Mental Models & Decision Science

Phase 6 · Week 54 · ~20 hours

Where it sits. Third module of the philosophy phase. The two before it gave you the schools: nine coherent ways to find and hold a bargain, each a temperament as much as a technique. But a school tells you what kind of thing to look for. It does not tell you how to decide when the facts are ambiguous, the odds are estimates, and the outcome will only be known years after you commit. What follows builds that operating system: the thinking tools a serious analyst uses to make good decisions under genuine uncertainty and to keep making them for decades without being ruined by variance or by his own mind. Inversion, base rates, expected value, Kelly, Bayes and ergodicity are not finance topics. They are how careful people think about anything consequential, which makes this the most transferable stretch of the program. It is also where the program's deepest spine finally gets stated outright: you are graded on the quality of your process, never on the luck of any single outcome. Everything from the app's process-scored leaderboards to your decision journal to the mastery gates exists to enforce that one idea, and this is where you learn why it is not a slogan but a mathematical necessity.

Suggested hour budget. Core content (the twelve tools): 8h · worked examples (outside-view forecast, EV table, Kelly sizing, Bayesian update): 3h · practice set (EV, Kelly, Bayes, inversion, ergodicity): 4.5h · applied mini-project (your circle of competence + a full decision workup): 2.5h · flashcards + mastery check: 2h.

App link. This module feeds the Philosophy & Behavior branch and wires directly into the machinery you'll use forever: the expected-value calculator and Kelly sizing sandbox (shared plumbing with the Phase 9 position-sizing tools), the Bayesian-update widget in the decision journal (enter a prior and evidence likelihoods, get a posterior; the app makes you do this every time new information hits a live thesis), the pre-mortem template attached to every case-engine thesis, and the calibration dashboard that scores your reasoning against outcomes over time (competency C9). The gate, applied decision problems ≥85%, is delivered as a mixed problem set in the app's assessment engine.

Learning objectives

By the end you can:

1. Explain Munger's latticework of mental models and the "man with a hammer" failure: why a small kit of big ideas from many disciplines, used in combination, out-reasons any single framework applied everywhere. Then start your own list of the models you actually use.
2. Define your circle of competence in writing: sort businesses into inside / edge / outside with honest justification, articulate why the perimeter (not the size) is what matters, and state where your own mistakes are most likely to come from.
3. Run inversion and a formal pre-mortem on an investment: attack a thesis by asking "how would I destroy this business / lose money here?", conduct a structured pre-mortem, and convert the failure modes into things to verify or size for.
4. Apply opportunity cost as your true hurdle: evaluate every buy as a rejection of your next-best idea, set a real hurdle rate rather than comparing to zero, and treat "sell A to buy B" as one decision.
5. Think in second order and locate the variant view: trace "and then what?" past the first consequence, and use the consensus-vs-variant grid to see that excess return requires being both non-consensus and right.
6. Separate risk from uncertainty (Knight): distinguish quantifiable odds from true unknowns, explain why markets and models conflate them, and respond correctly (margin of safety and robustness where you cannot compute probabilities).
7. Use base rates and the outside view: anchor a forecast on the reference class before the story, name the planning fallacy, pull Mauboussin's base-rate data, and blend an inside and outside view into a defensible estimate.
8. Compute expected value and separate decision quality from outcome quality: build an EV table for an asymmetric payoff, avoid "resulting," and explain why process is the only thing you can control and therefore the only thing worth scoring.
9. Size a position with the Kelly criterion and fractional Kelly: derive the formula's intuition, compute it for both even-money and capped-downside bets, and explain precisely why full Kelly is too aggressive in practice (the bridge to M9.01).
10. Update beliefs with Bayes' rule: turn a prior, a piece of evidence, and its likelihoods into a numeric posterior; explain base-rate neglect; and run your decision journal as a Bayesian instrument ("strong views, weakly held").
11. Reason about ergodicity and avoid ruin: explain why the time-average of a multiplicative process differs from its ensemble-average, why a 50% loss needs a 100% gain, why volatility drags the compound rate below the arithmetic mean, and why survival is therefore a first-order goal, not caution.

Prerequisites & connections

Builds on. M6.01 and M6.02 supply the raw material that gets organised here, and three pieces of M6.02 carry straight over. Pabrai's Dhandho asymmetry turns into expected value and Kelly. Marks's second-level thinking and his warning that "risk is highest when it feels lowest" turn into second-order thinking and the line between risk and uncertainty. And the concentration-versus-diversification debate finally gets its mathematical spine from Kelly and ergodicity. M3.01 to M3.03 give you the time value of money, the cost of capital and the ~12% INR / ~8 to 9% USD hurdle (as of mid-2026, check against your own WACC builds), which carry the opportunity-cost and hurdle-rate sections. M3.05's reverse DCF ("what is the price implying?") is the same instrument the variant-view and base-rate sections use, because the market's implied number is the consensus you have to beat. M2.07's Beneish M-Score and the other quant screens are the noisy diagnostic tests that Bayes teaches you to read without being fooled. And M4.05's mean reversion of ROIC and the Competitive Advantage Period are a base-rate story. "Great returns fade" is the outside view of any high-ROIC business.

Feeds forward. M6.04 is behavioural finance. What you learn here is the normative theory, how a rational decider should reason, and M6.04 is the descriptive one, how real human minds actually drift from it. The latticework's psychology models (Munger's 25 tendencies, the Lollapalooza effect) wait for M6.04 on purpose, and every tool here works as a pre-commitment device against a bias catalogued there. M6.05 is the checklist and temperament capstone. Your written circle of competence, your pre-mortem habit and your base-rate habit become standing lines in your personal checklist and in the capstone Mental-Model & Moat Checklist. M9.01 and M9.02 cover risk and position sizing. There, fractional Kelly and the ergodicity case for avoiding ruin are re-derived alongside hidden correlation, and turned into a live sizing policy. You get the theory now, and Phase 9 is the practice. M9.04 and the decision journal (installed back in M0.04) treat the journal as a Bayesian instrument and a calibration record, and process over outcome is the rule it enforces. In Phase 7, the gap between risk and uncertainty is why the macro modules end in humility. Macro is mostly Knightian, so you build robustness instead of forecasts. And competency C9 ("investor's temperament and process") is built here more than anywhere else.

In brief. Good investing is a long run of decisions made on incomplete information, with feedback that comes late and noisy. You win that game by reasoning with a kit of models from many disciplines and weighing outcomes by expected value, without ever mistaking a good decision for a good result. You size so that variance can never ruin you, and you update your beliefs honestly as evidence arrives. You only get to compound if you survive, and you only get better if you grade the process.

Core teaching content

4.1 From schools to decision science: the layer beneath the philosophy

Real decisions never arrive as "this is worth ₹100 and trades at ₹60, buy." They arrive as: there is maybe a 60% chance this is worth somewhere between ₹80 and ₹130, it trades at ₹70, the CEO said something reassuring but the channel checks are mixed, a competitor just raised money, and you have three other ideas competing for the same rupee.

Phases 1 through 5 taught you to answer what is this business worth? Phase 6's schools taught you to ask what kind of bargain am I hunting, and do I have the temperament to hold it? Neither one, on its own, tells you how to actually decide. That's the job of decision science, the layer beneath the philosophy. It gives you tools for turning messy, probabilistic, incomplete information into a sized action you can defend, and for getting better at it over a career.

Everything turns on one shift in thinking. An amateur asks "Was I right?" and checks the stock price. A professional asks "Was my reasoning sound, given what I could know at the time, and did I size it so being wrong couldn't hurt me too badly?" and checks the process. The price is the outcome. The reasoning and the sizing are the decision. In a domain thick with luck and thin with feedback, one decision's outcome tells you almost nothing about its quality. Organise your life around outcomes and you'll learn the wrong lessons, like the "lesson" that a reckless bet which happened to pay off was smart. Only organising it around process compounds skill. And this isn't soft advice. Once you reach expected value and ergodicity, you'll see that the mathematics of probability and of multiplicative returns forces it on you. Every tool below serves it.

The frame for the whole kit comes from Charlie Munger, so we start there.

4.5 Opportunity cost: your true and only hurdle

Munger again: "Intelligent people make decisions based on opportunity costs." Every rupee or dollar you put into idea A is one you didn't put into your next-best idea, B. So asking whether A is a good investment in the abstract gets you nowhere. Ask whether A beats B, your best available alternative, once you adjust for risk. It sounds obvious, and people violate it constantly, because the alternative is invisible. The cost of owning a mediocre 6% compounder appears on no statement, but it's brutally real. It's the 12% compounder you could have owned instead. Over twenty years, compounding at 6% versus 12% is roughly a threefold difference in ending wealth, ≈3.2× against ≈9.6× your money, so the 12% path ends about three times richer. Opportunity cost is the most important number in investing that never appears in any filing.

That reframes a few things you already know. Your hurdle rate is your best available opportunity. It isn't zero, and it isn't the bond yield either. If you can reliably find businesses that compound at 12% with acceptable risk, a new idea offering 9% is a bad decision, even though 9% is a "good return," because it's worse than what you can already get. Great investors with a deep bench of ideas are demanding for this reason. Their hurdle is high because their alternatives are good. Opportunity cost also dissolves the artificial split between buying and selling. "Should I sell A?" and "should I buy B?" are the same decision, which comes down to whether you'd rather own A or B with this capital. If B is meaningfully better per unit of risk, then holding A is the decision to forgo B, whether or not you frame it that way. You'll formalise this sell logic later in the course.

