Portfolio Construction

Phase 9 · Week 71 · ~20 hours

Where it sits. A book of thirty hand-picked stocks can secretly be a single bet on one factor. No single-position analysis can see that, which is why the work you spent sixty-five weeks learning to do on one business now turns into a decision about the whole pool of your capital. This is third-from-last of the process phase. M9.01 sized a single position: it answered "given this idea, how much?" with risk-as-permanent-impairment, conviction × downside, and fractional Kelly. What follows answers the questions that only appear once you own more than one thing. How many things should I own, and which things earn a place against the alternatives? The concentration-versus-diversification debate that M6.02 §4.9 set up philosophically turns quantitative here, with the actual mathematics of how correlation drives portfolio variance, where the benefit of the next name goes to zero, and what "diworsification" costs in numbers. Two disciplines then make a portfolio a system rather than a pile. Hurdle rates and opportunity cost demand that every rupee beat your best available alternative, including your own existing worst holding; cash as a position is the residual of discipline, not a failure of nerve. Next comes an honest, evenhanded tour of factor investing, the documented tilts (value, quality, momentum, size, low-volatility) that may already explain much of what you think is stock-picking, and of the active-versus-passive / efficient-markets debate, presented so you can judge for yourself whether your own edge is real enough to justify the game you are about to play. M9.03 will then teach you when to sell; M9.04 wires all of it into a written, repeatable process.

Suggested hour budget. Core content: 8h · worked examples (portfolio variance, the replace-that decision, factor-tilt diagnosis): 3h · practice set: 4.5h · applied mini-project (your Portfolio Construction Policy): 2.5h · flashcards + mastery check: 2h.

App link. The work here feeds the Portfolio & Risk branch: the portfolio-variance sandbox (enter weights, vols, and a correlation matrix; watch the diversification benefit and the systematic floor emerge, wired to the M9.01 sizing engine); the opportunity-cost "replace-that" challenge (a fully-invested book plus a new idea, graded on the after-friction decision); the factor-tilt diagnostic (tag a sample portfolio's holdings and read out its implicit value/quality/momentum/size/low-vol loadings); and the steelman drill (argue active and passive to a rubric). Its output, your written concentration, hurdle and cash policy, becomes the constitution the portfolio simulator scores you against for the rest of the program.

Learning objectives

By the end you can:

1. Compute portfolio variance from a correlation matrix for a real multi-asset book, decompose total risk into its diversifiable (idiosyncratic) and undiversifiable (systematic) parts, and show numerically why the risk floor is the average covariance, not zero.
2. Locate where the marginal benefit of the Nth position goes to zero using the σ√ρ̄ floor result, and quantify "diworsification": the point where each added name cuts expected return more than it cuts risk.
3. Operationalize the concentration-versus-diversification choice: run the return-dilution math for the concentration case, respect the Kelly-and-ruin caps on it, count independent bets rather than tickers, and set your own number-of-positions policy with stated reasons tied to your edge, temperament, and survival constraint.
4. Run the opportunity-cost / hurdle-rate discipline as Munger does it, setting an absolute hurdle and treating your existing worst holding as the bar a new idea must clear. Execute a full "should this replace that?" decision including taxes and frictions, knowing why the churn hurdle is higher than the raw expected-return gap.
5. Treat cash as a position: compute the expected drag of holding cash and the option value of dry powder, articulate the Klarman/Marks view that cash is the residual of a bottom-up process, and present the drag-versus-optionality tradeoff evenhandedly.
6. Explain the five documented factors (value, quality, momentum, size, low-volatility): what each is, the evidence and the honest caveats for each, and factor cyclicality. Then diagnose the implicit factor tilts of a bottom-up portfolio, answering the sharp question "are you just a value (or quality) factor, and could this be rented cheaply?"
7. Present the efficient-markets and active-versus-passive debate evenhandedly: state the three forms of the EMH, the after-fee underperformance evidence, the Grossman-Stiglitz paradox, and the specific places an active edge plausibly persists. Reach the honest conclusion about when active is justified.
8. Apply portfolio-level risk overlays (sector and factor concentration limits, plus correlation-aware sizing carried over from M9.01) so that no single shared factor can permanently impair the book.

Prerequisites & connections

Builds on. M9.01 gave you risk as probability × magnitude of permanent impairment, conviction × downside sizing, fractional Kelly, and the "30 names, one factor = one bet" hidden-correlation idea; those single-position tools now assemble into a whole-portfolio policy. M6.02 §4.9 set out the concentration-versus-diversification debate and its Kelly/ruin spine, read philosophically there and made quantitative here. M3.02 covers diversification, systematic versus idiosyncratic risk, and the "one free lunch" of removing idiosyncratic risk, plus the CAPM machinery and beta that the low-volatility factor (§4.6) contradicts. M2.03 and M3.03 built ROIC and the WACC/hurdle by hand (~12% INR, ~8–9% USD as of mid-2026, check against your own builds); the opportunity-cost hurdle here is that number, plus the honest floor of what an index would give you. M3.05–M3.06 hold the reverse-DCF and multiples work that produces the expected forward return every replace-that decision compares. M6.04 names the behavioral failure modes (loss aversion, overconfidence, recency) that turn a good cash policy into a bad one and a factor headwind into a self-firing.

Feeds forward. M9.03 covers selling: "a better opportunity exists" is one of the five legitimate sell reasons, and it is exactly the opportunity-cost decision of §4.4; the concentration policy you set here determines how ruthlessly you must prune. M9.04 is the process capstone, where your written concentration, hurdle, and cash policy become three sections of the investment-process document, and the simulator season scores whether you actually held to them. In the Phase 10 capstones, the two institutional deep dives end in a sizing and portfolio-fit recommendation that uses these overlays. Competency C9 in the master map, "an investor's temperament and process", is largely certified on whether the policy you write here is coherent and whether you can execute it under a drawdown.

The one-sentence version of this module. A portfolio is a system of independent bets rather than a collection of good ideas, each sized so its worst case cannot sink you and each holding its place only by beating your best available alternative, and building it well means knowing the math of how many bets you actually have, the discipline of what any new one must clear, the honest role of cash, and whether the returns you are chasing are real stock-picking or a factor tilt you could rent for a tenth of the effort.

Core teaching content

4.1 From one position to the whole book: the two questions

You have arrived at portfolio construction with a powerful single-position toolkit and a dangerous blind spot. The toolkit (M9.01): given an idea, you can estimate its downside, its conviction-weighted upside, and a fractional-Kelly size. The blind spot: every one of those judgments was made as if the position existed alone. It does not. The moment you own a second thing, three new facts appear that no single-position analysis can see.

• Positions interact. Two holdings that each look moderate-risk in isolation can be one large risk together if they move together, and genuinely safer together if they don't. Risk is not additive; it runs through correlation. This is §4.2.
• Capital is finite and rival. A rupee in your tenth-best idea is a rupee not in your best. Every position is therefore competing against every other and against the alternatives you don't own, including simply owning the index or holding cash. This is the opportunity-cost discipline of §4.4–4.5.
• You may not own what you think you own. A book of thirty hand-picked stocks can secretly be a single bet on one factor (all cheap, all small, all rate-sensitive). Your "stock-picking" returns may be a factor tilt in disguise (§4.6), and the whole enterprise of picking rather than indexing has to justify itself against the efficient-markets evidence (§4.7).

So portfolio construction reduces to two questions, and everything that follows is machinery for answering them well:

Q1: How many bets should I hold, and how large should each be? (Concentration policy: §4.2–4.3, with the correlation-aware sizing and limits of §4.8.) Q2: Does each holding still earn its place against the best alternative? (Opportunity-cost discipline: §4.4, with cash as the alternative of last resort in §4.5.)

Hold the spine of the whole program in view while you do this. Value is created when a business earns ROIC above its cost of capital and can reinvest at that spread; your job as a portfolio manager is the analogous discipline one level up: allocate your scarce capital only to bets whose expected return beats your own cost of capital (your hurdle) and whose risks are independent enough that no single mistake, or single shared factor, can end the game. Portfolio construction is capital allocation applied to yourself.

4.8 Portfolio-level risk overlays

The single-position tools of M9.01 and the concentration policy of §4.3 produce a proposed book. Before you own it, run three portfolio-level overlays that catch the risks no single-position analysis can see. They are the operational answer to hidden correlation.

Sector and factor concentration limits. Set explicit caps, written in advance, on how much of the portfolio may be exposed to any one sector, and (harder but more important) any one factor. A rule like "no more than ~25–30% in any one sector, no more than ~40% in any one factor tilt, no more than ~X% in a single name" is crude but powerful: it makes the hidden-correlation mistake structurally impossible. The 30-names-all-NBFCs portfolio (§4.3, M9.01) never gets built if you have a "≤25% financials" limit. Limits will occasionally stop you from buying your tenth attractive bank. That is exactly their job, because your tenth attractive bank is not a new bet, it is more of a bet you already have.

