Insurance: Life & General

Phase 5 · Week 44 · ~20 hours

Where it sits. Third of Phase 5's sector playbooks, and the last leg of the financials trio. The first playbook taught you banks (a spread earned on a leveraged, confidence-sensitive balance sheet); the second taught you NBFCs (the same spread minus the deposit franchise, plus a liability time bomb). Insurance completes the set with the strangest economics in the book: a business that gets paid before it delivers, whose cost of goods sold is an estimate that takes years to become a fact, and whose reported P&L is therefore closer to a forecast than a record. For life insurers the statutory accounts mislead so badly that an entire parallel actuarial reporting system, Embedded Value and Value of New Business, exists just to tell you the truth. For general insurers the truth hides in one ratio (the combined ratio) and one table (reserve development). Both playbooks follow: product economics, the full KPI panels with healthy/stress thresholds, valuation (P/EV for life, P/B-vs-ROE and combined-ratio quality for general), the red-flag panel, and the two great framings, Buffett's float and the persistency ratio as a mis-selling detector. On the master map's five-questions spine, insurance stresses questions 2 and 5 hardest: "is it a good business?" cannot be answered from the P&L alone, and "what is the price implying?" requires a multiple (P/EV) you have not used before.

Suggested hour budget. Core content: 8h · worked examples: 1.5h · practice set: 3h · applied mini-project: 4.5h · flashcards + mastery check: 2h · teach-back + journal: 1h.

App link. This module powers the Insurance wing of the sector case engine: the life KPI-panel drill (APE, VNB margin, RoEV, persistency, solvency computed from summary disclosures and classified healthy/watch/stress), the EV roll-forward builder (assemble opening EV → unwind → VNB → variances → closing EV, and separate operating from economic), the combined-ratio decomposer (loss + expense ratios under both Indian and US conventions), and the red-flag spotter (quarterly vignettes each hiding one deteriorating insurer). The gate is the mastery-check exam at ≥85%: the app's Insurance node unlocks the next sector playbook only on a passing score.

Learning objectives

By the end you can:

1. Explain float from first principles: why insurance inverts the normal working-capital cycle (cash in today, costs out years later), why that makes reported profit an estimate rather than a fact, and why Buffett calls durable, cheaply-generated float the engine of Berkshire Hathaway.
2. Classify any life-insurance product: term/protection, ULIP, participating (par), non-participating (non-par) savings, annuity, group. Sort them by who bears the investment, mortality, and longevity risk, and rank their VNB margins with the reasons why they differ so radically.
3. Compute and interpret the life KPI panel from public disclosures: APE, VNB and VNB margin (25–30%+ = strong), Embedded Value and RoEV, persistency by cohort (13th-month >85% = healthy), IRDAI solvency ratio (regulatory floor 150%), and product mix, each against its threshold.
4. Build a simplified Embedded Value by hand, adjusted net worth plus the discounted value of in-force profits, and run the RoEV roll-forward (unwind + VNB + operating variances vs economic variances), explaining what EV is, what it is not, and where its assumptions can be gamed.
5. Value a life insurer with P/EV and the implied new-business multiple, run the HDFC Life-style arithmetic end-to-end, and translate the lens for US/global insurers (P/B ex-AOCI vs ROE; IFRS 17's CSM as the EV cousin).
6. Decompose a general insurer's combined ratio into loss and expense ratios under both the Indian (NEP/NWP split) and US conventions, compute the cost of float, and judge when a 105% combined ratio is acceptable and when it is a slow-motion failure.
7. Read reserve development as the honesty meter: accident-year vs calendar-year ratios, prior-year strengthening and releases. Use the loss-development table to restate what an insurer really earned.
8. Run the insurance red-flag panel in fifteen minutes: VNB margin falling on ULIP tilt, persistency slide, combined ratio >100% masked by investment gains, solvency drifting toward 150%, reserve-release-propped earnings. Map India's motor/health dynamics against US P&C.