This ties straight back to Phase 3. The cost of capital, WACC, is the market's version of an opportunity-cost hurdle: the return available on alternatives of similar risk. The master test of value creation, ROIC > WACC, the spine of everything you've built so far, is an opportunity-cost statement too. A business creates value only when it earns more than its capital could earn elsewhere at that risk. Your personal portfolio hurdle is the same idea, applied to your own opportunity set. In practice, keep a ranked "bench" of ideas with their expected returns. Then require any new position to beat the marginal idea it would displace, which is a higher bar than clearing zero. Cash is an opportunity-cost decision as well. Holding it is right when nothing on the bench clears the hurdle. That's Klarman's approach. Holding it out of fear is wrong. The habit to build is making the invisible alternative visible, every time.

4.6 Second-order thinking and the variant view

Howard Marks, whom you met with the schools, draws a line between first-level and second-level thinking. It's the same line that separates people who beat the market from people who are the market. First-level thinking stops at the first consequence. "Good company, so buy it." "Rates are rising, so sell bonds." "The company beat earnings, so the stock goes up." Second-level thinking keeps asking and then what? It asks what everyone else already believes and whether that's already in the price. Then it asks how the chain of consequences actually plays out. "Good company, but everyone knows it, so it's overpriced. Avoid." "Rates are rising, but the whole market has already positioned for it and the last hike is in the price, so maybe buy the long bond." Second-order thinking is just refusing to stop at the first link of a causal chain. The money and the danger are in the second and third links.

Two separate skills hide inside it. The first is literally tracing the consequences of consequences. Say a commodity producer cuts prices to gain share. The first-order effect is that volumes rise, which looks good. The second-order effect is that rivals match, a price war breaks out and the whole industry's profit pool shrinks, leaving everyone worse off, the initiator included. That's the capital cycle seen from the demand side. Or a generic-drug maker wins a big product. Revenue jumps at first. Then competitors file their own versions, price erosion sets in, and next year's comparison is brutal. Or a country subsidises an industry. The industry booms, then capacity floods in and returns collapse, and the subsidy has created the glut. Training yourself to ask "and then what? and then what?" two or three links deep is most of what separates a naive read from an analytical one.

The second skill is finding the variant view. Being right isn't enough to earn excess return. You have to be right when the consensus is wrong, which means holding a view that's both correct and non-consensus. Lay it out as a 2×2 of your view against the market's:

Only one cell pays excess return, and it's variant and correct. The reverse DCF turns this into a working tool. It pulls out the consensus embedded in the price: the growth and margins the market is assuming. Then you can ask the question that decides it. Do I have a well-founded, differentiated reason to believe the truth is materially better or worse than what's priced? If your view matches the market's, you have no edge, however "good" the company is. If your view is merely different and no better founded, you're the sucker. Beginners resist the uncomfortable implication. A wonderful business can be a terrible investment, and an ugly business a wonderful one, depending entirely on the gap between reality and what the price already assumes. Analysing the business tells you the truth. The variant grid tells you whether the truth is news. You need both.

4.8 Base rates and the outside view

Kahneman and Tversky's most useful gift to an analyst is the distinction between the inside view and the outside view of a forecast. The inside view builds a prediction from the specifics of the case in front of you: this company, this management, this plan, this addressable market. The outside view sets the narrative aside at first and asks for the base rate. What reference class does this case belong to, and what usually happens to members of that class? Left to ourselves, we're wildly biased toward the inside view. The story is vivid and specific, and it seduces us. We keep ignoring how the broad class of similar cases actually turned out. The result is the planning fallacy. It's a well-established finding that plans are consistently and predictably too optimistic. So are forecasts, construction timelines and corporate projections. Each one is built from the inside, as if it were uniquely destined to beat the odds that governed all its predecessors.

The correction is mechanical, and it works. Anchor on the base rate first, then adjust for information that really is specific to the case, and adjust less than your gut wants to. Kahneman's startup founder, the one who "just knows" his venture will succeed, is the classic case. The base rate of startup survival should anchor the estimate, and the founder's particular quality should move it only modestly. In investing, base rates are everywhere once you look. The fraction of companies that sustain 20%+ revenue growth for a decade is small. The fraction of high-ROIC businesses still earning high returns ten years later shrinks as returns fade, which is the mean reversion you studied with moats. Large acquisitions have a poor base rate of creating value for the acquirer. The base rate of turnarounds that actually turn is low. "Turnarounds seldom turn," as Buffett puts it. Michael Mauboussin has done the analyst an enormous service here. His Base Rate Book, a free Counterpoint Global / Credit Suisse publication, compiles the empirical distributions of sales growth, margin change and ROIC persistence by starting size and starting level. So you can look up the actual reference class instead of guessing it. Download the current edition, and don't freeze the figures from any summary.

A worked outside-view forecast. Take an Indian consumer-durables company on the scale of a Titan or an Asian Paints. Treat the numbers as illustrative. It has grown revenue ~18% a year for five years, and management guides to "sustained ~18 to 20% growth for the next decade." The inside view makes a compelling story: strong brand, rising incomes, category tailwinds, a proven management team.

Now take the outside view. Pull the base rate Mauboussin-style, or build your own from screener.in histories of the fifty largest Indian consumer names. Of companies already at this revenue base, say ₹30,000 to 40,000 crore, what fraction grew 18%+ for a full decade? Illustratively, it's a single-digit-percent event. The median large consumer company grows more like high-single to low-double digits over ten years, because the law of large numbers and market saturation bite. So the reference class says ~9 to 11%, and the story wants ~19%. A careful estimate doesn't naively split the difference, and it doesn't just adopt the story either. It anchors near the base rate and adjusts modestly upward for this company's above-average quality (a real brand and runway). That gets you to perhaps 12 to 14%, with a wide band.

Then close the loop. Run the reverse DCF and read the growth the price implies. If the market is pricing 18 to 20%, the price embeds the inside view, and the base rate says that's a poor bet. If it's pricing 8%, the base rate says the price is too pessimistic, and you may hold a variant view that's also correct. The base rate kept your judgment in check without replacing it, and it told you which side of the consensus the evidence actually supports.

Two cautions. Choosing the reference class is itself a judgment call. Make it too broad ("all companies") and it's useless. Make it too narrow ("companies exactly like this one") and it collapses back into the inside view. Aim for the tightest class that still has enough members to give a stable base rate. And base rates are where you start. They aren't the verdict. A really exceptional case can beat its class, but the burden of proof is on the specific evidence, and you should demand a lot of it before straying far from the reference class.

4.10 The Kelly criterion and fractional Kelly

Expected value ranks bets, and the Kelly criterion sizes them. It answers one question. Given an edge, what fraction of your capital should you stake to maximise the long-run growth rate of your wealth? John Kelly derived the answer at Bell Labs in 1956, and Ed Thorp made it famous in investing. Poundstone's Fortune's Formula is the readable history. The answer is beautifully simple, and it deserves to be understood as well as quoted.

The intuition. Bet too little and you leave growth on the table. Bet too much and volatility eats you. Push it far enough and a single bad run wipes you out, ending all future compounding. The last tool, ergodicity, explains why that asymmetry is so violent. Kelly is the unique fraction that balances these. It maximises the expected logarithm of wealth, which is the same as maximising the long-run compound growth rate. Take a simple bet that pays net odds b on a win, so you win b per unit staked. The win probability is p, the loss probability is q = 1 − p, and on a loss you lose your whole stake. Then the Kelly fraction is:

f* = (b·p − q) / b

You can remember it as edge / odds. The numerator (b·p − q) is your expected profit per unit staked, the edge. Dividing it by the odds b scales it into a fraction of capital. Here's a worked case. A bet wins +100% (b = 1) with p = 0.60 and loses 100% with q = 0.40, so f* = (1×0.60 − 0.40)/1 = 0.20. Full Kelly says stake 20% of capital.

The investing generalisation. Real investments rarely wipe out the whole stake on a loss. They fall by some fraction. Say a position gains fraction u of the amount invested on a win, with probability p, and loses fraction d on a loss, with probability q. Then the Kelly fraction of wealth to allocate is:

f* = p/d − q/u

When d = 1, a total loss, this reduces to the gambling formula. Check it: you get p/1 − q/u, and with u = b you recover (bp − q)/b. This form carries the most important sizing lesson here. Take a position that could gain +100% (u = 1) with p = 0.60, and watch what happens to the size as we cap the downside. If the loss is a total wipeout (d = 1), f* = 0.60/1 − 0.40/1 = 0.20. Now cap the downside at −50% (d = 0.5). The edge and the upside are the same, but the loss is survivable, and f* = 0.60/0.5 − 0.40/1 = 1.20 − 0.40 = 0.80. Capping the downside quadrupled the Kelly size, from 20% to 80%. That's the mathematics behind Pabrai's obsession with "heads I win, tails I don't lose much." Asymmetric, capped-downside bets can safely be sized far larger. So everything rides on checking that the downside floor is real. A floor that's really hiding leverage or fraud is an illusion.

Why full Kelly is too aggressive, and fractional Kelly. Look at the raw numbers above, 20% and 80%. Full Kelly is savage, and there are three reasons no thoughtful investor bets it:

1. You don't know the true odds. Kelly assumes p, u, and d are known. In investing they're subjective estimates, and Kelly punishes you hard for overestimating your edge. Bet as if your edge is 60/40 when it's really 52/48 and you'll overbet systematically until you're eventually ruined. The penalty is lopsided. Underbetting Kelly costs you some growth. Overbetting it destroys growth and can bankrupt you, because past a point more size means lower long-run growth, and then negative growth. With an uncertain edge, you have to err low.
2. The drawdowns are unbearable. Full Kelly tolerates gut-wrenching interim losses, and a full-Kelly bettor routinely sees his capital halve. Almost no human, and no client, can hold a position through drawdowns like that. Being forced to sell at the bottom turns a mathematical optimum into a real-world catastrophe.
3. Real positions are correlated. Kelly for a single bet ignores that your positions share factors, the hidden-correlation problem from the concentration debate. Add up full-Kelly sizes across correlated ideas and you massively overbet the portfolio's true independent risk.