Correlation-aware sizing. Carry M9.01's insight into every new-position decision: a name's contribution to portfolio risk depends not on its own volatility but on its correlation with what you already own. A new idea that is highly correlated with your existing holdings (another software name in a software-heavy book) adds far more risk per rupee than an equally-volatile idea that is uncorrelated (a domestic utility in that same book), so the correlated name deserves a smaller size, or must clear a higher hurdle, than its standalone conviction suggests. Concretely: two ideas with identical standalone attractiveness are not equally sizable if one duplicates a factor you're already loaded on and the other diversifies it. The uncorrelated one is worth more to the portfolio than to itself; size accordingly.

The independent-bet count and the stress test. Before finalizing, do two things. First, count your independent bets, not your names (§4.2–4.3): group holdings by dominant shared factor and ask "how many genuinely different things am I betting on?" If a 15-stock book has only 5 independent bets, you are running a 5-bet portfolio and should size and diversify as if you held 5, not 15. Second, run a crude factor stress test: pick the one factor or macro variable your book is most exposed to (rates, the rupee, a single commodity, the credit cycle) and ask "if that turns hard against me, how much of the portfolio moves together, and is the drawdown survivable?" This is the M9.01 ruin constraint applied to the whole book at once: the question is not whether any single position can ruin you, but whether any single shared factor can. If the answer is "yes, a rate shock takes 60% of my book down together," you have a concentration problem dressed as a diversified portfolio, and the sector/factor limits above are how you fix it.

These overlays are the bridge from "a set of well-sized individual ideas" to "a portfolio built to survive", which is the whole point of Phase 9, and the discipline the simulator will test for the rest of the program.

Common mistakes & how experts think differently

Counting names instead of independent bets. The commonest and most dangerous error: owning 30 tickers that share one factor and calling it diversified. The novice counts to 30 and feels safe; the expert asks "how many independent things am I betting on?" and counts the 30-name NBFC book as roughly one bet (§4.2's ρ̄ ≈ 1 collapses the systematic floor to a single name's risk; IL&FS/DHFL took the whole complex down at once, M5.02). Diversification is measured in independent risk factors, never in ticker count.

Believing diversification drives risk to zero. The variance math (§4.2) is explicit: diversification removes only the idiosyncratic part; it converges to the systematic floor σ√ρ̄, not to zero. The expert knows the floor is set by average correlation, so the way to lower portfolio risk is to hold less-correlated things, not merely more things, and that adding your 40th correlated name is nearly pure diworsification (return dilution with negligible risk benefit).

Over-diversifying out of hidden fear (or under-diversifying out of hidden ego). Both errors masquerade as philosophy. The person who owns 60 names because they can't bear to concentrate is usually hiding from the work of ranking ideas; the person who owns 3 names at full-Kelly conviction is usually hiding overconfidence behind Buffett quotes. The expert sets the number by the reliability of their edge and the survival constraint (§4.3), sizes by fractional Kelly (never full), and holds a hard rule that no single name's or single factor's worst case can end the game.

Ignoring the tax and friction cost of churn. The novice sees a 13% idea and a 12% holding and switches, feeling diligent. The expert runs the after-tax, after-friction arithmetic (§4.4) and sees that a 1-point gross edge is destroyed by the LTCG tax on the switch, so doing nothing is the higher-return action. The churn hurdle is higher than the raw expected-return gap, and most portfolio activity is value-destroying.

Treating cash as failure. The novice feels that any cash is money "not working" and lowers standards to stay fully invested: the institutional imperative in miniature. The expert treats cash as the disciplined residual of a bottom-up process (Klarman/Marks, §4.5): held not as a market forecast but because nothing currently clears the hurdle, priced honestly for its drag, and valued for the option it represents when the panic comes and everyone else is a forced seller. (The same expert also knows the opposite trap: using "cash for optionality" as a cover for a market-timing habit that mostly hurts.)

Mistaking a factor tilt for stock-picking skill. The novice who owns quality compounders and beats a value-heavy index in a quality-led year concludes they are brilliant; the same novice, lagging in a value-led year, concludes they are broken and abandons the approach at the worst moment. The expert diagnoses their factor tilts (§4.6), benchmarks against the matching factor blend, and locates their real edge in what remains, holding process through factor headwinds and refusing to take credit for factor tailwinds.

Assuming an edge instead of proving one. The deepest error, and the one this whole argument is built to prevent: playing the active game without honestly confronting the efficient-markets evidence (§4.7). The novice assumes they can beat the market because they're smart and trying hard; the expert knows the arithmetic guarantees the average active dollar loses to passive after fees, names the specific corner where their edge plausibly lives (small-caps, special situations, time-horizon arbitrage, India's less-efficient pockets), demands evidence they're in the justified minority, and indexes the capital for which they can't make that case.

Confusing volatility with risk at the portfolio level. §4.2's math is all in standard deviation. Useful, but it is volatility, not the permanent impairment M9.01 defined as real risk. The expert uses the variance math to understand diversification and correlation, but never forgets that the deeper portfolio risk is a shared factor causing permanent loss across many holdings at once (a debt crisis, a fraud in a concentrated position, a regime the whole book was implicitly betting against), which is why the overlays of §4.8 stress-test for survivable drawdowns, not just low volatility.

Worked examples

Worked Example 1: Portfolio variance from a correlation matrix, and the hidden-correlation contrast

This is the computation the mini-project and the app sandbox are built on. Do it by hand once.

You are choosing between two three-stock portfolios, each equally weighted (1/3 in each name). Portfolio D (diversified) holds one bank, one IT-services exporter, and one paint/consumer company, all genuinely different businesses. Portfolio C (concentrated in one factor) holds three private-sector banks. Illustrative inputs (modeled on realistic single-stock volatilities; pull your own from price history):

Portfolio D: σ_bank = 30%, σ_IT = 25%, σ_paint = 20%; correlations ρ(bank,IT) = 0.20, ρ(bank,paint) = 0.30, ρ(IT,paint) = 0.25. Portfolio C: three banks, each σ = 30%, all pairwise ρ = 0.75.

Step 1: Portfolio D variance. Use σ_p² = Σ wᵢ²σᵢ² + Σ_{i≠j} wᵢwⱼρᵢⱼσᵢσⱼ with w = 1/3 (so w² = 1/9).

Variance (diagonal) terms: (1/9)(0.30² + 0.25² + 0.20²) = (1/9)(0.09 + 0.0625 + 0.04) = (1/9)(0.1925) = 0.021389.

Covariance (off-diagonal) terms, where each pair's covariance is ρσᵢσⱼ and each appears twice (2wᵢwⱼ):

• bank–IT: ρ σ_bank σ_IT = 0.20 × 0.30 × 0.25 = 0.0150
• bank–paint: ρ σ_bank σ_paint = 0.30 × 0.30 × 0.20 = 0.0180
• IT–paint: ρ σ_IT σ_paint = 0.25 × 0.25 × 0.20 = 0.0125

Sum of the three covariances = 0.0150 + 0.0180 + 0.0125 = 0.0455; times 2 × (1/9): 2 × (1/9) × 0.0455 = 0.010111.

σ_p²(D) = 0.021389 + 0.010111 = 0.031500 → σ_p(D) = √0.0315 = 0.1775 = **17.8%**.

Compare to the weighted-average SD (what you'd bear with zero diversification, ρ = 1): (30 + 25 + 20)/3 = 25.0%. Diversification took 25.0% → 17.8%, a 7.2-point reduction, because the businesses are genuinely uncorrelated.

Step 2: Portfolio C variance (three correlated banks).

Variance terms: 3 × (1/9)(0.30²) = 3 × (1/9)(0.09) = 0.030000. Covariance terms: each pair = 0.75 × 0.30 × 0.30 = 0.0675; three pairs sum to 0.2025; times 2 × (1/9): 2 × (1/9) × 0.2025 = 0.045000. σ_p²(C) = 0.030000 + 0.045000 = 0.075000 → σ_p(C) = √0.075 = 0.2739 = **27.4%**.

Step 3: read the result. Portfolio C holds three names but sits at 27.4% SD, barely below a single bank's 30%. Portfolio D holds the same number of names at 17.8%. The difference is entirely correlation: three bank tickers are close to one bet (their high ρ̄ keeps the systematic term huge), while three uncorrelated businesses genuinely cancel each other's idiosyncratic swings. This is hidden correlation in hard numbers, and it is why §4.8's sector limits exist. Owning "three stocks" told you nothing; owning three independent stocks told you everything.

Worked Example 2: The opportunity-cost "should this replace that?" decision, with tax

The Munger discipline made concrete, including the friction beginners omit. You run a fully-invested US book with an 8% real hurdle. Holding X, a mature and perfectly fine business, now offers an estimated 8% forward return (right at your hurdle, no better). A new idea Y off your deep-dive pipeline offers an estimated 14% forward return. Replace X with Y?

Step 1: the naive vs. the disciplined frame. Naive: "14 > 8, obviously switch." Disciplined: switching sells X, which triggers tax, so compare the after-friction five-year outcomes of keeping X versus switching to Y.

