized a position by enthusiasm, big because the upside thrilled you. Reconstruct it through this module's lens: what was the real downside, and what would conviction × downside (and the ruin overlay) have told you to do instead? Commit the lesson as a line in your checklist, so the next time excitement and size arrive together, the written rule is already there to overrule the feeling.

This module's flashcards and mastery quiz are wired into the app: see the node's Quiz and Reviews.