The practical answer is fractional Kelly. You bet a fixed fraction of the Kelly number. Half is common, and sometimes it's a quarter. Half-Kelly captures roughly three-quarters of full Kelly's growth with far lower volatility and drawdown, and it builds in a margin of safety against overestimating your edge. So the 20% full-Kelly bet becomes a ~10% or ~5% position. The 80% one gets capped hard, by the fraction and by sane concentration limits. For an investor, Kelly is less a precise position size to three decimals than a principle that keeps you in line. Size in proportion to your edge and inversely to your risk of loss. Always err smaller than the formula. And never let the math tempt you past the size a bad outcome can survive. Later you'll turn this principle into a live sizing policy. For now, own the logic and the two formulas.

Common mistakes & how experts think differently

Resulting: judging the decision by the outcome. The novice looks at a stock that fell and concludes "bad call," and at one that rose and concludes "good call." The expert knows that in a domain this noisy, one outcome carries almost no information. A well-reasoned, well-sized bet that lost money was probably still the right decision, and a reckless bet that tripled was probably still wrong. Experts grade the process. They ask whether the odds were honest, whether the EV was positive and whether the size was survivable. They can say "I lost money and it was the right decision" and mean it. They can say "I made money and it was a bad decision" and learn from it. It's the hardest habit here and the most important. Everything else degrades without it.

Confusing a great business with a great investment. The beginner reasons "wonderful company, so buy." The expert runs the variant grid and the reverse DCF. A wonderful company whose wonderfulness is fully in the price is a market-return holding at best, and a loss if expectations are too high. An ugly company priced for catastrophe can be a superb investment if reality is merely bad. So don't ask "is it good?" Ask "is it better than the price assumes, and does the consensus disagree with me?"

Ignoring the base rate because the story is vivid. The inside view is seductive and the outside view is boring, so untrained analysts forecast every promising company as an exception. The expert anchors on the reference class first. How many companies this size sustain 20% for a decade? Very few. Then the expert demands extraordinary evidence before straying far from it, and treats management guidance as an inside view to check against base rates. Guidance isn't data. "This time is different" is the most expensive sentence in finance because it's a refusal to look at the base rate.

Thinking in point estimates instead of distributions. Amateurs say "it's worth ₹100." Professionals say "there's maybe a 25% chance it's worth under ₹70, a 50% chance ₹90 to 120, and a 25% chance over ₹150, so the EV is ~₹105 but the downside tail is real." Valuation is a range, and that's a guardrail across the whole course. The shape of the distribution, especially the left tail, drives sizing more than the central estimate does.

Treating uncertainty as if it were measurable risk. The dangerous analyst pins a confident probability on an unknowable event (a technology transition, a policy regime change, a first-of-its-kind product) because a number feels rigorous. The expert asks "is this risk I can estimate, or uncertainty I'm pretending to estimate?" Where it's uncertainty, the expert responds with margin of safety, robustness and small size instead of false precision. The confidence of a precise-looking model is inversely related to its reliability in the tails.

Over-sizing the good idea. Having found a real edge, the untrained investor bets the farm. Eventually an unlucky run ruins them, because they optimised the arithmetic mean and ignored the geometric one. The expert sizes with fractional Kelly and remembers that overbetting punishes far more than underbetting. The expert also accounts for correlation across positions, and never lets a single outcome, however attractive its EV, put the portfolio within reach of ruin. Getting rich slowly requires not going broke first.

Comparing to zero instead of to the next-best idea. The beginner asks "will this make money?" and buys anything with a positive answer. The expert asks "is this better than what I'd have to give up to own it?" They keep a ranked bench and hold cash without guilt when nothing clears the hurdle. A "good" 8% idea is a bad decision if a 12% idea of equal risk is on offer.

Updating too slowly, or not at all (and its opposite). Ego-defence makes some people cling to a thesis while disconfirming evidence piles up. "The market just doesn't get it yet." Overreaction makes others abandon a sound thesis on one noisy data point. The Bayesian expert stays in between. They move their probability in proportion to the strength of the evidence, neither anchoring stubbornly nor flip-flopping. And because they wrote down in advance what would change their mind, they notice when it happens.

Staying at the boundary of the circle. You already know to avoid being outside your competence. The trap is being at its edge while feeling like you're inside it. Experts are almost eager to say "I don't understand this well enough" and pass. The number of things Buffett and Munger put in the "too hard" pile is a feature of their record. It isn't a gap in it. "No" is a complete analysis.

Worked examples

These four are the core computations, each carried end to end. One is Indian (WE1) and one is US (WE2), while WE3 and WE4 are currency-neutral. Work each one with a pen before you read the solution.

WE1: The outside-view forecast (Indian consumer company)

Setup. The figures are illustrative, so pull the real filings and build your own base rate. A large Indian consumer company, one you can model on the ~₹35,000 crore revenue scale of a Titan or Asian Paints, has compounded revenue at ~18%/yr for five years. Management guides to "sustained ~18 to 20% growth for the coming decade," and the sell-side has largely adopted that inside view. Your job is a ten-year growth estimate you can defend.

Step 1: State the inside view and why it's seductive. Strong brand, premiumisation, rising Indian incomes, category runway, a proven capital-allocating management. Every point is true, and together they feel like a guarantee of ~19%.

Step 2: Define the reference class and pull the base rate. The right class is "companies already at roughly ₹30,000 to 40,000 crore revenue." That means large companies. Small-caps don't belong in it, and neither does "all companies." Use Mauboussin's Base Rate Book, which is US-heavy but directionally global (download the current edition), plus a hand-built history of the fifty largest Indian consumer names on screener.in. Among companies at this scale, sustaining 18%+ revenue growth for a full ten years is a single-digit-percent event. The median large consumer company lands around high-single to low-double-digit ten-year growth, because saturation and the law of large numbers bite as the base gets large. Call the base rate ~9 to 11%.

Step 3: Anchor, then adjust modestly. Anchor on ~10%, the reference class. Adjust upward for this company's above-median quality (a real brand, a demonstrated reinvestment runway, category tailwinds). Keep the adjustment modest, though, because the base rate already includes plenty of good companies that still slowed. Land at ~12 to 14%, with a wide band (say 9% bear / 13% base / 17% bull). That's well short of the ~19% the story wants.

Step 4: Close the loop with the reverse DCF. Read the growth the price implies. If the market is pricing ~18 to 20%, the price embeds the inside view, which the base rate says is unlikely. That's a poor entry even for a wonderful business. There's no variant edge, and expectations are too high. If the market is pricing ~8%, the price is below your base-rate-anchored estimate, and you may have a variant view that's correct and bullish. The base rate doesn't decide for you, but it shows which side of the consensus the evidence supports and how much of the story you're being asked to pay for.

Takeaway. The outside view turns an unanchored story ("great company, buy") into a falsifiable estimate with a band, and hands the last step to the reverse DCF. Do this on every growth stock before you fall in love with it.

WE2: Expected value, asymmetry, and resulting (US special situation)

Setup. A US specialty-pharma stock trades at $40 ahead of a binary FDA decision about 4 months out. Here are your base-rate-informed estimates:

Step 1: Compute EV and expected return. EV = 0.20×120 + 0.30×55 + 0.50×25 = 24.0 + 16.5 + 12.5 = $53.0. Against the $40 price, expected return = 53/40 − 1 = +32.5%. The edge is real, and it's positive.

Step 2: Read the shape as well as the average. The modal outcome, the most likely one at 50%, is a loss of 37.5%. The positive EV is carried entirely by a 20% tail that pays +200%. It's an asymmetric, capped-downside-ish bet. It's attractive on EV because the up-case dwarfs the down-case, even though you lose more often than you win.

Step 3: Size it with Kelly in mind. Don't bet the portfolio. A 50% chance of −37.5% at large size is a portfolio-wrecking risk. Treat it as one modest, uncorrelated position among many such bets. The +EV only compounds in your favour if you make many of them and no single one can hurt you badly. A rough Kelly on u = 2.0, d = 0.375, p = 0.20 for the big outcome, or a blended calc, comes out at a small full-Kelly fraction. Fractional Kelly and sane single-name limits cut it further. Size for survival across repetition. You aren't sizing for this one flip.

Step 4: The resulting test. Suppose the drug is rejected and you lose 37.5%. Was it a bad decision? No. If the 20/30/50 odds were estimated in good faith, it was a good decision with a bad outcome, and you should take the identical bet every time it's offered. Now flip it. If the drug had been approved and you'd tripled your money, that wouldn't have proven the decision brilliant. It would have been a good decision with a good outcome, no different in quality from the losing case. Grade the process (honest odds, positive EV, survivable size), journal it before the result, and refuse to let the outcome rewrite the decision's quality. That refusal is the spine of everything you're learning.

WE3: Kelly sizing and the value of a capped downside

Setup. Here are two versions of the same opportunity. Both win +100% (u = 1.0) with probability p = 0.60, and they differ only in the downside.