Step 2: the tax on the switch. X is worth $200 against a cost basis of $120, an $80 long-term gain. US long-term capital-gains tax at 15% (ignore the possible 3.8% NIIT for simplicity; add it if it applies to you) is 0.15 × 80 = $12. You redeploy $188, not $200.

Step 3: compound both paths five years.

• Keep X: $200 × 1.08⁵ = 200 × 1.46933 = $293.87.
• Switch to Y: $188 × 1.14⁵ = 188 × 1.92541 = $361.98.
• Y wins by ~$68 after paying the tax, a clear and decisive replacement. The 6-point edge easily covers the ~6% capital you lost to tax on the switch.

Step 4: now find the churn hurdle. Re-run with Y offering only 9.5% (a 1.5-point edge over X's 8%):

• Switch to Y: $188 × 1.095⁵ = 188 × 1.574238 = $295.96, versus keeping X at $293.87. Essentially a dead heat.
• So a 1.5-point gross expected-return edge is entirely consumed by the tax on the switch. Below roughly that gap, the higher-return action is to do nothing.

Step 5: the lesson. The bar to replace a holding is not "the new idea has a higher number." It is "the new idea beats the old by enough to clear the tax and frictions of switching and to compensate for trading a return you partly understand for one that is still an estimate." This is why disciplined opportunity-cost investors turn over their portfolios slowly, why most churn is value-destroying, and why the rising bar of the opportunity-cost portfolio (§4.4) quietly compounds quality without constant trading.

Worked Example 3: Diagnosing a portfolio's implicit factor tilts

Answering "am I just a factor?" for a real book. You hold eight quality-compounder names. India: HDFC Bank, Asian Paints, TCS, Titan. US: Costco, Visa, Moody's, Adobe. Score each holding's exposure to each factor's "long leg" as +1 (strong), 0 (neutral), −1 (opposite). (Scores are illustrative analyst judgments; in practice you'd estimate loadings from a regression against factor returns or from the stock's characteristics.)

Read the profile. This portfolio is, whatever its owner calls it, strongly long Quality (+1.00) and Low-Volatility (+0.63), strongly short Value (−0.75, i.e., you are paying up for quality) and short Size (−1.00, all large-caps), and roughly Momentum-neutral (+0.13). In factor language: long QMJ and BAB, short HML and SMB. Two conclusions follow, and they are the whole point of the exercise:

1. Your benchmark is a quality + low-vol factor blend, not the index. A cheap quality/min-vol ETF captures much of what this book delivers. Your stock-picking is earning its keep only to the extent you beat that blend, not to the extent you beat the Nifty or S&P. That is the honest yardstick for "do I have alpha beyond my tilt?"
2. You have pre-diagnosed your bad years. When value violently outperforms quality (as in 2021–2022, or off a deep-cycle bottom), this book will lag for reasons unrelated to your company analysis. Knowing that in advance is what lets you hold the process through the headwind instead of firing yourself at the worst possible moment, and stops you from mistaking a quality tailwind for genius. If, after honest benchmarking, nothing remains beyond the factor blend, intellectual honesty says rent the blend and reclaim your research time (§4.7).

Worked Example 4: The effective number of bets, and what a position cap buys you (India, ₹)

"How many names do I really own?" The answer is almost never the ticker count, because concentration hides in the weights. A synthetic ₹50,00,000 book, Meridian Capital (synthetic), holds eight Indian stocks:

Step 1: the effective number of holdings. The concentration of a set of weights is captured by the Herfindahl index HHI = Σ wᵢ², and its reciprocal 1/HHI is the effective number of equally-weighted holdings, the count of equal positions that would be exactly this concentrated.

HHI = 0.28² + 0.16² + 0.14² + 0.12² + 0.11² + 0.08² + 0.06² + 0.05² = 0.0784 + 0.0256 + 0.0196 + 0.0144 + 0.0121 + 0.0064 + 0.0036 + 0.0025 = 0.1626.

Effective N = 1 ÷ 0.1626 = **6.15**. You hold eight names, but your capital is deployed as though you held only ~6.1 equal ones: the top two names alone (A + B = 44%) carry nearly half the book. The eight-count flatters your true diversification.

Step 2: impose a per-name cap and re-measure. §4.8's per-name limit is not only a ruin control; it mechanically raises your effective bet count. Apply a hard 20% cap. Only A (28%) breaches it. Trim A to 20%, selling ₹4,00,000 (the 8-point excess, 8% of the book), and redeploy that ₹4,00,000 pro-rata across the other seven names (which sum to 72%):

New weights: A 20.00%, B 17.78%, C 15.56%, D 13.33%, E 12.22%, F 8.89%, G 6.67%, H 5.56% (sum 100%, none above the cap). Now HHI = 0.1440, so Effective N = 1 ÷ 0.1440 = **6.95**. One cap on one name moved you from 6.15 → 6.95 effective bets, a ~13% gain in genuine diversification from a single ₹4,00,000 trade, no new research required.

Step 3: the honest caveat. 1/HHI measures weight concentration only; it assumes the names are independent. If several of these eight share a factor (all the same sector, all rate-sensitive), the number of independent bets (§4.2–4.3) is lower still, because weight-concentration and correlation-concentration stack. The effective-N number is a ceiling on your diversification, not a floor; §4.2's ρ̄ and Worked Example 5's risk contributions tell you how much lower the truth sits.

Worked Example 5: Contribution to risk: the 10% position that is 21% of your risk (US, $)

A dollar of capital and a unit of risk are not the same currency; §4.8's correlation-aware sizing is exactly this arithmetic. A synthetic $1,000,000 US book, Cascade Partners (synthetic), holds three positions:

Correlations: ρ₁₂ = 0.40, ρ₁₃ = 0.50, ρ₂₃ = 0.70 (positions 2 and 3 move closely together).

Step 1: portfolio variance and SD. Build the covariance matrix Cᵢⱼ = ρᵢⱼσᵢσⱼ (diagonal = σᵢ²) and compute σ_p² = Σᵢ Σⱼ wᵢwⱼCᵢⱼ:

σ_p² = 0.6²(0.15²) + 0.3²(0.25²) + 0.1²(0.45²) + 2(0.6)(0.3)(0.40)(0.15)(0.25) + 2(0.6)(0.1)(0.50)(0.15)(0.45) + 2(0.3)(0.1)(0.70)(0.25)(0.45) = 0.0081 + 0.005625 + 0.002025 + 0.0054 + 0.00405 + 0.004725 = 0.029925, so σ_p = √0.029925 = **17.30%**.

Step 2: each position's contribution to risk. A holding's share of risk is not its share of capital. The component contribution to variance is CCVᵢ = wᵢ · Cov(i,p), where Cov(i,p) = Σⱼ wⱼCᵢⱼ is the covariance of the holding with the whole portfolio; by construction the CCVᵢ sum exactly to σ_p².

• Cov(1,p) = 0.6(0.0225) + 0.3(0.0150) + 0.1(0.03375) = 0.021375; CCV₁ = 0.6 × 0.021375 = 0.012825.
• Cov(2,p) = 0.6(0.0150) + 0.3(0.0625) + 0.1(0.07875) = 0.035625; CCV₂ = 0.3 × 0.035625 = 0.0106875.
• Cov(3,p) = 0.6(0.03375) + 0.3(0.07875) + 0.1(0.2025) = 0.064125; CCV₃ = 0.1 × 0.064125 = 0.0064125.

Check: 0.012825 + 0.0106875 + 0.0064125 = 0.029925 = σ_p² ✓. As shares of total risk:

Step 3: read it. Position 3 is 10% of the money but 21.4% of the risk, more than double its capital weight, because it is both the most volatile name and highly correlated (0.70) with position 2. The "safe-looking" 60% core carries only 42.9% of the risk. This is why §4.8 sizes a name by its contribution to portfolio risk, not its standalone conviction: to bring position 3's risk share down to its 10% capital share you must cut it, or cut position 2, its correlated twin. Sizing by capital alone silently overweights your most volatile, most correlated ideas.

Practice set

Work each problem completely before reading its solution. Numeric answers are graded to the stated tolerance. Formula reminders: two-asset σ_p² = w₁²σ₁² + w₂²σ₂² + 2w₁w₂ρσ₁σ₂; equal-weight N-asset with average correlation ρ̄: σ_p = σ·√[ρ̄ + (1−ρ̄)/N]; systematic floor = σ√ρ̄.

P1 (guided, 8 min, two-asset variance). A portfolio holds 60% in Asset A (σ = 20%) and 40% in Asset B (σ = 35%), with correlation ρ = 0.25. (a) Compute σ_p. (b) What would σ_p be if ρ = 1? if ρ = 0? (c) In one sentence, what is the diversification benefit here?

