06.8 · Margin of safety, expected value & asymmetry¶

Why this matters: you have now built a value of ₹320 for a stock priced at ₹390, with a bear case of ₹168 and a bull case of ₹416. The decision is not "is ₹320 right?" — it never will be exactly — but "given how wrong I could be, does this price pay me to take the risk?" That is what margin of safety means when you stop treating it as a slogan and start treating it as a distribution. This lesson turns the valuation outputs of Module 06 into a decision input for Module 11.

Learning objectives — after this lesson you can:

Restate Graham's margin of safety as a statement about the distribution of value, not a fixed discount.
Compute expected value and payoff asymmetry across scenarios, and explain why downside analysis comes first.
Run a simple Monte-Carlo simulation of a DCF's inputs and read the resulting value distribution.
Connect uncertainty in value to position sizing (a preview of Kelly and the max-loss rule in 11.4).
State, for Kaveri at ₹390, exactly what would have to change for it to become attractive — and at what price it already is.

Prerequisites: 06.4 DCF in practice, 06.6 Reverse DCF & expectations · Time: ~80 min

1. Margin of safety, restated¶

Benjamin Graham's idea: buy only when price is well below your estimate of value, so that errors in the estimate, bad luck and the unknowable are absorbed before you lose money. The traditional rule of thumb — "a third off" — is a crude way of saying something precise:

Your estimate of value is a random variable with a mean and a spread. The margin of safety is the distance between price and that distribution, measured in units of its spread.

Three consequences that the slogan hides:

The required margin depends on the spread. A 20% discount to a value you can estimate within ±10% (a regulated utility) is a large margin; a 40% discount to a value you can only place within ±60% (a biotech, a turnaround) is a small one.
The margin is about the downside, not the mean. A stock priced at your mean value with a fat left tail is dangerous; one priced above your mean with almost no downside (net cash, hard assets, contracted revenue) may be fine.
Uncertainty is not symmetric. Business risks are skewed: things break faster than they are built. The left tail of a value distribution is usually longer than the right — which is why downside analysis comes first.

Trader's lens

Margin of safety is the edge divided by the standard error of the edge — a Sharpe ratio for one position. A 2% mispricing with a 1% estimation error is a better trade than a 30% mispricing with a 40% error. And because business outcomes are negatively skewed, the "expected value" of an equity position is like the P&L of a short-put book: mostly small positive, occasionally large negative. The margin of safety is the premium you demand for that skew.

2. Downside first¶

Before any upside analysis, answer: what can I lose, and how likely is it? Not the worst imaginable case — the plausible bad case — and the mechanism by which it happens.

Downside source	Kaveri	Sizing
Operating bear case	State dues stuck; growth 3–8%; margin 12–13%; NWC stays high	₹168/share (−57% from ₹390) — the reference bear case
Balance-sheet stress	Working-capital lines cut; forced provisioning; rights issue	The stress test in 04.5 showed no insolvency, but a ₹60 Cr write-off would cut PAT by ~50% in the year it lands
Governance jump	Pledge invocation; related-party revelations; auditor issues	Historical base rate for Indian small caps with a governance score of ~7: a meaningful minority suffer a 30%+ gap-down within three years (order-of-magnitude judgement, not a statistic)
De-rating	Multiple falls from 26x to the 15–18x an ex-growth industrial commands, even with flat earnings	₹15.1 × 16 = ₹240 (−38%) with no change in profit
Liquidity	Small cap; exit takes days in a drawdown	Impact cost of 2–5% on the way out

The de-rating line is the one most investors forget. Kaveri does not need to have a crisis to lose you a third of your money; it only needs to stop being seen as a growth company.

3. Expected value and asymmetry¶

With scenario values and probabilities, the expected value and the payoff asymmetry are arithmetic:

Scenario	Probability	Value	Return from ₹390	Prob. × return
Bear	25%	₹168	−56.9%	−14.2%
Base	50%	₹320	−17.9%	−9.0%
Bull	25%	₹416	+6.7%	+1.7%
Expected		₹306	−21.5%

\[\text{Upside/downside ratio} = \frac{P(\text{up}) \times \text{gain}}{P(\text{down}) \times \text{loss}} = \frac{0.25 \times 6.7\%}{0.75 \times \bar{L}}\]

where the probability-weighted loss across bear and base is (0.25 × 56.9 + 0.50 × 17.9)/0.75 = 30.9%, giving a ratio of 1.7 / 23.2 ≈ 0.07. For a long position you want this ratio comfortably above 1 — typically 2–3 for a concentrated position. At ₹390 Kaveri is not a marginal call; it is the wrong side of a lopsided bet.