Case A (total loss on the downside, d = 1.0). With the gambling form and net odds b = 1, f* = (b·p − q)/b = (1×0.60 − 0.40)/1 = 0.20. Full Kelly is 20% of capital. The general form agrees: f* = p/d − q/u = 0.60/1.0 − 0.40/1.0 = 0.20.

Case B (downside capped at −50%, d = 0.5). The edge and the upside are the same, and the loss is survivable: f* = p/d − q/u = 0.60/0.50 − 0.40/1.0 = 1.20 − 0.40 = **0.80**. Full Kelly is 80% of capital.

Step: Interpret. Capping the downside from −100% to −50% quadrupled the Kelly size, from 20% to 80%, even though the edge and the upside were identical. That's the mathematics under Pabrai's Dhandho. Asymmetric, capped-downside bets can safely be sized far larger, provided the floor is real. So for any concentrated bet, your first analytical job is checking that the downside floor holds. That means asset backing, balance-sheet strength, no fraud, no leverage that forces selling. A fake floor turns an 80% "Kelly-justified" position into a route to ruin.

Step: Apply fractional Kelly. You wouldn't actually stake 20% or 80%. Your p is an estimate, and full Kelly's drawdowns are unbearable, so don't go above half-Kelly. That takes Case A to ~10% and Case B to ~40%. Case B then gets capped further by a hard single-name limit (say 15 to 25%) and by its correlation with everything else you own. Half-Kelly keeps roughly three-quarters of the growth with far smaller drawdowns and a buffer against overestimating your edge. Kelly gives you the ceiling and the ranking logic. Prudence and correlation set the actual size, and it's always below the formula.

WE4: A Bayesian update on a live thesis

Setup. Your prior belief that the company has a durable cost moat is P(H) = 0.60.

Update 1: confirming evidence E. During an input-cost spike, the company held gross margin flat while a smaller rival's margin collapsed. The likelihoods are P(E | moat) = 0.80 and P(E | no moat) = 0.30.

P(H | E) = (0.80 × 0.60) / [(0.80 × 0.60) + (0.30 × 0.40)] = 0.48 / (0.48 + 0.12) = 0.48/0.60 = **0.80**

Belief rises from 60% to 80%. It didn't jump to certainty, because the evidence was more likely under the moat but still possible without it.

Update 2: disconfirming evidence D. A quarter later the company cut prices to defend share, a sign of weaker pricing power. Take the 80% as the new prior, with P(D | moat) = 0.20 and P(D | no moat) = 0.70:

P(H | D) = (0.20 × 0.80) / [(0.20 × 0.80) + (0.70 × 0.20)] = 0.16 / (0.16 + 0.14) = 0.16/0.30 = **0.53**

Belief falls from 80% to 53%. This is a thesis that breathes. It goes up on confirming evidence and down on disconfirming evidence. It never leaps to certainty, and it always moves in proportion to how diagnostic the evidence is. Record each update in the decision journal against the "what would change my mind" you wrote at entry. Do that until it's a habit, and over a career it's what calibrates you (competency C9).

The base-rate footnote (why a screen flag ≠ guilt). Suppose that instead of a moat you're testing "is this a manipulator?" A quant screen, the Beneish M-Score, flags the company. The test catches 80% of true manipulators (P(flag | manip) = 0.80) with a 10% false-positive rate (P(flag | clean) = 0.10). But the base rate of manipulators is only 5%:

P(manip | flag) = (0.80 × 0.05) / [(0.80 × 0.05) + (0.10 × 0.95)] = 0.04 / (0.04 + 0.095) ≈ **0.30**

Only ~30%. Clean companies are so common that a 10% false-positive rate throws off more false alarms than the rare true positives do. A flag is a reason to dig, never a verdict. Base-rate neglect is what makes untrained intuition scream "fraud" at a number that really means "one-in-three, go investigate."

WE5: Blending the inside and outside view into one number (India)

Setup (synthetic, with illustrative figures). Build your own base rate from real filings. Deccan Consumer Ltd (synthetic), a large Indian consumer company at ~₹35,000 crore revenue, has compounded revenue ~18%/yr for five years. Your bottoms-up inside-view model (brand, premiumisation, distribution runway) also lands near 18% for the coming decade. The outside-view base rate for companies already at this revenue scale is a ten-year CAGR of about 9%, from Mauboussin's Base Rate Book plus a hand-built screener.in history of the fifty largest Indian consumer names. WE1 told you to "anchor on the base rate and adjust modestly." Here, "modestly" becomes a number.

Step 1: write the blend as a weighted average. A defensible estimate is estimate = w × inside + (1 − w) × base, where w is the weight you put on your differentiated, case-specific evidence. All the judgment goes into w. The base-rate class already contains plenty of strong, well-managed companies that still decelerated, so a defensible w is low. You aren't the first analyst to notice a good brand.

Step 2: compute the band.

• Real, differentiated evidence, w = 0.35: 0.35 × 18 + 0.65 × 9 = 6.30 + 5.85 = **12.15%**.
• Thin evidence, w = 0.20: 0.20 × 18 + 0.80 × 9 = 3.60 + 7.20 = **10.8%**.
• Even a naïve 50/50 split, w = 0.50: 0.50 × 18 + 0.50 × 9 = **13.5%**. Still nowhere near the 18% story.

So the defensible estimate is roughly 11 to 13%, with ~12% as the base case. The story's 18% is off the top of that band even under the most generous weighting.

Step 3: price the gap (why it matters). Over ten years, revenue compounds to 1.18¹⁰ = 5.23× on the story against 1.12¹⁰ = 3.11× on your estimate. So the story assumes year-10 revenue about 68% higher (5.23 ÷ 3.11 = 1.69). A price built on the guidance is asking you to pay for that 68%, and the base rate says it's a single-digit-percent event.

Step 4: close with the reverse DCF. Read the growth the price implies. If it implies ~18%, you're buying the inside view the base rate rejects, with no variant edge, so pass. If it implies ~8%, the price is below even your base-rate-anchored estimate, and you may have a variant, bullish view you can defend. The blend kept the judgment and forced it into one visible knob, w, which makes the story's cost explicit.

Takeaway. "Anchor and adjust modestly" becomes w × inside + (1 − w) × base with a deliberately low w. Report a band, then let the reverse DCF say which side of consensus you're on.

WE6: Quantifying a pre-mortem: the probability of ruin from independent kill-shots (US)

Setup (synthetic). Cordero Materials Inc. (synthetic), a US single-plant specialty-materials maker at ~$800m revenue with net debt near 3.5× EBITDA, trades at a tempting multiple. Run the pre-mortem, but don't stop at a list. Put rough probabilities on each independent failure mode over a five-year horizon:

Step 1: the first-level read (the trap). Taken one at a time, every risk looks survivable. "Only 15%… only 20%… only 10%… only 8%." A first-level thinker dismisses each in turn and buys.

Step 2: invert and combine. The business lives only if it dodges all of them. Treating the modes as roughly independent, P(survive all) = (1 − 0.15)(1 − 0.20)(1 − 0.10)(1 − 0.08) = 0.85 × 0.80 × 0.90 × 0.92 = **0.563**. So the probability that at least one kill-shot lands is 1 − 0.563 = **0.437, nearly a coin flip.** Risks that each look minor compound into a ~44% chance of disaster.

Step 3: attack the biggest terms. The arithmetic ranks your diligence for you. The two largest terms, the 20% customer concentration and the 15% refinancing wall, dominate. Suppose work could cut each to 5%, with a diversified customer base and the debt termed out. Then P(survive) = 0.95 × 0.95 × 0.90 × 0.92 = **0.747**, and ruin drops from ~44% to ~25%. That ~19-point gap is the value of the actions the pre-mortem generated, and it names the two facts to verify before you size anything.

Step 4: the honesty caveat (why the real number is worse). Independence is a convenience. It isn't a fact. The failure modes are positively correlated in the tail, because a recession fires the funding shock and the customer loss together. So the true joint probability of ruin is higher than 0.437, and it's higher right when it hurts most. That's the hidden-correlation lesson from the concentration debate, and it comes back when you size positions. Quantifying a pre-mortem turns "these each seem unlikely" into "the combined chance of ruin is ~44%, and probably worse." Ergodicity demands that honesty, because ruin is an absorbing state and can't be treated as just another row in a table.

Takeaway. Inversion plus one line of probability arithmetic turns a vague risk list into a survival probability. It ranks the risks by how much each one moves it, and it exposes the sizing implication a first-level read hides.

Practice set

Problems 1 to 4 are guided (framework applied with scaffolding) and 5 to 9 are independent. Problems 10 to 12 are exam-condition: 12 minutes each, closed book. Give numeric answers to the stated tolerance. Currencies are illustrative, and all companies are synthetic unless named. The required set covers EV in P1/P10, Kelly in P4/P11, Bayes in P6/P12 and inversion in P5.

P1 (guided: expected value). A stock trades at ₹60. Over ~2 years you estimate a 30% chance of ₹150, a 50% chance of ₹65 and a 20% chance of ₹30. (a) Compute the EV and the expected return. (b) What is the most likely single outcome, and what does it earn? (c) Does (b) change whether this is a good bet?

Solution. (a) EV = 0.30×150 + 0.50×65 + 0.20×30 = 45 + 32.5 + 6 = ₹83.5. Against ₹60, that's a +39.2% expected return. (b) The modal outcome, at 50%, is ₹65, a mere +8.3%. The most likely single result is a small gain, nowhere near the ₹150 dream. (c) No. You judge the bet by the whole probability-weighted distribution, and the modal outcome is only one part of it. The 30% upside tail carries the +39.2% edge. As long as the odds are sound and the position is sized so the 20% downside (−50%) can't ruin you, the positive EV makes it a good bet, whichever single outcome is likeliest.