Solution. (a) σ_p² = 0.6²(0.20²) + 0.4²(0.35²) + 2(0.6)(0.4)(0.25)(0.20)(0.35). Terms: 0.36 × 0.04 = 0.0144; 0.16 × 0.1225 = 0.0196; 2 × 0.24 × 0.25 × 0.07 = 0.0084. Sum = 0.0424; σ_p = √0.0424 = **20.6%**. (b) If ρ = 1, σ_p = weighted-average SD = 0.6(20) + 0.4(35) = **26.0%**. If ρ = 0, drop the last term: σ_p² = 0.0144 + 0.0196 = 0.0340, σ_p = **18.4%**. (c) At the actual ρ = 0.25, diversification cut risk from the 26.0% you'd bear if they moved together to 20.6%, a 5.4-point reduction that comes entirely from the two assets not moving in lockstep.

P2 (guided, 8 min, the systematic floor and the Nth name). Your holdings each have σ = 28% and an average pairwise correlation ρ̄ = 0.25. (a) What is the portfolio-SD floor as N → ∞? (b) Compute σ_p for N = 10 and N = 30. (c) How much SD does going from 10 to 30 names buy you, and what does that say about adding your 30th name?

Solution. (a) Floor = σ√ρ̄ = 28% × √0.25 = 28% × 0.5 = **14.0%**. (b) N = 10: σ_p = 28 × √(0.25 + 0.75/10) = 28 × √0.325 = 28 × 0.5701 = **16.0%**. N = 30: σ_p = 28 × √(0.25 + 0.75/30) = 28 × √0.275 = 28 × 0.5244 = **14.7%**. (c) Tripling from 10 to 30 names buys only 16.0 − 14.7 = **1.3 points** of SD, and at 30 names you are already within 0.7 points of the 14.0% floor. Past ~10–15 names you are chasing basis points: the marginal benefit of the Nth name has gone to near-zero, exactly where diworsification begins.

P3 (independent, 10 min, hidden correlation). Portfolio A and Portfolio B each hold four equally-weighted stocks with σ = 30% each. In A the average correlation is ρ̄ = 0.20; in B (four names in one sector) it is ρ̄ = 0.80. (a) Compute σ_p for each. (b) How many "independent bets" does B really contain? (c) State the portfolio rule this illustrates.

Solution. (a) A: σ_p = 30 × √(0.20 + 0.80/4) = 30 × √(0.20 + 0.20) = 30 × √0.40 = 30 × 0.6325 = **19.0%**. B: σ_p = 30 × √(0.80 + 0.20/4) = 30 × √(0.80 + 0.05) = 30 × √0.85 = 30 × 0.9220 = **27.7%**. (b) B sits at 27.7%, barely below a single name's 30%, so its four names behave as roughly one bet held four times; the shared sector factor dominates. (c) Count independent bets, not tickers, and impose sector/factor limits (§4.8) so a "diversified" name count can't hide a single-factor concentration.

P4 (independent, 10 min, diworsification arithmetic). Your top 12 ideas have an average expected return of 14%; ideas 13–25 average 8%. Assume every name has σ = 30% and ρ̄ = 0.25. (a) Compute portfolio expected return and σ_p at N = 12 and at N = 25 (equal-weighted). (b) Quantify the trade of going 12 → 25. (c) Under what single condition is this trade actually worth making?

Solution. (a) Expected return is the average of holdings' expected returns. N = 12: 14.0%. N = 25: (12 × 14 + 13 × 8)/25 = (168 + 104)/25 = 272/25 = **10.9%**. SDs: N = 12: 30 × √(0.25 + 0.75/12) = 30 × √0.3125 = 30 × 0.5590 = **16.8%**. N = 25: 30 × √(0.25 + 0.75/25) = 30 × √0.28 = 30 × 0.5292 = **15.9%**. (b) Going 12 → 25 costs 14.0 − 10.9 = **3.1 points of expected return** to buy only 16.8 − 15.9 = **0.9 points of SD**. A terrible trade (over 3 points of return per point of volatility removed). This is diworsification. (c) It is worth making only if you cannot reliably tell your "14% ideas" from your "8% ideas" — i.e., if your ranking is noise, the return "cost" is illusory (they're really all the same expected return) and the risk reduction is a real free gain. The whole concentration decision hinges on the honesty of that self-assessment.

P5 (independent, 12 min, replace-that with tax). A fully-invested Indian book. Holding X: current value ₹300, cost basis ₹200, estimated forward return 9%. New idea Y: estimated forward return 15%. LTCG tax 12.5% on the gain (ignore the exemption for this position; as of mid-2026). (a) Compute the redeployable amount after selling X. (b) Compare 5-year outcomes of keeping X vs switching to Y. (c) Find the approximate Y return at which the switch becomes a wash, and state the lesson.

Solution. (a) Gain = 300 − 200 = ₹100; tax = 0.125 × 100 = ₹12.5; redeployable = 300 − 12.5 = **₹287.5**. (b) Keep X: 300 × 1.09⁵ = 300 × 1.53862 = **₹461.6**. Switch to Y: 287.5 × 1.15⁵ = 287.5 × 2.01136 = **₹578.3**. Y wins by ₹116.7, so switch. (c) Solve 287.5 × (1+r)⁵ = 461.6 → (1+r)⁵ = 1.6056 → 1+r = 1.6056^(0.2) = 1.0989 → r ≈ **9.9%**. So Y must beat X's 9% by roughly a full point just to break even after the tax. That is the churn hurdle. Lesson: the tax on switching raises the replacement bar well above the raw expected-return gap; only switch when the new idea beats the old by enough to clear the friction plus the loss of certainty.

P6 (independent, 8 min, the hurdle stack and cash). Your required return (hurdle) is 12%. A broad index is expected to return ~10%. You hold cash and have three candidate ideas with estimated forward returns of 11%, 9%, and 14%. (a) Which ideas clear your hurdle? (b) The 11% idea beats the index. Should you buy it? (c) If only the 14% idea clears and it is already fully sized, what do you do with the rest, and what is that position called?

Solution. (a) Only the 14% idea clears the 12% hurdle. (b) Not necessarily: 11% beats the index's 10% but not your 12% required return, so a single stock's extra risk isn't being compensated relative to your bar. The disciplined choices are to buy the index (if 11% single-name risk isn't worth 1 point over a 10% index) or to pass and keep the capital for the 14% idea; buying an 11% stock because it beats the index but ignoring your own hurdle is lowering your standard. (c) Hold the remainder in cash, a legitimate position and the bottom-up residual of applying your hurdle (§4.5), not a failure. It carries a drag but also the option to deploy when a better idea (or a panic) appears.

P7 (independent, 12 min, cash drag vs dry powder). Equities are expected to return 11%/year, cash 6%/year. You hold 25% cash. (a) Compute the expected annual drag. (b) Over a 5-year bull market (equities compounding at 11% throughout), how much terminal wealth do you give up versus being fully invested? (c) Sketch the offsetting dry-powder case and state the honest evenhanded conclusion.

Solution. (a) Expected annual drag = 0.25 × (11% − 6%) = **1.25%/year**. (b) Fully invested: 1.11⁵ = 1.6851. With 25% cash: 0.75 × 1.11⁵ + 0.25 × 1.06⁵ = 0.75 × 1.6851 + 0.25 × 1.3382 = 1.2638 + 0.3346 = 1.5984. Ratio = 1.5984 / 1.6851 = 0.9485 → you end with about 5.1% less wealth after five bull years. The drag is real and certain. (c) Dry-powder case: if instead a panic strikes and you deploy the 25% at a 40% discount into an idea that then compounds at 20% from the trough, that cash effectively buys 1/0.60 ≈ 1.67× as much "pre-panic value" and compounds off a depressed base, precisely when the fully-invested are forced sellers and can't act. Honest conclusion: the drag is certain and hurts in bull markets; the dry-powder payoff is state-contingent (needs a panic to arrive and you to deploy into it) and can be large. Cash held as a bottom-up residual is defensible; cash held as a market-timing bet usually is not. Most investors misuse the lever, so know which kind you are.

P8 (independent, 12 min, factor-tilt diagnosis). Score this six-stock book on the five factors (+1/0/−1 for exposure to each factor's long leg): a cheap PSU bank, a cheap cyclical steelmaker, a small-cap value industrial, a cheap regulated utility, an out-of-favor pharma name, and a small-cap textile maker. Use these analyst scores:

(a) Compute each factor's average score. (b) Name the portfolio's dominant tilts. (c) What is this book's correct benchmark, and when will it hurt and help?

Solution. (a) Sums: Value +6; Quality −3; Momentum −2; Size +2; Low-Vol −2. Averages (÷6): Value +1.00, Quality −0.50, Momentum −0.33, Size +0.33, Low-Vol −0.33. (b) A deep-value, small-tilted, low-quality, anti-momentum, higher-volatility book: long HML and (mildly) SMB, short quality, momentum, and low-vol. (c) Its correct benchmark is a value + small-cap factor blend, not the broad index. It will hurt in growth/quality-led markets (like 2007–2020's value drought) and help sharply in value rebounds (2021–2022, recoveries off deep-cycle bottoms). Those cyclical swings are the factor tilt, not the stock selection, and must be held through rather than fired.