Prerequisites & connections

Builds on. M5.01–M5.02, for the financials mindset: the balance sheet is the business, valuation is equity-direct, and the regulator is a permanent character. Phase 3's financials-valuation module gave you excess-return logic and P/B against ROE, and P/EV against RoEV is the same idea one layer up. Phase 2's quality-of-earnings toolkit matters too. An insurance loss reserve is the largest single management estimate you'll ever meet, and everything you learned about accrual discretion applies at triple strength. From Phase 1 you bring statement mechanics. The formats will look unfamiliar (policyholder vs shareholder accounts in India), but the articulation logic is unchanged. M4.06's capital cycle comes back as the underwriting cycle.

Feeds into. M5.10's cross-sector capstone, whose right-lens exam includes insurance items. Phase 8 teardowns, where financials rotate through your weekly reps. And Phase 9, where float is the cleanest real-world case study in "leverage that can be safe", and in why that safety depends entirely on how carefully the insurer underwrites. The EV mindset (valuing a book of long-duration contracts by discounting its future profits) will also sharpen your DCF instincts in general.

Core teaching content

4.1 The inverted machine: float, or why you get paid to hold other people's money

Every business you studied before Phase 5 runs on the same working-capital treadmill. It spends cash first, on inventory, on salaries, on receivables. Customers pay later. In Phase 2 you measured that gap with the cash conversion cycle and learned to prize companies that push it negative. A retailer that sells goods before it pays suppliers is partly financed by its vendors, and for free.

Insurance is that inversion raised to an industrial scale and made the whole business model. The customer pays the full price, the premium, on day one. The product, a claim payment, is delivered later, sometimes decades later, and often never (most term-life policies expire without a death claim; most car-years pass without an accident). Between premium and claim, the insurer holds a pool of money that belongs, economically, to future claimants. Warren Buffett named this pool float: "money we hold but don't own." Berkshire Hathaway's float grew from $19m when Buffett bought National Indemnity in 1967 to roughly $170bn by the mid-2020s (per the 2024 shareholder letter; check the current figure in the latest one), and his letters return to the same two-part test:

1. How much float, and how durable? Long-tail lines such as liability and motor third-party hold premiums for years. Short-tail lines such as crop and health pay out within months. Growing float is a rising tide of investable money. Shrinking float is a repayment schedule.
2. What does the float cost? Cost of float = underwriting loss ÷ average float. If the insurer breaks even on underwriting, the float is a zero-interest loan. If it earns an underwriting profit (a combined ratio below 100%, which comes up again with general insurance), the cost is negative. In that case the insurer is being paid to hold the money it invests. That's better than free leverage. It's leverage with a negative interest rate. Buffett built an empire on it, and he calls float better than equity when the underwriting is disciplined.

Hold this frame all the way through. A life insurer is a machine that manufactures very-long-duration float plus mortality risk pooling. A general insurer manufactures shorter float plus annual risk pooling. For either one, you keep asking the same quality questions. Is the float growing, and will it last? Is it cheap (which takes careful underwriting), and is it invested sanely?

4.6 General insurance: the combined ratio decomposed

General insurance (P&C, property & casualty) sells one-year promises: motor, health, fire, marine, crop, liability. There's no 20-year profit emergence, so the accounting is less alien. But the estimate problem from the accounting section is at its worst here. A whole year's profit hinges on reserves for claims that haven't finished happening.

The premium cascade first. You'll meet four different "premium" lines, and using the wrong one corrupts every ratio:

• GWP / GDPI (gross written / gross direct premium income): everything sold. The market-share number.
• NWP (net written premium): GWP minus reinsurance ceded. What the insurer keeps at risk.
• NEP (net earned premium): NWP adjusted for the unearned premium reserve. A policy sold in March is only 1/12th "earned" by fiscal year-end. NEP is the true revenue of the year.
• The gap between written and earned matters most when growth is fast: written races ahead, earned lags.

The master metric:

Combined ratio (CR) = Loss ratio + Expense ratio · CR < 100% = underwriting profit, and the insurer's float is negative-cost. CR > 100% = underwriting loss. Then the float has a positive cost, which is tolerable only under the conditions in the float section below.