P2 (guided: inside vs outside view). A five-year-old SaaS company grew revenue 60%/yr and management guides to "50%+ for five more years." (a) State the inside view. (b) State how you'd build the outside view and roughly what it says. (c) How should the two combine?

Solution. (a) Inside view: the specific story (great product, huge TAM, strong net revenue retention, a visionary founder) projected forward at ~50%. (b) Outside view: define the reference class as software companies that had already reached this ARR scale. Then pull the base rate for sustaining 50%+ for five years from Mauboussin's base-rate work and from screener/stockanalysis histories of comparable names. That base rate is a low-single-digit-percent event, and deceleration as the base grows is near-universal, the "law of large numbers" of revenue. (c) Anchor on the base rate, then adjust modestly up for above-median quality. You'll land well below 50%, perhaps 25 to 35% with a wide band. Demand extraordinary, specific evidence before straying far toward the story. Then read the reverse-DCF implied growth. If the price already assumes 50%, the base rate says pass.

P3 (guided: opportunity cost). You own idea A expected to compound at 7%. Idea B, of similar risk and inside your circle, is expected to compound at 13%. (a) Over 15 years, how much does the choice matter? (b) Frame the decision correctly. (c) When is holding A nonetheless right?

Solution. (a) ₹1 at 7% for 15 years grows to 1.07¹⁵ ≈ 2.76×. At 13% it grows to 1.13¹⁵ ≈ 6.25×, so the 13% path ends about 2.27 times richer. The invisible opportunity cost of owning A is enormous, and it compounds. (b) "Hold A or buy B" is one decision. With fixed capital, keeping A means choosing to forgo B, so the question is simply "would I rather own A or B per unit of risk?" Given equal risk and a 6-point edge, B wins. (c) Holding A is right only if the comparison isn't apples-to-apples. Maybe B carries materially more risk or uncertainty, or lies outside your circle. Maybe it has large frictional or tax costs to switch into, or your 13% estimate for B is far less reliable than your 7% for A. Absent those, the opportunity cost makes holding A the worse decision.

P4 (guided: Kelly, even-money-style). A bet wins +150% (net odds b = 1.5) with probability p = 0.50 and loses the whole stake with probability q = 0.50. (a) Compute full Kelly, f* = (b·p − q)/b (tolerance ±0.01). (b) What would you actually stake? (c) State the two reasons for (b).

Solution. (a) f* = (1.5×0.50 − 0.50)/1.5 = (0.75 − 0.50)/1.5 = 0.25/1.5 = 0.167, so full Kelly is ≈ 16.7% of capital. (b) I'd stake fractional Kelly, either half (~8%) or a quarter (~4%). (c) There are two reasons. First, p = 0.50 is an estimate, and overestimating edge turns full Kelly into a ruin machine, since the penalty for overbetting is far worse than for underbetting. Second, full Kelly's drawdowns are unbearable, and correlated positions compound the portfolio's true risk, so I discount for both.

P5 (independent: inversion / pre-mortem). You are about to buy a rapidly growing Indian NBFC at 3× book on a thesis of "20% AUM growth and 3% ROA for a decade." Run an inversion: it's five years later and the position lost 80%. Write the four most likely failure stories, then convert each into an action.

Solution (model answer; many valid). How it died: (1) A funding shock (ALM mismatch). It funded long assets with short/CP borrowing. When a peer defaulted in an IL&FS/DHFL-style event, its funding froze. It couldn't roll its debt, and a solvent-but-illiquid lender died in weeks. Action: dissect the asset-liability maturity ladder and borrowing mix before buying. Size for this being the real kill-switch, and check that the company isn't on the wrong side of it. (2) Credit costs normalised. The 3% ROA was earned in a benign credit environment. When the cycle turned, provisions spiked and ROA halved. For high-growth lenders, the base rate is that fast AUM growth buys future NPAs. Action: stress the ROA to through-cycle credit costs, and distrust ROA earned only in good years. (3) Growth was quality-blind. The 20% AUM growth came from lending to borrowers others had rejected, so the book was adverse-selected. Action: check whether growth outran the addressable prime market. Then read the disbursement-quality and provisioning notes. (4) Governance or promoter risk. Related-party lending or a promoter share-pledge (an M2/M8 red flag) unwound. Action: read the related-party notes and the pledge disclosure, and treat 3× book as demanding proof of clean governance. The inversion's payoff: it turned a rosy thesis into a list of what to verify, what would change your mind for the journal, and the size the risks really support, all before the market taught you the same lesson expensively.

P6 (independent: Bayesian update). Prior that a thesis is correct: P(H) = 0.50. Evidence E (a strong, hard-to-fake data point) arrives; P(E | H) = 0.90, P(E | not H) = 0.40. (a) Compute the posterior (±0.01). (b) Then disconfirming evidence D arrives with P(D | H) = 0.30, P(D | not H) = 0.75; update again from your new prior. (c) What habit does this illustrate?

Solution. (a) P(H|E) = (0.90×0.50) / [(0.90×0.50) + (0.40×0.50)] = 0.45 / (0.45 + 0.20) = 0.45/0.65 = 0.69. Belief rises from 50% to 69%. (b) With the new prior of 0.69, P(H|D) = (0.30×0.69) / [(0.30×0.69) + (0.75×0.31)] = 0.208 / (0.208 + 0.234) = 0.208/0.442 ≈ 0.47. Belief falls from 69% to 47%. (c) This is a thesis that breathes. It rises on confirming evidence and falls on disconfirming evidence, in proportion to how diagnostic each piece is, and it never leaps to certainty. That's "strong opinions, weakly held," recorded against your pre-committed "what would change my mind."

P7 (independent: ergodicity). A game each round: heads (p = 0.5) multiplies wealth by 1.8 (+80%); tails multiplies by 0.5 (−50%). (a) Compute the ensemble (one-round) average return. (b) Compute the per-round time-average (geometric) growth. (c) Reconcile: should you play repeatedly with your whole wealth?

Solution. (a) Ensemble = 0.5×1.8 + 0.5×0.5 = 0.9 + 0.25 = 1.15, or +15% expected per round. It looks superb. (b) Time-average multiplier ≈ (1.8 × 0.5)^0.5 = (0.90)^0.5 = 0.9487, which is −5.1% per round. Played repeatedly with full wealth, you almost surely go broke. (c) No. A few astronomically lucky all-heads paths drag the +15% ensemble average up. The typical path, the one you experience because you live sequentially, decays at −5.1% compounded. Wealth is multiplicative and non-ergodic. So the geometric (time-average) rate is what matters, and the arithmetic expectation misleads you. A positive-EV game can still be a ruin machine when it's over-sized. The fix is sizing: bet a small fraction instead of the whole stake, which is what Kelly enforces.

P8 (independent: recovery math and drawdowns). (a) A fund falls 60%. What return just restores the original capital? (b) One falls 25%. (c) Explain, using these, why avoiding large drawdowns raises long-run compound returns even at the same average return.

Solution. (a) You need L/(1−L) = 0.60/0.40 = 1.50, so +150%. (b) 0.25/0.75 = 0.3333, so +33.3%. (c) Losses and gains are asymmetric under compounding. A 60% loss needs a 150% gain to recover, and a 60% gain won't do it. So deep drawdowns destroy compounding out of all proportion. Formally, geometric ≈ arithmetic − variance/2, which makes volatility a direct drag on compound returns. Two strategies with the same arithmetic mean and different volatility end at different wealth, and the smoother one wins. Cutting drawdown depth does more than ease your nerves. It mechanically raises the geometric rate you actually compound at.

P9 (independent: second-order thinking and the variant view). A beloved, high-quality US company (call it a 40% ROIC compounder) trades at 38× earnings; consensus is glowing. (a) Give the first-level and second-level thoughts. (b) Use the consensus-vs-variant grid. (c) Name the single test that decides the case.

Solution. (a) First-level: "wonderful company, buy it." Second-level: "wonderful, but everyone knows it. At 38× the price already embeds many years of flawless compounding, so unless reality beats an already-optimistic bar, it's an overpriced stock even though it's a great business." (b) On the grid, you earn excess return in only one cell, variant and correct. If your view ("it will compound beautifully") equals the consensus, you have no edge, however good the company is. You'd have an edge only with a well-founded, differentiated view, say that the moat is even more durable, or more fragile, than the price assumes. (c) The reverse DCF. Solve for the growth and duration the 38× multiple implies, then ask whether the moat evidence on the competitive advantage period supports more or less than that. That one number separates "fine price for quality" from "priced for perfection," and tells you whether you're variant.

P10 (exam, 12 min: EV and resulting). A biotech at $20 faces a binary trial readout, with a 25% chance of $60, a 35% chance of $22 and a 40% chance of $8. (a) Compute EV and expected return (±$0.1). (b) Is it a good bet on the numbers? (c) The trial fails and you lose 60%. A colleague says "terrible call." Rebut him point by point.

Solution. (a) EV = 0.25×60 + 0.35×22 + 0.40×8 = 15 + 7.7 + 3.2 = $25.9. Against $20, that's a +29.5% expected return. (b) Yes, on the numbers. The EV is positive and sizeable, from an asymmetric payoff in which a 25% shot at +200% outweighs the 40% chance of −60%. That holds provided the position was sized small enough that the likely-ish loss can't ruin the portfolio. (c) He's resulting, judging the decision by its outcome. If the 25/35/40 odds were estimated in good faith and the size was survivable, this was a good decision with a bad outcome. I should take the identical bet every time it's offered, because across many such bets the +29.5% edge compounds. One unlucky result doesn't make a sound, well-sized, positive-EV decision "terrible." And had it hit $60, that wouldn't have made a reckless, oversized version "brilliant." Grade the process and ignore the roll of the dice.