P9 (independent, 6 min, name the factor). For each strategy, name the single factor it primarily harvests: (a) buy the cheapest quintile by price-to-book; (b) buy the past year's biggest winners and rotate monthly; (c) buy firms with high gross profitability and low leverage; (d) buy the lowest-beta quintile; (e) tilt to the smallest-cap quintile.

Solution. (a) Value (HML). (b) Momentum (UMD). (c) Quality (profitability/QMJ/RMW). (d) Low-volatility / betting-against-beta (BAB). (e) Size (SMB). Bonus honesty: (c) and (d) heavily overlap, and (e) is the weakest and most fragile of the five once you control for quality.

P10 (timed, 15 min, steelman both sides). In two tight paragraphs, argue first the strongest case that a smart, hard-working analyst should index rather than pick stocks, and then the strongest case that active management can be justified. End with one honest sentence on how to decide.

Solution (model answer). For passive: Active management is a negative-sum game after costs by arithmetic identity (Sharpe): active dollars collectively are the market before fees, so they must underperform passive dollars by their fees after them, and the empirical record (SPIVA: ~85–90% of US large-cap active funds trail the S&P over 15 years) confirms it, with almost no persistence among the winners. Indexing captures the equity premium, removes idiosyncratic risk for free, costs almost nothing, requires no skill you may lack, and protects you from your own behavioral errors. For anyone without a demonstrable edge, it is the intellectually honest choice, and most people don't have the edge they think they do. For active: Markets cannot be perfectly efficient. Grossman-Stiglitz shows that if prices reflected all information no one would be paid to gather it, so no one would, so prices couldn't; there must be enough mispricing to reward the informed at the margin. That reward is real but localized, living in small-caps, special situations, time-horizon arbitrage (patience the quarterly crowd can't offer), and less-efficient markets like India's mid-caps, where a disciplined process on public information genuinely wins. To decide: index your core by default, and run active capital only where you can name your specific, persistent edge and show, benchmarked against the matching factor blend rather than the broad index, that it is real.

P11 (timed, 15 min, set a concentration policy). An investor can run genuine 3-day deep dives on about 12–14 businesses and keep them current; invests mostly in Indian mid-caps where governance and accounting risk are elevated; and admits that in March 2020 he panicked and sold near the bottom. Recommend a concentration policy (number of positions, per-name cap, sector/factor limits, and cash stance) with reasons tied to his edge, his information reliability, and his temperament.

Solution (model answer). Number of positions: lean to the higher end of the deep-work range, ~14–18 names. His analytical capacity supports concentration (he can truly know ~12–14 businesses), but two forces push up the count: elevated Indian mid-cap governance/fraud risk means information is unreliable per name (a structural reason to diversify more, §4.3–4.7), and his 2020 capitulation flags a temperament that needs enough breadth and quality to hold through drawdowns rather than sell. Per-name cap: a hard ~7–8% maximum, so that even a total loss from a fraud (a live risk in this universe) is a survivable dent, not a mortal wound. That is the ruin cap of §4.3. Sector/factor limits: ≤25% in any one sector, ≤~35–40% in any one factor, to prevent a "diversified" 16-name book from secretly being one bet on, say, financials or the mid-cap-value factor (§4.8). Cash: no target and no cap, since cash is the bottom-up residual when nothing clears his hurdle; but given his panic history, pre-commit in writing to a rule (a checklist and a decision journal, M9.04) that stops him from converting that cash into a market-timing lever he'll misuse at the bottom. Honest coda: index the core capital he cannot cover at deep-dive depth (§4.7); run the concentrated active sleeve only where his edge, deep work in an under-covered and less-efficient corner, plausibly persists.

P12 (guided, 8 min, the effective number of bets). A six-stock book has weights 35%, 25%, 20%, 10%, 6%, 4%. (a) Compute the Herfindahl index HHI = Σ wᵢ². (b) Compute the effective number of holdings, 1/HHI. (c) In one sentence, what does the gap between six and that number tell you?

Solution. (a) HHI = 0.35² + 0.25² + 0.20² + 0.10² + 0.06² + 0.04² = 0.1225 + 0.0625 + 0.0400 + 0.0100 + 0.0036 + 0.0016 = **0.2402**. (b) 1 ÷ 0.2402 = **4.16**. (c) You hold six names but your capital is concentrated as if in only ~4.2 equal ones: the top three (80% of the book) dominate, so the ticker count overstates your diversification; a per-name cap (§4.8) would push the effective number back up toward six.

P13 (guided, 10 min, contribution to risk). A two-asset book: 70% in A (σ = 20%), 30% in B (σ = 40%), ρ = 0.30. (a) Compute σ_p. (b) Compute each asset's component contribution to variance CCVᵢ = wᵢ·Cov(i,p) and its share of total risk. (c) Does B's risk share exceed its capital weight, and why?

Solution. (a) Cov(A,B) = 0.20 × 0.40 × 0.30 = 0.024; σ_p² = 0.7²(0.04) + 0.3²(0.16) + 2(0.7)(0.3)(0.024) = 0.0196 + 0.0144 + 0.01008 = 0.04408; σ_p = √0.04408 = **21.0%**. (b) Cov(A,p) = 0.7(0.04) + 0.3(0.024) = 0.0352, so CCV_A = 0.7 × 0.0352 = 0.02464; Cov(B,p) = 0.7(0.024) + 0.3(0.16) = 0.0648, so CCV_B = 0.3 × 0.0648 = 0.01944 (they sum to 0.04408 = σ_p² ✓). Shares: A = 0.02464 ÷ 0.04408 = **55.9%**, B = **44.1%**. (c) Yes. B is only 30% of the capital but 44.1% of the risk, because its 40% volatility (double A's) makes each rupee in B contribute far more to portfolio variance. Risk share ≠ capital share.

P14 (independent, 10 min, rebalancing with tax). You hold a ₹20,00,000 book split 50/50 between A and B. A year later A is worth ₹13,00,000 and B ₹9,00,000. (a) Compute the current weights and the drift from target. (b) Compute the trades to rebalance to 50/50. (c) If the amount you sell from A carries an embedded long-term gain equal to 40% of the sale proceeds, taxed at 12.5% LTCG (as of mid-2026), compute the tax cost of rebalancing and state the discipline it argues for.

Solution. (a) Total = 13 + 9 = ₹22,00,000; weights A = 13 ÷ 22 = **59.1%**, B = **40.9%**, a +9.1-point drift in A. (b) Target = ₹11,00,000 each, so sell ₹2,00,000 of A and buy ₹2,00,000 of B. (c) Embedded gain = 40% × 2,00,000 = ₹80,000; LTCG tax = 12.5% × 80,000 = **₹10,000** (0.45% of the book). Lesson: mechanical rebalancing crystallizes tax on winners, so disciplined investors rebalance with new cash and dividends first, use threshold bands (act only when drift exceeds, say, ±5–10 points) rather than the calendar, and, echoing §4.4, never let a small drift trigger a taxable trade costing more than the risk it corrects.

P15 (independent, 10 min, factor-exposure netting). A five-stock book has these (weight, value-factor loading) pairs: (30%, +0.8), (25%, +0.6), (20%, −0.3), (15%, +0.5), (10%, +0.2). (a) Compute the portfolio's net value-factor loading, Σ wᵢ × loadingᵢ. (b) Which holding offsets the tilt, and by how much? (c) If policy caps any single net factor loading at 0.40, are you within it?

Solution. (a) Net = 0.30(0.8) + 0.25(0.6) + 0.20(−0.3) + 0.15(0.5) + 0.10(0.2) = 0.24 + 0.15 − 0.06 + 0.075 + 0.02 = **+0.425**, so the book is net long the value factor. (b) The third holding (loading −0.3) is a growth-leaning name; its contribution is 0.20 × (−0.3) = **−0.06**, so it nets the aggregate tilt down by 0.06. Factor exposures add across the book, and an offsetting position genuinely reduces the tilt. (c) No. 0.425 exceeds the 0.40 cap, so you are in breach and must trim value exposure (cut a high-loading name or add a low/negative-loading one, §4.8).

P16 (independent, 8 min, correlation-aware sizing). Two candidate ideas are identical on a standalone basis: each has σ = 30% and would enter as a small position. Your existing book has σ_book = 18%. Idea U is uncorrelated with your book (ρ = 0.0); idea C is highly correlated with it (ρ = 0.80). (a) Using the marginal contribution to risk ≈ ρ(i,book) × σᵢ, compute it for each. (b) Which idea adds more portfolio risk per rupee? (c) State the sizing rule.

Solution. (a) Idea U: 0.0 × 30% = **0%** per unit weight; idea C: 0.80 × 30% = **24%** per unit weight. (b) Idea C. Despite identical standalone volatility, the correlated idea's marginal contribution to portfolio risk is ~24 percentage points per unit weight versus ~0 for the uncorrelated one, because a name's contribution to portfolio risk runs through its covariance with what you already own, not its own σ. (c) Size the uncorrelated idea larger (it diversifies) and the correlated idea smaller, or demand a higher hurdle of it (§4.8): two ideas with equal standalone appeal are not equally sizable if one duplicates a factor you already hold.