Now find the price at which it becomes a reasonable bet. Holding the scenario values fixed:

Entry price	Expected return	Upside/downside ratio	Verdict
₹390	−21.5%	0.07	Avoid
₹306 (= EV)	0%	1.0	Fair — no margin
₹260	+17.7%	3.0	Attractive: base and bull pay well; bear loses 35%
₹220	+39.1%	7.6	Very attractive: bear loses only 24%
₹168 (= bear)	+82%	large	Priced for the bear case

scen = {"bear": (0.25, 168), "base": (0.50, 320), "bull": (0.25, 416)}
def ev_ratio(price):
    rets = {k: (p, v / price - 1) for k, (p, v) in scen.items()}
    er = sum(p * r for p, r in rets.values())
    up = sum(p * r for p, r in rets.values() if r > 0); dn = -sum(p * r for p, r in rets.values() if r < 0)
    return round(er, 3), round(up / dn, 2) if dn else float("inf")
for px in (390, 306, 260, 220, 168): print(px, ev_ratio(px))

This is the practical meaning of margin of safety for Kaveri: around ₹220–260, the price at which even the base case pays a double-digit return and the bear case is survivable. It is also a watchlist trigger — a number to write down now, before the price gets there and the narrative changes.

4. From scenarios to a distribution: Monte Carlo¶

Three scenarios are a coarse distribution. A Monte-Carlo simulation draws every uncertain input from a distribution, runs the DCF thousands of times and produces the full histogram of value. It is not more accurate — the input distributions are your guesses — but it exposes interactions (a bad growth draw and a bad margin draw) and gives tail probabilities the three-point tree cannot.

import numpy as np
from fi.valuation import dcf_from_drivers, kaveri_base_drivers
d = kaveri_base_drivers(); rng = np.random.default_rng(42); vals = []
for _ in range(10_000):
    g_scale = rng.normal(1.0, 0.20)                 # growth path scaled ±20%
    m_shift = rng.normal(0.0, 0.012)                # margin path shifted ±1.2 pp
    nwc = rng.normal(0.22, 0.03)                    # terminal NWC/revenue
    w = max(0.10, rng.normal(0.1219, 0.006))        # WACC ±60 bps
    ronic = max(0.13, rng.normal(0.18, 0.03))       # terminal return on new capital
    growth = [x * g_scale for x in d["growth"]]
    margin = [max(0.08, x + m_shift) for x in d["ebitda_margin"]]
    nwc_path = [max(0.15, nwc + 0.03), max(0.15, nwc + 0.01)] + [max(0.15, nwc)] * 8
    vals.append(dcf_from_drivers(**{**d, "growth": growth, "ebitda_margin": margin,
                                    "nwc_pct": nwc_path, "wacc": w, "ronic": ronic})["per_share"])
v = np.array(vals)
print(f"mean ₹{v.mean():.0f}  median ₹{np.median(v):.0f}  P10 ₹{np.percentile(v,10):.0f}  P90 ₹{np.percentile(v,90):.0f}")
print(f"P(value < ₹390) = {(v < 390).mean():.0%}   P(value < ₹250) = {(v < 250).mean():.0%}")
# mean ₹324  median ₹316  P10 ₹236  P90 ₹420
# P(value < ₹390) = 83%   P(value < ₹250) = 15%

Read the output as a trader reads a P&L distribution: the median (₹316) sits near the base case; the 10th–90th percentile band is ₹236–420 — a ±30% range that is the honest precision of this valuation; and 83% of the simulated values are below the ₹390 price. The simulation does not say Kaveri is worth ₹316; it says that under these input distributions, paying ₹390 is a bet on the top fifth of outcomes.

Caveats that matter: the inputs were drawn independently (in reality bad growth and bad margins go together — correlating them fattens the left tail); the distributions are normal (business inputs are skewed); and the whole exercise inherits the base drivers. Use it to see the shape, never to claim a probability to two decimals.

5. Uncertainty and position size — a preview¶

The width of the distribution should set the position size, not just the decision to buy. Two rules from 11.4, previewed here:

Kelly-style sizing says the optimal fraction of capital to bet is proportional to edge ÷ variance: a stock with 30% expected return and a ±30% value range deserves a much larger position than one with 30% expected return and a ±80% range. Most investors use half or quarter Kelly because the inputs are estimates.
Max-loss sizing says: size the position so that the bear case loses no more than a fixed fraction of the portfolio (say 2–3%). For Kaveri at ₹260 (bear loss −35%), a 3% max-loss rule caps the position at ~8.5% of the portfolio; at ₹390 (bear loss −57%) the cap is ~5% — and the expected return is negative anyway.

The point of doing valuation as a distribution is that it feeds both rules directly. A single target price feeds neither.

6. Kaveri: what would have to change¶

Bring the module together in one table — the conditions under which the conclusion flips, each tied to an observable:

For Kaveri to be attractive at ₹390, you need…	Observable test	Where you would see it
Bull-case operating outcome as the base case: 15%+ growth, margin → 16.5%, NWC → 19%	Two quarters of 15%+ growth with receivable days < 80 and EBITDA margin > 15%	Q2–Q3 FY27 results; investor presentation (if the KPI returns)
Or a lower market cost of equity (~11.5%) applied consistently	Nifty and peer multiples holding or rising while G-secs fall	Macro — 12.1
Or a price near ₹220–260 with the base case intact	The stock falling ~35–45% without the bear case materialising (e.g., a market-wide small-cap sell-off)	Price; check that receivables and provisioning have not worsened
And no governance deterioration	Pledge not rising; RPT purchases stable as % of costs; no auditor/KMP changes	SAST filings; annual report; announcements

That is the output of an entire valuation module for one company: not a target price, but a map of prices and events to decisions. It goes into the memo in 11.3 and the monitoring plan in 11.5.