P11 (exam, 12 min: Kelly with a capped downside). A concentrated idea: win +120% (u = 1.2) with p = 0.55; lose −40% (d = 0.40) with q = 0.45. (a) Compute raw full Kelly, f* = p/d − q/u (±0.01). (b) Interpret the number. (c) What do you actually do, and what one fact must you verify?

Solution. (a) f* = 0.55/0.40 − 0.45/1.20 = 1.375 − 0.375 = 1.00, so raw full Kelly is 100% of capital. (b) The formula says "all-in." Taken literally that's absurd, but it shows how Kelly behaves. Raw Kelly on a favourable, capped-downside bet routinely spits out enormous fractions, here 100%, and with leverage assumptions it can exceed 100%. A survivable downside plus a real edge mathematically wants size. Treat the number as a ceiling and a ranking signal. It isn't a position size. (c) In practice, apply fractional Kelly (half gives ~50%, and a quarter is likely given the uncertainty in your estimates). Then cap hard at a sane single-name limit (say 15 to 25%) and adjust for correlation with the rest of the book. The one fact to verify is that the −40% floor is real. That means real asset and earnings backing, no leverage that forces selling and no fraud. The whole case for a large size rests on the capped downside, so if the floor is illusory, the "Kelly-justified" big position becomes a route to ruin.

P12 (exam, 12 min: base rate + Bayesian update on a turnaround). The base rate of corporate turnarounds actually succeeding is ~10%. A specific turnaround shows an early encouraging sign; you assess P(sign | success) = 0.60 and P(sign | failure) = 0.20. (a) Compute the posterior probability of success (±0.01). (b) Interpret it against "turnarounds seldom turn." (c) What does this imply for sizing and for how you'd treat management's optimism?

Solution. (a) P(success | sign) = (0.60×0.10) / [(0.60×0.10) + (0.20×0.90)] = 0.06 / (0.06 + 0.18) = 0.06/0.24 = 0.25. (b) Even a really encouraging early sign lifts the probability only from the 10% base rate to ~25%. The base rate is so low that one positive signal, which failing turnarounds also sometimes give off, can't move it far. Buffett's "turnarounds seldom turn" is a base-rate statement, and Bayes shows why one good data point doesn't overturn it. (c) Size small. A ~25% success probability still means a probable failure, so treat it as a low-conviction, capped-downside bet or pass entirely. Demand a wider margin of safety. And treat management's optimism as an inside view to check against the outside-view base rate. It isn't evidence in itself. Require several strong, independent, hard-to-fake signals, each one a further Bayesian update, before your conviction (and your size) goes up.

P13 (guided: expected value on a decision tree). Theertha Pharma (synthetic) trades at ₹600. Its value turns on two sequential gates. A plant re-inspection clears with probability 0.65, and only if it clears, a key product approval comes through with probability 0.40. Prices are ₹1,050 for clear-and-approved, ₹680 for clear-but-rejected and ₹350 for not cleared (import alert). (a) Compute the three joint probabilities. (b) Compute the EV and expected return vs ₹600. (c) What is the modal outcome, and does it decide the bet?

Solution. (a) Multiply along the branches. Clear-and-approved = 0.65 × 0.40 = 0.26, clear-but-rejected = 0.65 × 0.60 = 0.39, and not cleared = 0.35 (sum = 1.00). The approval probability is conditional. It applies only in the 65% of worlds where the plant clears first, and that structure is why you draw a tree at all. (b) EV = 0.26×1,050 + 0.39×680 + 0.35×350 = 273 + 265.2 + 122.5 = ₹660.7. Against ₹600, that's a +10.1% expected return. (c) The modal outcome is clear-but-rejected (p = 0.39) at ₹680, a +13.3% gain, but it doesn't decide the bet. You weigh the whole distribution, where the ₹1,050 upside tail and the −41.7% ₹350 downside both count. And you size so the 35% import-alert branch can't ruin you.

P14 (guided: Bayesian updating and diagnosticity). Your prior that a high-ROIC company still has a real moat a decade out is the base rate from the ROIC persistence data in the moats work, P(H) = 0.30. It posts another year of margin expansion (evidence E). Compute the posterior under two readings of that evidence: (i) weakly diagnostic, P(E|H) = 0.60, P(E|not H) = 0.50; (ii) strongly diagnostic, P(E|H) = 0.90, P(E|not H) = 0.20. (c) What does the contrast teach?

Solution. (i) P(H|E) = (0.60×0.30) / [(0.60×0.30) + (0.50×0.70)] = 0.18 / (0.18 + 0.35) = 0.18/0.53 = **0.34**, and belief crawls from 0.30 to 0.34. (ii) P(H|E) = (0.90×0.30) / [(0.90×0.30) + (0.20×0.70)] = 0.27 / (0.27 + 0.14) = 0.27/0.41 = **0.66**, and belief jumps from 0.30 to 0.66. (c) Same prior, same direction of evidence, but diagnosticity governs the size of the update. That's the likelihood ratio P(E|H) ÷ P(E|not H), 1.2 in case i and 4.5 in case ii. Margin expansion that a no-moat company would also often show, as in (i), is nearly worthless. Margin expansion that essentially only a real moat produces, as in (ii), is powerful. Before updating on any "good news," ask how likely that same news would be if your thesis were wrong.

P15 (independent: base-rate blending). Aster Cloud Inc. (synthetic), a US SaaS firm, has grown revenue ~40%/yr, and your inside-view model also says ~40% for five more years. The outside-view base rate for software firms already at its ARR scale is ~12%. Using estimate = w × inside + (1 − w) × base: (a) compute the blend at w = 0.25 and w = 0.40; (b) what does a naïve 50/50 give; (c) which weight is defensible, and why isn't the answer 40%?

Solution. (a) w = 0.25: 0.25×40 + 0.75×12 = 10 + 9 = 19%. w = 0.40: 0.40×40 + 0.60×12 = 16 + 7.2 = 23.2%. (b) 50/50: 0.5×40 + 0.5×12 = 26%, already far below the 40% story. (c) A low weight (≈0.20 to 0.30, i.e. ~15 to 19%) is defensible. The 12% base rate comes from a class that already includes many fast-growers that decelerated as their base grew, the law of large numbers of revenue. Differentiated evidence has to overcome that, and it rarely earns more than a modest weight. The answer isn't 40%, because adopting the inside view whole is the planning-fallacy trap: the story feels uniquely destined to beat the odds its predecessors faced. Then read the reverse-DCF implied growth. A price assuming ~40% means pass, and a price assuming ~10% means you may be variant.

P16 (independent: inversion / probability of ruin). Run a quantified pre-mortem on a synthetic levered cyclical. Four independent failure modes over five years: funding shock 25%, key-customer loss 20%, technology substitution 12%, regulatory hit 8%. (a) Compute the probability the business survives all four, and the probability at least one lands. (b) If diligence could cut the two largest to 8% each, recompute survival. (c) Why is the true probability of ruin likely worse than your (a) figure?

Solution. (a) P(survive) = (1 − 0.25)(1 − 0.20)(1 − 0.12)(1 − 0.08) = 0.75 × 0.80 × 0.88 × 0.92 = **0.486**, so P(at least one kill-shot) = 1 − 0.486 = **0.514**. The "individually survivable" risks compound into a >50% chance of disaster. (b) Cut the funding and customer terms to 0.08 each, and 0.92 × 0.92 × 0.88 × 0.92 = **0.685**. Ruin falls from ~51% to ~31%, and that ~20-point improvement is the payoff from the two diligence actions the pre-mortem named. (c) Assuming independence is optimistic. In the tail the modes are positively correlated, since a recession fires the funding shock and the customer loss together. So the real joint probability of ruin exceeds the product-rule figure right when it matters, and that's why you size for correlated stress instead of the independent case.

P17 (independent: volatility drag). Two synthetic strategies each have an arithmetic-mean annual return of 10%. Strategy Smooth has return volatility (standard deviation) of 10%; strategy Jumpy has 30%. Using geometric ≈ arithmetic − variance/2: (a) estimate each one's compound (geometric) return; (b) compound each over 20 years from ₹1; (c) state the lesson for drawdowns.

Solution. (a) Smooth: variance = 0.10² = 0.01, so geometric ≈ 0.10 − 0.01/2 = 9.5%. Jumpy: variance = 0.30² = 0.09, so geometric ≈ 0.10 − 0.09/2 = 5.5%. Same average, and a four-point gap in what actually compounds. (b) Smooth: 1.095²⁰ ≈ 6.14×. Jumpy: 1.055²⁰ ≈ 2.92×. The smoother strategy ends ~2.1× richer on an identical arithmetic mean. (c) Volatility is a direct tax on compound growth (the variance/2 term). So cutting the depth of drawdowns mechanically raises long-run wealth, on top of any emotional comfort it brings. Two managers with the same average return aren't equally rich, and the calmer one wins.

P18 (stretch, 12 min: full decision workup, US). Nyx Bio (synthetic) trades at $30 ahead of a binary trial readout, with a 25% chance of $90, a 30% chance of $35 and a 45% chance of $12. (a) EV and expected return (±$0.1). (b) For sizing, collapse to the binary that matters, the bad ($12) outcome versus everything else, by computing the blended up-fraction u as the probability-weighted price given a non-rejection, then raw Kelly f* = p/d − q/u with d = 0.60. (c) What do you actually stake, and does any single outcome threaten ruin at that size? (d) The trial fails and you lose 60%. Rebut "terrible call."