P17 (timed, 15 min, build and diagnose a four-name book). A synthetic book: weights A 40%, B 30%, C 20%, D 10%; volatilities 20%, 22%, 35%, 40%; A and B are the same sector (ρ_AB = 0.70); other correlations ρ_AC = 0.20, ρ_AD = 0.10, ρ_BC = 0.25, ρ_BD = 0.15, ρ_CD = 0.30. (a) Compute the effective number of holdings by weight. (b) Compute σ_p. (c) Compute each name's risk share (CCVᵢ ÷ σ_p²) and identify the holding whose risk share most exceeds its weight. (d) Recommend one action.

Solution. (a) HHI = 0.4² + 0.3² + 0.2² + 0.1² = 0.16 + 0.09 + 0.04 + 0.01 = 0.30; effective N = 1 ÷ 0.30 = **3.33**. (b) Building Cᵢⱼ = ρᵢⱼσᵢσⱼ and summing Σ wᵢwⱼCᵢⱼ gives σ_p² = 0.03231, so σ_p = **17.97%**. (c) Risk shares: A 35.7% (weight 40%), B 29.7% (30%), C 24.8% (20%), D 9.8% (10%). C's risk share (24.8%) most exceeds its weight (20%), a +4.8-point gap, because it is the second-most-volatile name. (d) Either trim C toward its risk budget, or recognize that A and B (70% of capital, ρ = 0.70) are effectively one sector bet, so the honest independent-bet count is nearer three than four and the deeper fix is to replace one of A/B with something uncorrelated (§4.8). The effective-N-by-weight (3.33) already warned you the four names were closer to three bets; the correlation confirms it.

P18 (independent, 12 min, look-through earnings on a five-position book). (numeric; tolerance ±0.01 percentage points on each yield.) A ₹50,00,000 book holds five names. Weight, share price, earnings per share and payout ratio, in order: 30% at ₹2,400 with EPS ₹96 and a 25% payout; 25% at ₹1,800 with EPS ₹60 and 40%; 20% at ₹650 with EPS ₹26 and no dividend; 15% at ₹3,200 with EPS ₹80 and 50%; 10% at ₹480 with EPS ₹12 and 30%. (a) Compute total look-through earnings and the look-through earnings yield. (b) Compute the book's look-through price-to-earnings ratio and verify it a second way. (c) Compute the dividends actually received and the share of look-through earnings that was retained on your behalf. (d) If retained capital earns a 15% incremental return, compute what it adds to next year's look-through earnings, and state the one condition under which that addition does not happen.

Solution. (a) Share counts are value ÷ price: 15,00,000 ÷ 2,400 = 625.000000, 12,50,000 ÷ 1,800 = 694.444444, 10,00,000 ÷ 650 = 1538.461538, 7,50,000 ÷ 3,200 = 234.375000, 5,00,000 ÷ 480 = 1041.666667. Multiplied by EPS: ₹60,000.00, ₹41,666.67, ₹40,000.00, ₹18,750.00, ₹12,500.00, summing to ₹1,72,916.67. The yield is 1,72,916.67 ÷ 50,00,000 = 3.458333%. (b) 50,00,000 ÷ 1,72,916.67 = 28.915663. The second route is the value-weighted earnings yield: 0.30(4.000000%) + 0.25(3.333333%) + 0.20(4.000000%) + 0.15(2.500000%) + 0.10(2.500000%) = 3.458333%, identical, which is the check that the share counts are right. (c) Dividends are look-through earnings times payout: 15,000.00 + 16,666.67 + 0 + 9,375.00 + 3,750.00 = ₹44,791.67, a dividend yield of 0.895833%. Retained is the remainder, ₹1,28,125.00, or 74.0964% of the total. (d) 1,28,125.00 × 0.15 = ₹19,218.75 of additional look-through earnings, without buying a share. The condition: the incremental return has to clear the cost of capital. Retained at a return below the hurdle, that ₹1,28,125 destroys value rather than compounding it, which is why the 74% figure measures how much of your outcome rests on capital allocation you do not control (M3.07) rather than promising anything about it.

Applied mini-project: your Portfolio Construction Policy

Task. Take a real model portfolio of 8–12 stocks you could defend, built from your watchlist or the program's starter bank (India: Asian Paints, DMart, TCS, Titan, ITC, HDFC Bank, Bajaj Finance, Divi's, InterGlobe; US: Costco, Coca-Cola, Visa, Adobe, Moody's, Copart, Old Dominion, Chipotle, Deere, Apple), mixing India and US. Then write a four-to-five-page Portfolio Construction Policy that turns all of this into your own operating rules. This is the "policy memo" half of the M9.02 gate. The discipline it teaches is the one the simulator will score for the rest of the program: a portfolio is a system with written rules, not a pile of ideas.

Data to pull. For each holding: a rough single-stock volatility (from 3–5 years of monthly returns on screener.in / stockanalysis.com / your data source), a plausible pairwise-correlation matrix (estimate it, or bucket holdings by dominant factor and assign high within-bucket / low across-bucket correlations), and your own estimated forward 5-year return (from your M3.05 reverse-DCF / base-case work). For the factor diagnosis, use the +1/0/−1 scoring of Worked Example 3.

Required contents.

1. The variance and independent-bet analysis. Compute the portfolio's σ_p (or a careful approximation) from your matrix; state your systematic floor σ√ρ̄; and, the real deliverable, count your independent bets, grouping holdings by shared factor, and name any hidden correlation.
2. Your concentration policy, with reasons. State your target number-of-positions range and defend it explicitly against your edge (how many businesses can you truly keep current?), your temperament (what drawdown can you hold without capitulating?), and the survival constraint (fractional-Kelly sizing; the ruin cap). Reference where you sit on the M6.02 §4.9 dial and why.
3. Your hurdle and opportunity-cost rule. State your absolute hurdle (from your WACC build and inflation), your index floor, and the rule by which a new idea must beat your existing worst holding after tax and frictions. Then run one full "replace-that" decision on a real candidate versus your current weakest holding, with the after-tax five-year arithmetic.
4. Your cash policy. State whether cash is a target, a cap, or a pure bottom-up residual, and be honest about which behavioral trap (drag-inflicting timidity or market-timing overconfidence) you are more prone to, and how your rule guards against it.
5. Your factor diagnosis and honest benchmark. Score the book's five factor tilts, name its dominant loadings, state the factor blend that is its true benchmark, and answer in one honest paragraph: how much of this portfolio's expected return is a rentable factor tilt, and where does your genuine stock-selection edge live (§4.7's map: small-caps, special situations, time-horizon, India's pockets)?
6. Your risk overlays. Write your sector cap, factor cap, and per-name cap, and run the one-factor stress test: name the single variable your book is most exposed to and confirm the drawdown it would cause is survivable.

Scoring rubric (self-score honestly; pass ≥11/14):

Case Lab: run case XL-9.1 in the app's Cases tab. Take a fully-invested book with a fresh idea in the pipeline and a pile of idle cash, and work the whole construction decision end to end: count the independent bets (not the tickers), size by correlation-aware contribution to risk, run the after-tax replace-that call against your weakest holding, and state the cash policy that falls out.