India notes

Indian small caps have historically shown larger drawdowns than the index in sell-offs (2008, 2018, the 2024–25 correction — verify magnitudes), and liquidity vanishes at exactly the moment a margin of safety is tested. Add an explicit liquidity haircut to the bear case for any stock where your position would exceed a few days of average traded value, and prefer to set watchlist prices in calm markets rather than decide in falling ones.

Common mistakes

Treating "a third off my DCF" as a margin of safety regardless of how uncertain the DCF is.
Computing expected value with probabilities chosen to make the answer positive.
Ignoring the de-rating path — a stock can lose 40% with earnings flat.
Running a Monte Carlo with independent normal inputs and quoting the tail probabilities as if they were real.
Buying at fair value with no margin because "the base case is conservative" — the base case is the mean, and you need the price below the distribution.
Letting the decision drift with the narrative instead of the pre-written triggers.

Key terms¶

Term	Meaning
Margin of safety	The gap between price and estimated value, judged relative to the uncertainty of the estimate
Expected value	Probability-weighted average of scenario values
Upside/downside ratio	Probability-weighted gain ÷ probability-weighted loss from the current price
Payoff asymmetry	The shape of the return distribution — skew — from an entry price
Downside analysis	Identifying and sizing the plausible bad cases before the upside
De-rating	A fall in the valuation multiple independent of earnings
Monte-Carlo simulation	Repeated random sampling of inputs to produce a distribution of outputs
Percentile band (P10–P90)	The range containing the central 80% of simulated values
Kelly criterion	Bet size proportional to edge ÷ variance; usually applied fractionally
Max-loss rule	Position size chosen so the bear case costs no more than a fixed share of the portfolio
Watchlist trigger	A pre-set price or event at which a decision is re-examined

Check your understanding¶

Compute the expected return and upside/downside ratio for Kaveri at ₹280 with the reference scenarios.

Answer

Returns: bear −40.0%, base +14.3%, bull +48.6%. Expected = 0.25 × (−40.0) + 0.5 × 14.3 + 0.25 × 48.6 = −10.0 + 7.1 + 12.1 = +9.3%. Up = 7.1 + 12.1 = 19.3; down = 10.0; ratio 1.9. Borderline interesting — a modest margin.

Two stocks each have expected return +20%. A: values ₹80/₹120/₹150 (25/50/25) at price ₹100. B: values ₹40/₹130/₹200 at ₹100. Which has the larger margin of safety, and why is the expected return not enough?

Answer

A: EV = 20 + 60 + 37.5 = 117.5 (+17.5%); B: EV = 10 + 65 + 50 = 125 (+25%). B has the higher expected return but a 60% loss in its bear case vs 20% for A. Margin of safety is about the distribution: A's worst case is survivable and its estimate is tighter; B needs a smaller position for the same risk budget. Expected return ignores the variance and skew that sizing depends on.

Why does a Monte-Carlo with independent inputs understate the left tail?

Answer

Because bad outcomes cluster: a demand shock lowers growth *and* margins *and* raises working capital at the same time. Independent draws average these out; correlated draws (or scenario trees, which are implicitly correlated) produce fatter tails.

Kaveri's P10–P90 band is ₹236–420. What is the "precision" of the valuation, and how should that affect the required discount?

Answer

Roughly ±30% around the median — the honest precision of a small-cap industrial DCF. With that spread, a 20% discount to the median is within the noise; a meaningful margin needs the price near or below the P10–P25 range (₹236–272), which is consistent with the ₹220–260 trigger from the scenario analysis.

State the difference between a margin of safety and a low P/E.

Answer

A low P/E is a price relative to one year's earnings; a margin of safety is a price relative to the distribution of intrinsic value. A 6x P/E on peak cyclical earnings or on a governance-flagged company can carry no margin at all; a 30x P/E can carry one if the value distribution sits well above the price.

Go deeper¶

Benjamin Graham, The Intelligent Investor, chapter 20, "Margin of Safety as the Central Concept of Investment".
Seth Klarman, Margin of Safety (1991) — out of print; the chapters on risk and on the "value of uncertainty" are the modern statement.
Michael Mauboussin, The Success Equation and his notes on base rates — for the probabilities.
Howard Marks, memos "Risk Revisited" (2014) and "The Most Important Thing" — on downside first.
11.4 Position sizing & portfolio construction — where this lesson's distribution becomes a position.

← Previous: 06.7 Other valuation methods · Module index · Next: 07.1 Banks & NBFCs →