Solution. (a) EV = 0.25×90 + 0.30×35 + 0.45×12 = 22.5 + 10.5 + 5.4 = $38.4. Against $30, that's +28.0%. (b) The non-rejection probability is p = 0.25 + 0.30 = 0.55. The blended up-price given non-rejection is (0.25×90 + 0.30×35) ÷ 0.55 = 33 ÷ 0.55 = $60, so u = 60/30 − 1 = **1.0** (+100%). With q = 0.45 and d = 0.60, f* = 0.55/0.60 − 0.45/1.0 = 0.9167 − 0.45 = **0.467**, which makes raw full Kelly ≈ 47%. (c) Never 47%. Apply fractional Kelly (half gives ~23%) and cap hard at a sane single-name limit, say 15%, adjusted for correlation. At 15%, a −60% move is a 0.15 × 0.60 = **9%** portfolio hit. That's painful but survivable, and it's nowhere near ruin. (d) Not a terrible call. That's resulting. If the 25/30/45 odds were honest, the EV was positive (+28%) and the size was survivable, this was a good decision with a bad outcome. You'd take the identical bet again, because the edge compounds across many such independent bets and no single one can ruin you. Had it hit $90, that wouldn't have made an oversized version wise. Grade the process, and journal it before the readout.

Applied mini-project: your circle of competence + a full decision workup

This project produces two lasting artifacts that feed straight into your personal checklist and the capstone Mental-Model & Moat Checklist. Part A is the headline deliverable, and it meets the "define your circle of competence in writing" objective. Part B runs the whole decision-science kit on one real case.

Part A: Your circle of competence, in writing (1 to 2 pages). Sort your investable world into three honest lists. Inside holds the sectors and business models you can really evaluate. For each, say why: you can spread the financials by hand, you know the KPIs and thresholds from Phase 5, and you've studied the failure cases. Then name the one thing that would most likely make you wrong even here. Edge holds what you're actively learning and can hold only small, with an extra margin of safety. Outside is a long, unashamed list of what you'll pass on. For two of those, write one sentence on why you can't evaluate them yet, which is the most valuable part. Finish with one paragraph on where your own mistakes are most likely to come from, meaning the boundary of your circle. It becomes a standing checklist line.

Part B: A decision workup on one real company (3–4 pages). Pick one name you can get good data on, Indian or US, from the starter bank (India: Asian Paints, DMart, Titan, HDFC Bank, Bajaj Finance, Divi's; US: Costco, Visa, Moody's, Copart, Adobe) or another inside your circle. Produce, in order:

1. A pre-mortem. "It's five years later and this lost 60%+. What happened?" Write four specific failure stories, and turn each into a thing to verify or size for.
2. An outside-view growth estimate. Give the inside-view story first. Then give the reference class and its base rate, using Mauboussin's Base Rate Book plus a screener history. Add your base-rate-anchored estimate with a bear/base/bull band, and the reverse-DCF implied growth from the current price. State which side of the consensus you're on.
3. An EV framing. Build a three-outcome table (bear/base/bull) with realistic probabilities and target values. Compute EV against price, and note the modal outcome and whether the payoff is asymmetric.
4. A size. Sketch a fractional-Kelly size. State your u, d, p. Compute raw f*, then apply half-Kelly plus a single-name cap. Name the one fact your downside floor depends on.
5. A Bayesian plan. Give your prior that the thesis is right and the two or three pieces of future evidence that would move it most, with rough likelihoods. Then commit in advance to what would change your mind.
6. A ruin check. Could any single outcome permanently impair the portfolio at your chosen size? If so, cut the size until the answer is no.

Data to pull. For India: the latest annual report and investor presentation (company site), the screener.in ten-year page (revenue, margins, ROCE, growth) and the shareholding and related-party disclosures. For the US: the latest 10-K and DEF 14A proxy (EDGAR), plus stockanalysis.com or TIKR for the ten-year series. For both: Mauboussin's Base Rate Book (free Counterpoint Global PDF, current edition) and your own reverse-DCF sheet.

Scoring rubric (score yourself honestly, pass at ≥12/16):

Reading & resources

• Charlie Munger, Poor Charlie's Almanack. "The Psychology of Human Misjudgment" (the misjudgment speech; its behavioural content is the spine of M6.04) and the worldly-wisdom talks ("A Lesson on Elementary, Worldly Wisdom") for the latticework and the man-with-a-hammer idea. The founding text here. [Paid] [Intermediate]
• Michael Mauboussin, the free Counterpoint Global / Credit Suisse reports. The Base Rate Book (sales growth, margin, ROIC persistence distributions; download the current edition and use the actual tables), Think Twice (inside vs outside view, base rates, "resulting" in decisions), and More Than You Know (probabilistic thinking, decision-making). The best free resource here. [Free] [Intermediate]
• Annie Duke, Thinking in Bets (2018). Decision quality vs outcome quality, "resulting," thinking in probabilities, and running your beliefs as bets. The clearest single book on the process-over-outcome spine. [Paid] [Beginner]
• Daniel Kahneman, Thinking, Fast and Slow. Part III (overconfidence, the inside vs outside view, the planning fallacy) and Part IV (prospect theory; read now for base rates, revisit in M6.04). [Paid] [Intermediate]
• Howard Marks, The Most Important Thing (ch. on second-level thinking) and the Oaktree memos (oaktreecapital.com/insights), which are free, especially "You Bet!" (probabilistic thinking, decision quality vs luck) and "Risk." [Book Paid / Memos Free] [Beginner→Intermediate]
• Frank Knight, Risk, Uncertainty and Profit (1921). The original risk-vs-uncertainty distinction; read summaries or the early chapters instead of the whole. Pair with Nassim Taleb, Fooled by Randomness and The Black Swan for the modern, market-facing version (uncertainty, fat tails, avoiding ruin) and Skin in the Game for the ergodicity argument. [Knight Free (public domain) / Taleb Paid] [Intermediate→Advanced]
• William Poundstone, Fortune's Formula (2005). The history and intuition of the Kelly criterion (Kelly, Shannon, Thorp) told as narrative; the most enjoyable way to internalise why you size by edge and why full Kelly is too much. Pair with Ed Thorp, A Man for All Markets (also in M9.01). [Paid] [Beginner→Intermediate]
• Ole Peters, "The ergodicity problem in economics" (Nature Physics, 2019) and the Ergodicity Economics lecture notes (ergodicityeconomics.com, free): the rigorous source for time-average vs ensemble-average and the multiplicative-returns trap; technical but the core coin-flip example is accessible. [Free] [Advanced]
• Philip Tetlock, Superforecasting (2015). Base rates, the outside view, calibration, and updating in practice; the empirical case that good forecasting is a trainable process of the kind taught here. [Paid] [Beginner→Intermediate]
• Gerd Gigerenzer, Calculated Risks (a.k.a. Reckoning with Risk). Natural-frequency Bayesian reasoning made intuitive (the medical-test / base-rate examples that mirror the fraud-screen case here). [Paid] [Beginner]
• India ties. Sanjay Bakshi's "Behavioural Finance & Business Valuation" (BFBV) lecture notes and Fundoo Professor blog (free): the finest Indian treatment of Munger's latticework, mental models, and base rates applied to Indian companies; Gautam Baid, The Joys of Compounding (mental models, circle of competence, process, with an Indian investor's lens). [Bakshi Free / Baid Paid] [Intermediate]
• Going deeper: Edward O. Thorp, "The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market" (1997, rev. 2007; free PDF at edwardothorp.com). The rigorous derivation of Kelly and the quantitative case for fractional Kelly and drawdown control, a level below Poundstone's narrative into the math of Kelly sizing. [Free] [Advanced]
• Going deeper: Daniel Kahneman, Olivier Sibony & Cass Sunstein, Noise: A Flaw in Human Judgment (2021). Why independent expert judgments of the same case scatter (noise, an error distinct from bias) and how decision hygiene, structured protocols, and aggregation reduce it; a direct deepening of the process-over-outcome and calibration spine (competency C9). [Paid] [Intermediate]