Reading & resources

• Howard Marks, The Most Important Thing (2011). The chapters on risk, on cycles, and on defensive investing are the philosophical backbone of §4.5 and the concentration/cash discipline. Also the Oaktree memos (oaktreecapital.com/insights), free and decades deep, especially "Liquidity," "The Value of Predictions," and "You Can't Predict. You Can Prepare." [Book Paid / Memos Free] [Beginner→Intermediate]
• Seth Klarman, Margin of Safety. Out of print (PDFs circulate); read the chapters on why to hold cash and on the institutional imperative for the definitive statement of cash-as-residual. [Hard to find] [Advanced]
• John C. Bogle, The Little Book of Common Sense Investing. The clearest statement of the passive/indexing case and the "cost matters hypothesis"; pair with Charles Ellis, Winning the Loser's Game for the argument that active is a loser's game for most. [Paid] [Beginner]
• William F. Sharpe, "The Arithmetic of Active Management" (1991, ~2 pages, free online). The mathematical identity that active must underperform passive after costs in aggregate. Read it once; it is decisive. [Free] [Intermediate]
• Grossman & Stiglitz, "On the Impossibility of Informationally Efficient Markets" (1980, AER, free via search). The paradox that markets can't be perfectly efficient; skim the setup and the conclusion even if the algebra is heavy. [Free] [Advanced]
• S&P Dow Jones Indices, SPIVA Scorecards (spglobal.com/spdji, free, updated twice a year). The after-fee active-vs-benchmark record for the US and India (and the SPIVA "Persistence" reports). Pull the current edition; the numbers move. [Free] [Beginner]
• Fama & French, "The Cross-Section of Expected Stock Returns" (1992) and "A Five-Factor Asset Pricing Model" (2015); Carhart (1997) on momentum; Jegadeesh & Titman (1993) on momentum; Frazzini & Pedersen, "Betting Against Beta" (2014); Asness-Frazzini-Pedersen, "Quality Minus Junk." The primary factor literature; most are free on SSRN. [Free] [Advanced]
• AQR Capital research library (aqr.com, free). The best plain-English factor writing anywhere: "Fact, Fiction, and Value Investing," "The Devil in HML's Details," "Craftsmanship Alpha," and the factor-cyclicality pieces. [Free] [Intermediate]
• Antti Ilmanen, Expected Returns (2011). The advanced, comprehensive treatment of factors, risk premia, and portfolio construction; the reference book when you want depth. [Paid] [Advanced]
• Meir Statman, "How Many Stocks Make a Diversified Portfolio?" (1987) and Evans & Archer (1968). The original diversification-count studies behind §4.2's table. [Free via search] [Intermediate]
• Burton Malkiel, A Random Walk Down Wall Street. The readable case for the EMH and indexing, to steelman the other side of your own active ambitions. [Paid] [Beginner]
• CFA Program Level III, Portfolio Management readings. The formal treatment of portfolio construction, risk budgeting, and the institutional process; the most systematic single source. [Paid] [Advanced]
• India-specific. SPIVA India (the large-cap vs mid/small-cap active split is directly relevant to §4.7's "where edge persists"); NSE/AMFI factor-index methodologies (Nifty Alpha, Quality, Value, Low-Volatility, Momentum indices; read one methodology document to see a factor defined precisely for the Indian market); Aswath Damodaran's blog posts on the small-cap premium ("Where is the beef?") and on factor investing. [Free] [Intermediate]
• Going deeper: Attilio Meucci, "Managing Diversification" (Risk, 2009; free on SSRN). It formalizes diversification as the effective number of uncorrelated bets, the rigorous, correlation-aware version of Worked Example 4's 1/HHI weight measure; the paper behind "count independent bets, not tickers." [Free] [Advanced]
• Going deeper: Thierry Roncalli, Introduction to Risk Parity and Budgeting (2013). The definitive treatment of marginal and component contributions to risk (Worked Example 5) and risk-budgeted construction; the CFA Level III "risk budgeting" readings cover the same machinery more briefly. [Paid] [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": "Portfolio variance from a correlation matrix, the diversification curve and its systematic floor, the concentration dial and its three caps, opportunity-cost hurdles with switching frictions, cash as a residual, factor tilt diagnosis, and contribution-to-risk sizing",
  "use": "This module has one question that AI answers well and one it answers dangerously, and telling them apart is the skill. WELL: the qualitative half of the factor diagnosis in Worked Example 3. Paste your holdings and ask it to argue, per name, which of value / quality / momentum / size / low-volatility the position is long or short and why, then to summarise the book's implicit tilt and name the market regime in which the whole book lags for reasons unrelated to your company analysis. That is a hypothesis-generation task on public characteristics, and it is fast and good at it. Also well: drafting the Portfolio Construction Policy itself — hand it your rules and ask it to find the case your policy does not cover (a position that breaches two caps at once, a rights issue, a spin-off, an idea that clears the hurdle only after tax). DANGEROUSLY: anything numeric about correlation. It will happily assert that two stocks are 'moderately correlated, around 0.4' and it has not measured anything. Correlations are computed from a return panel you assemble, never recalled.",
  "tools": ["Chat assistants — Claude, ChatGPT for factor-tilt hypotheses, policy stress-cases and the pre-mortem on your own construction rules", "Coding assistants — for the covariance engine, the contribution-to-risk decomposition and the estimation-error simulation below", "Document/PDF AI — for reading a fund factsheet or an index methodology document to find what a benchmark actually holds before you benchmark yourself against it"],
  "prompt": "Here are my holdings with weights: <PASTE>. For each name, state which of the five factors (value, quality, momentum, size, low-volatility) it is long or short, in +1 / 0 / -1 form, and give the one characteristic that justifies each score. Then: (a) the book's average loading per factor; (b) the two market regimes in which this book underperforms a broad index for reasons that have nothing to do with my company analysis; (c) which cheap factor index or ETF is the honest benchmark for this tilt, in both India and the US, named. Do not estimate correlations, volatilities or expected returns — I compute those from price history.",
  "verify": "Score the tilts yourself against a regression before you believe them: regress your book's monthly excess returns on the Fama-French factors (Ken French Data Library) for the US sleeve, and on the NSE strategy indices (Nifty200 Quality 30, Nifty100 Low Volatility 30, Nifty500 Value 50, Nifty200 Momentum 30) for the Indian sleeve. A judgement score of +1 and a regression loading of 0.2 are different claims, and only the second survives a bad year. Every correlation, volatility and covariance in this module must come from a return panel you built and can rebuild; a number a model produced about co-movement is a fluent guess, and the whole of Section 4.2 turns on that number. And check the benchmark it names actually holds what it says — read the index methodology, because 'quality' has several incompatible definitions (Section 4.6) and a benchmark chosen by name rather than by construction will flatter or damn you at random.",
  "diy": "The gate is unaided: you build a covariance matrix from correlations and volatilities, compute portfolio SD by hand for two and three assets, derive sigma_p = sigma*sqrt(rho_bar + (1-rho_bar)/N) and read the systematic floor sigma*sqrt(rho_bar) off it, compute 1/HHI, run the after-tax replace-that arithmetic, and defend a concentration number in writing. No tool is in that room."
}

Modern Data Analysis

By hand first. You built the three-name covariance matrix twice, once for genuinely different businesses (17.75% SD) and once for three banks (27.39%), then derived the equal-weight formula and its floor σ√ρ̄, computed 1/HHI before and after a position cap, and decomposed a book into contributions to risk that summed exactly to σ_p². Keep those. The by-hand version is what makes the code trustworthy, because the covariance decomposition is easy to get subtly wrong and the only reliable check is a case you can verify on paper.

Today's workflow. With the arithmetic in code, three things become answerable that the paper method cannot reach.

• The diversification curve as a curve, not a table. §4.2's table gives seven points; the function gives the marginal value of every name and lets you sweep ρ̄, which is where the actual lesson lives.
• Contribution to risk on a real book, recomputed every quarter. WE5 does three positions by hand. Fifteen positions is 120 covariance pairs, and the risk shares move as prices move. A spreadsheet that is right in January is stale by April.
• And the question none of the above asks: how well do you know the covariance matrix? Every figure in §4.2 and WE5 treats ρ and σ as given. In practice they are estimated from a short return history, and the gap between "estimated" and "known" is large enough to invert the module's own advice about what to do with the numbers. That is the Quantitative lens below.

Tools & sources (IN + US). numpy for the covariance algebra, pandas for the return panel and the alignment problem that eats most of the time, numpy.linalg for the eigen-decomposition that shows you how much of your correlation matrix is noise, matplotlib for the diversification curve. India: NSE and BSE daily bhavcopy for adjusted price history (mind corporate actions; an unadjusted split will manufacture a spurious correlation), the NSE strategy indices as factor proxies for the WE3 diagnosis (Nifty200 Quality 30, Nifty100 Low Volatility 30, Nifty500 Value 50, Nifty200 Momentum 30, Nifty Alpha 50), NSE's own index methodology documents for what each of those actually screens on, screener.in for the fundamentals behind your hurdle-rate estimates, AMFI for scheme-level holdings if you want to know what a "quality fund" is really long, and the Income Tax Act's Section 112A rules for the LTCG arithmetic in §4.4. US: the Ken French Data Library (Dartmouth, free) for the Fama-French 3- and 5-factor series plus momentum, which is the rigorous version of WE3's judgement scores; EDGAR XBRL companyfacts for the profitability and investment characteristics behind a quality score; FRED (DGS3MO for the cash rate in §4.5's drag arithmetic, DGS10 for the hurdle build); Damodaran's industry cost-of-capital and ERP tables at pages.stern.nyu.edu for the hurdle stack in §4.4.