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": "The multidisciplinary latticework and circle of competence, inversion and the pre-mortem, opportunity cost as the true hurdle, second-order thinking, risk vs uncertainty (Knight), base rates and the outside view, expected value under 'resulting', the Kelly criterion and fractional Kelly, Bayesian updating, and ergodicity/ruin avoidance",
  "use": "Two narrow jobs, and one absolute ban. (1) BAYES SCAFFOLDING: give it a thesis and a new piece of evidence and ask it to lay out the three empty cells a Bayesian update needs — P(H) the prior, P(E|H), P(E|not H) — as a table for YOU to fill in from the disclosure, then have it show the update arithmetic once you supply the numbers. (2) DIVERGENT LISTS: ask it for a first-pass, deliberately unfinished list of candidate pre-mortem kill-shots, or a first-pass list of reference-class companies for a base rate — raw material for you to prune, verify against primary sources, and take ownership of, never a finished pre-mortem or a finished base rate. THE BAN: never accept a probability from it as usable input — a prior, a P(E|H), a base-rate percentage, a Kelly p — because a fluent-sounding probability is indistinguishable from a correct one until you trace it to a source, and every probability this module's toolkit runs on must come from your own reference class, your own likelihood reasoning, or a disclosure you can cite.",
  "tools": ["Chat assistants — Claude, ChatGPT for Bayes-table scaffolding and first-pass brainstorming lists", "A spreadsheet or notebook — where the Kelly fraction, Bayes update, base-rate blend, and pre-mortem ruin arithmetic actually get computed", "screener.in's peer-comparison and ten-year data pages, and Mauboussin's Base Rate Book (Counterpoint Global) — for assembling a real reference class", "SEC EDGAR's companyfacts XBRL API — for a US reference class's exact revenue history"],
  "prompt": "Here is my thesis: [state it in one sentence]. Here is a new piece of evidence: [state it]. Lay out the three numbers a Bayesian update needs as blank cells for me to fill in from the disclosure myself — P(H) the prior, P(E|H) how likely this evidence is if the thesis is true, P(E|not H) how likely it is if the thesis is false — then show me the update formula with those cells still blank. Do not suggest values for any of the three.",
  "verify": "Fill in every probability yourself from a source you can name, then compute the posterior, the Kelly fraction, or the ruin probability by hand once before trusting a spreadsheet formula or an AI's arithmetic; open the primary filing for any AI-suggested reference-class company rather than accepting its characterization of that company.",
  "diy": "Without a tool in the room: state the Kelly fraction (gambling and capped-downside forms) from given odds; run a Bayesian update by hand from a stated prior and two likelihoods; blend an inside view and an outside view into one number and defend the weight you chose; compute the probability of ruin from a list of independent kill-shot probabilities; and read a bet's payoff structure well enough to say, unaided, whether to size it up or walk away."
}

Modern Data Analysis

By hand first. WE5 builds the outside view from a short, given list of comparable companies' growth history. In practice nobody hands you that list. You assemble a reference class from real disclosures, one filing at a time, before you can blend anything.

Today's workflow. Pick a reference class, for example "listed Indian consumer companies that crossed ₹1,000 crore revenue while growing above 25%." Pull each member's revenue for the crossing year and the five years after it from real filings. Bucket the class by starting growth rate, and read off the distribution of subsequent growth in each bucket. That's the same table Mauboussin's Base Rate Book publishes for the US large-cap universe. Apply WE5's blend formula against that table, and sensitivity-test the weight w instead of asserting one value once.

Tools & sources (IN + US). For India, use screener.in's "Compare companies" tool and its ten-year Profit & Loss data pages to get a hand-picked peer set's revenue by year, and the BSE/NSE annual-report archives for years outside screener.in's default window. For the US, SEC EDGAR's companyfacts XBRL API (data.sec.gov/api/xbrl/companyfacts/CIK##########.json, the Revenues tag) gives a reference class's exact revenue history. Michael Mauboussin's Base Rate Book (Morgan Stanley Counterpoint Global, a free downloadable PDF) is a professionally built reference-class base rate to sanity-check your hand-built one against.

# recomputed in-session — R3
import pandas as pd

def subsequent_growth_by_bucket(df: pd.DataFrame, bucket_edges, growth_col="growth_next"):
    """Bucket a reference class by its CURRENT growth rate and report the
    distribution of growth realised over the following period — the shape
    of table the Base Rate Book publishes."""
    df = df.copy()
    df["bucket"] = pd.cut(df["growth_now"], bucket_edges)
    return df.groupby("bucket", observed=True)[growth_col].agg(["count", "mean", "median", "std"])

def blend_estimate(inside_view_pct: float, outside_view_base_rate_pct: float, w: float) -> float:
    """WE5's formula, codified: estimate = w*inside + (1-w)*outside."""
    return w * inside_view_pct + (1 - w) * outside_view_base_rate_pct

# WE5's own figures, run through the function as the verification anchor
inside, outside = 18.0, 9.0
for w in (0.35, 0.20, 0.50):
    print(f"w={w}: blended estimate = {blend_estimate(inside, outside, w):.2f}%")

Verify. blend_estimate(18.0, 9.0, w) reproduces WE5's Step 2: w=0.35 gives 12.15%, w=0.20 gives 10.80% and w=0.50 gives 13.50%. And 1.18**10 / 1.12**10 = 1.685 matches the "68% higher" in Step 3 to the rounding stated there. The function adds nothing WE5 didn't already do by hand. Its value is that a base-rate blend built from forty real reference-class rows and one built from three sentences of story both reduce to the same one-line, auditable formula. So the arithmetic is settled, and all that's left to defend is w and the reference class it came from.

Quantitative lens

The coin game from the ergodicity discussion (heads ×1.5, tails ×0.6) proves ergodicity with one clean closed form: ensemble average +5%, time average −5.1%. A closed form shows you where the paths end up. It can't show you how they get there: how a mean gets dragged upward, how unstable that mean is to estimate, or what happens near the Kelly boundary. Monte Carlo shows all three.

The ensemble/time-average gap, path by path. Simulate 200,000 independent 50-round paths of the coin game. The sample mean final wealth is 13.21× starting wealth (closed form 1.05^50 = 11.4674×), while the sample median is 0.071790× (closed form 0.9487^50 = 0.071789×, matching to five significant figures). Of all the paths, 75.89% end below where they started and 23.78% end below 1% of starting wealth. The single luckiest path of the 200,000 ends at 417,872×. The 99.9th-percentile path alone reaches 1,712×, and dropping just the top 0.1% of paths collapses the sample mean from 13.21× to 5.47×. A handful of astronomical outliers carries the entire "ensemble average looks fine" story, just as the ergodicity argument said in words.

The mean is itself unstable, a second and sharper way to see non-ergodicity. Re-run the same 200,000-path simulation across five different random seeds and you get sample means of 9.81×, 10.41×, 10.84×, 11.04×, 11.08×. That's a spread of more than 1.3× on a supposedly fixed target, while every seed's median comes out at 0.07179×, agreeing to four decimal places every time. The analytic standard deviation of final wealth is ≈776, roughly 68 times the analytic mean of 11.47. So the mean is a real number, but from any finite sample it's practically un-estimable. That's a second, independent demonstration of the same non-ergodicity you saw derived from the coin's two multipliers. The median (the time-average proxy) is stable and estimable. It's also the one number that describes what actually happens to a real sequential player.

Fractional Kelly, simulated at WE3's own numbers. WE3 Case A (p = 0.60, win +100%, lose 100%, full Kelly f* = 0.20) run for 100 rounds and 100,000 paths at four sizes:

Every row shares the identical positive-EV bet (0.60 × 1.0 − 0.40 × 1.0 = +0.20 expected value per unit staked). Only the stake changes. Double the full-Kelly fraction and the compound growth rate goes negative. After 100 rounds the typical, median path is left with 78% of starting capital, a loss, even though every individual bet has positive expected value. The chance of falling below 10% of starting capital jumps from 1.72% at full Kelly to 30.63%. That's the Kelly warning, "overbetting is punished far more than underbetting," turned into a number. Past its boundary, the Kelly fraction doesn't degrade gracefully like a soft suggestion. It's close to a cliff edge.

Know the limits here. The coin game's ×1.5/×0.6 payoffs and WE3's p = 0.60 are stylised teaching parameters, chosen because they give clean closed forms. Nobody fitted them to a real security's return distribution. So the lesson is the shape of these results, the mean/median divergence and overbetting's cliff. Hold the specific percentages loosely. The Kelly simulation also assumes p is known for certain. In practice p is itself an estimate with error. That's why the Kelly section recommended fractional Kelly before you saw any simulation, and the Monte Carlo only shows in numbers what the argument there already made in words. And every path here is independent of every other. A real portfolio's bets often share risk factors, which is the hidden-correlation warning from the concentration debate, applied to sizing instead of to basket composition.

Where this goes next: galaxy cross-links

• G1 · QM1.01 (Probability, Distributions, Sampling & Estimation): the formal treatment of sampling distributions, of which the seed-to-seed instability above is a special, heavy-tailed case.
• G1 · QM1.03 (Time Series, Monte Carlo & ML): the general Monte Carlo simulation toolkit (path generation, percentile/tail estimation) that the simulations above apply to one specific non-ergodic process.
• G2 · DA1.05 (Statistics in Code): coding the base-rate bucket table and the Bayes/Kelly arithmetic as reusable, testable functions instead of one-off spreadsheet cells.

Teach it back & journal

Feynman prompt. Write a one-page explainer for a smart 15-year-old titled "Why a good decision can lose money (and a bad one can win)." Get three ideas across without jargon. First, when luck is involved you can't judge a choice by how it turned out. You judge it by whether the reasoning and the bet size were sensible, given what could be known. Second, smart people weigh all the things that might happen, each by how likely it is (expected value), instead of fixating on one. Third, you must never bet so much on any one thing that a bad roll could wipe you out, because you can only keep playing, and keep compounding, if you survive. If your page needs the words "expected value," "Kelly," or "ergodicity," you don't own it yet. Bonus: explain, in one sentence a child would get, why losing half your money means you then have to double what's left just to get back to even.

Journal prompt. Three entries. (1) Write the first draft of your circle of competence: three sectors clearly inside, two clearly outside and one that's really at the edge. For the edge one, add the single thing you don't yet understand well enough. You'll formalise this in the checklist capstone. (2) Take the last investment decision you actually made (or seriously considered) and separate its decision quality from its outcome quality. Was the reasoning sound and the size survivable regardless of how it turned out? Were you ever "resulting," praising or blaming yourself for a result instead of a process? (3) Pick one belief you currently hold about a company or market and write it as a Bayesian statement. Give your honest prior probability that it's true, and the two pieces of future evidence that would move it most, one that would raise it and one that would lower it. Commit to them now, before you know the answer. Come back to this entry when the evidence arrives and update explicitly. Repeated for years, that habit is what calibration is made of.

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