# The diversification curve, and how badly a covariance matrix is known. R3, recomputed in-session.
import numpy as np
rng = np.random.default_rng(902)
s = 0.30
sd = lambda N, r: s*np.sqrt(r + (1-r)/N)                 # 4.2's equal-weight formula
print([round(sd(N,0.30)*100,1) for N in (1,2,5,10,20,30,50)])   # 30.0 24.2 19.9 18.2 17.4 17.1 16.8
print([round((sd(N-1,0.30)-sd(N,0.30))*1e4,1) for N in (3,10,20,30)])  # 227.8  19.1  4.8  2.1 bp

# how much of an ESTIMATED covariance matrix is signal? N names, T months of history
N, T = 30, 60
beta = rng.normal(1.0, 0.25, N); sf = 0.18/np.sqrt(12)
idio = rng.uniform(0.20, 0.40, N)/np.sqrt(12)
Sig  = np.outer(beta, beta)*sf**2 + np.diag(idio**2)      # the TRUE world
one, mu = np.ones(N), np.full(N, 0.09/12)                 # equal true means: any tilt is noise
ann = lambda v: np.sqrt(v*12)*100
out = {"est_minvar": [], "equal": [], "est_mvo": []}
for _ in range(2000):
    X = rng.multivariate_normal(mu, Sig, T); S = np.cov(X, rowvar=False)
    w  = np.linalg.solve(S, one);      w  /= w.sum()
    wm = np.linalg.solve(S, X.mean(0)); wm /= wm.sum()
    out["est_minvar"].append(w@Sig@w); out["equal"].append((one/N)@Sig@(one/N))
    out["est_mvo"].append(wm@Sig@wm)
print({k: round(ann(np.median(v)), 2) for k, v in out.items()})
# {'est_minvar': 21.27, 'equal': 18.57, 'est_mvo': 87.42}  — true optimum is 15.32

Verify. Reproduce the module's own numbers before trusting the engine on yours. §4.2's table returns 30.0 / 24.2 / 19.9 / 18.2 / 17.4 / 17.1 / 16.8% for N = 1, 2, 5, 10, 20, 30, 50 with a floor of 16.43%, and every row matches. WE1: Portfolio D variance 0.031500, SD 17.75%, against a weighted-average SD of 25.0%; Portfolio C variance 0.075000, SD 27.39%. WE4: HHI = 0.1626, effective N 6.15; after the 20% cap the weights become 20.00 / 17.78 / 15.56 / 13.33 / 12.22 / 8.89 / 6.67 / 5.56%, HHI = 0.1440, effective N 6.947, a 12.96% gain. WE5: σ_p² = 0.029925, σ_p 17.30%, component contributions 0.012825 / 0.010688 / 0.006413 summing exactly to σ_p², risk shares 42.9 / 35.7 / 21.4%. §4.4's India switch: keep X ₹230.8, switch ₹289.1, and at a 10% Y ₹231.5, the dead heat the prose describes. WE2's US switch: $293.87 / $361.98 / $295.96. §4.2's diworsification arithmetic: 30 names average 11.0% expected return against ten names' 15%, for a 1.19-point SD saving. Everything in the prose above reproduces exactly.

Quantitative lens

Two results, both computed on the module's own parameterisation (σ = 30%, equal weights) and on a simulated 30-name world whose true covariance matrix is known by construction. The second is a statement about estimators, not about any market.

Correlation governs the curve; count barely touches it past the first handful. The marginal name is the cleanest way to see it. At σ = 30% and ρ̄ = 0.30, the third name removes 227.8 basis points of portfolio SD, the tenth removes 19.1 bp, the twentieth 4.8 bp, and the thirtieth 2.1 bp. The third name is worth 107 times the thirtieth. Going from one name to ten buys 11.75 points of SD; ten to thirty buys 1.19; thirty to infinity buys 0.63. Now hold the count fixed and move the correlation instead: five names at ρ̄ = 0.10 sit at 15.87%, while one hundred names at ρ̄ = 0.50 sit at 21.32% and can never go below their floor of 21.21%. Five genuinely different businesses beat a hundred similar ones, and no amount of counting closes the gap, which is the same fact §4.3's "independent-bets cap" states in words and M9.01 §4.10 states as N_eff. The practical instruction is uncomfortable for a research process organised around finding more ideas: after roughly the tenth name, additional diversification is bought by lowering ρ̄, not by raising N, and lowering ρ̄ means owning businesses whose fortunes genuinely do not share a driver, a much harder research problem than finding another good company.

The optimiser's "optimal" weights are mostly estimation error, and this is why equal weight is hard to beat. §4.2 and WE5 treat the covariance matrix as known. It never is. A 30-name covariance matrix has N(N+1)/2 = 465 free parameters; five years of monthly returns gives N × T = 1,800 observations, or 3.87 data points per parameter, and the estimation noise does not average out. It concentrates in exactly the directions the optimiser leans on. Simulate it: build a true one-factor world for 30 names, draw a 60-month history, estimate the covariance matrix, solve for the unconstrained minimum-variance weights, and score them against the true covariance. Median out-of-sample volatility is 21.27%/yr. Naive equal weighting, which uses no estimate at all, delivers 18.57%. The true optimum is 15.32%. So the optimiser does not merely fail to capture the available 3.25 points of variance reduction. It gives back 2.70 points relative to doing nothing. Add estimated means and it becomes farce: the estimated tangency portfolio delivers 87.42%/yr, because expected returns are the noisiest input in finance and the optimiser treats them as facts. The instability is visible directly in the weights: across draws, the average absolute weight is 0.0905 against 1/N = 0.0333, 100% of draws contain at least one short position, and the median largest single weight is 0.299 — a 30% position in a thirty-name "diversified" book. Most damning, run the optimiser on two non-overlapping 60-month windows drawn from the same world: the two "optimal" weight vectors correlate at 0.270, and moving from one to the other requires 160.4% one-way turnover. Nothing about the world changed between those windows. That is what "mostly estimation error" means, and it is the quantitative reason §4.3 lands on a written policy and a hard cap rather than an optimiser's output.

Honest limits, and one of them changes the conclusion. The simulation's world is a single-factor model with equal true means and Gaussian returns. Real returns are fat-tailed and correlations rise in stress, both of which make estimation harder, so the figures above are optimistic. But the comparison is also unfair to the optimiser in one specific way that matters: the weights above are unconstrained, and the standard practitioner fix is a no-shorts constraint, which acts as implicit shrinkage. Rerun with a long-only constraint and the picture changes: at 36 months of history the constrained optimiser delivers 17.60% against equal weight's 18.57%; at 60 months, 16.94%; at 240 months, 15.96%. It beats equal weight at every history length tested, while the unconstrained version needs about 120 months merely to draw level (17.54%). The honest statement is therefore narrower and more useful than "optimisers do not work": an unconstrained optimiser fed a short sample is a noise amplifier, the constraint is doing most of the work that the optimisation is credited with, and the returns to estimating better are smaller than the returns to constraining sensibly, which is precisely the argument for §4.8's caps. None of this touches expected returns, where the estimation problem is far worse and where no constraint saves you. And none of it addresses permanent impairment: every number here is about the dispersion of outcomes, and §4.2 already flags that this is not the risk M9.01 taught you to fear.

Do it in code: build the return panel first and check it for corporate actions, compute the correlation matrix over a calm window and a stress window, plot the diversification curve at your book's actual ρ̄ rather than at a textbook 0.30, decompose contribution to risk and compare each name's risk share to its capital share, and, before you ever act on an optimiser, re-run it on a second, non-overlapping window and look at how much the weights moved. See G1/AA1.01, G1/AA1.02 and G2/QD1.04.

Where this goes next: galaxy cross-links

• G1 · AA1.01 Mean-Variance Optimization & the Efficient Frontier. The formal version of §4.2, including exactly where the frontier's inputs come from and why the frontier is drawn with an uncertainty band or not at all.
• G1 · AA1.02 Black-Litterman & Risk Budgeting. Shrinkage, priors and covariance estimators built for short samples, plus risk budgeting, which is WE5's contribution-to-risk turned into the thing you actually set rather than the thing you measure afterwards.
• G2 · QD1.04 Portfolio Optimization in Code. Constrained optimisation as an engineering problem: the no-shorts and cap constraints that, as the lens above shows, are doing more work than the objective function.
• G1 · QM1.02 Hypothesis Testing & Regression. The regression that turns WE3's judgement scores into estimated factor loadings with standard errors, so "am I just a factor?" gets a number and a confidence interval instead of a tally.
• G2 · DA1.02 Pandas Data Wrangling. Assembling the aligned, corporate-action-adjusted return panel that every correlation here silently assumes you already have.

Teach it back & journal

Feynman prompt. Write a one-page explainer for a smart 15-year-old titled "Why owning more stocks isn't the same as being safer — and how a real investor decides what to own." You must convey, without jargon: that safety comes from owning things that don't move together, not just from owning more things (three lottery tickets on the same number are still one bet); that every new thing you buy has to be better than the worst thing you already own, or you shouldn't bother; that holding some cash isn't lazy — it's keeping money ready for when everything goes on sale; and that sometimes the honest answer is "I don't actually know how to beat the market, so I'll just buy a little piece of all of it." If your page needs the words "covariance," "factor," or "hurdle rate," you don't own it yet.

Journal prompt. Three entries. (1) Look at whatever you actually own (or your top watchlist). Group the holdings by what would hurt them at the same time: the same interest rate, the same sector, the same one factor. How many independent bets do you really have, versus how many names? Were you surprised? (2) Take your current weakest holding and one new idea you find attractive. Run the after-tax replace-that arithmetic honestly. Did the tax change your answer, and had you been ignoring that friction before? (3) Write, in one sentence each, the concentration policy you think you want and the one your temperament actually survived (or would survive) in a 40% drawdown, and note the gap. This is the policy you will formalize in the mini-project and live under in the M9.04 simulator season; the honesty of entry (3) is the whole point of Phase 9.

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