• Loss ratio = net incurred claims ÷ NEP (incurred = paid + change in outstanding reserves incl. IBNR, an estimate!)
• Expense ratio = (net commissions + operating expenses) ÷ NWP. That's the Indian convention (per ICICI Lombard and peers), which puts expenses on written premium while losses go on earned. US statutory does the same split, but US GAAP presentations often put both on earned. For a fast grower, the mixed-denominator detail moves reported CR by 1 to 3 points. So when you compare an Indian insurer with a US peer, recompute both on a common basis before you conclude anything.

Worked computation (synthetic "Meghdoot General"). Start with GWP of ₹10,000 cr. Take off ₹2,000 cr ceded and NWP is ₹8,000 cr. Unearned premium rose ₹400 cr, so NEP is ₹7,600 cr. Net incurred claims are ₹5,624 cr, net commissions ₹608 cr and opex ₹1,672 cr.

• Loss ratio = 5,624 ÷ 7,600 = 74.0%
• Expense ratio = (608 + 1,672) ÷ 8,000 = 2,280 ÷ 8,000 = 28.5%
• Combined ratio = 102.5%, so the underwriting result = 7,600 − 5,624 − 2,280 = −₹304 cr. On the all-earned basis it's 74.0% + 2,280/7,600 = 74.0% + 30.0% = 104.0%. That's the same company under a different convention, 1.5 points apart, so know which one you're reading.

Meghdoot loses 2.5 paise per rupee of premium on underwriting. Whether that makes it a bad business is the next question.

The underwriting cycle. P&C pricing is the capital cycle you met earlier, in miniature. Good years pull in a flood of capital from new entrants and reinsurance capacity. Price competition follows and CRs creep over 100. Then a shock burns capital. It might be a catastrophe or an inflation surprise. Capacity exits and prices harden. CRs fall, and the whole thing repeats every 5 to 10 years. A disciplined insurer shrinks when prices are soft. So premium growth well above the market in a soft phase is a red flag. Progressive's stated rule is the industry's gold standard of capital-cycle behaviour: grow as fast as possible at or below a 96% combined ratio, which means never buying growth with an underwriting loss.

4.8 Reserve development: the honesty meter

Every year-end reserve is a guess, and every year after that, reality grades the guess. Prior-year development (PYD) is the grade. If last year's reserves prove too low, the insurer books adverse development (strengthening), an expense running through this year's P&L for old business. If they prove too high, it books favourable development (releases), a credit that flatters this year.

The disclosure is the loss-development triangle: for each accident year, the estimate of ultimate losses re-stated at each subsequent year-end. US 10-Ks must show ten years of it, plus statutory Schedule P. Indian general insurers publish claims-development tables in annual reports and actuarial disclosures, though coverage varies, so verify company by company. Read it like this:

Mini-example. Accident-year 2023 ultimate losses were estimated at ₹1,000 cr at end-2023, ₹1,080 cr at end-2024, ₹1,150 cr at end-2025. So the company under-reserved by 15%. The ₹150 cr emerged as adverse PYD smeared over 2024–25, adding ~2 points to each of those years' reported CRs. It also means 2023's glossy reported CR of, say, 96% was really ~101% on an accident-year basis.

Hence the two ratios experts keep separate:

• Calendar-year CR (what's reported): this year's premiums vs this year's booked claims including PYD from old years.
• Accident-year CR: this year's business priced against this year's ultimate losses, excluding old-year noise. It shows how the insurer is really pricing today.

A company can hold its reported CR at 98% for years while its accident-year CR runs at 104%, just by releasing old reserves. That gap is finite. When the cookie jar empties, the reported CR snaps up 5+ points, "suddenly". Persistent modest favourable development is the fingerprint of a conservative reserver (Berkshire, Chubb, Progressive as archetypes). Persistent adverse development is institutionalised optimism, and it compounds, because growth makes each year's under-reserving bigger than the last correction.

The cautionary catalogue. GEICO nearly died in 1976 by under-reserving into an inflation spike. Rate hikes, reinsurance, and Buffett's investment rescued it. AIG booked a $5.6bn reserve strengthening for Q4 2016, then paid Berkshire's National Indemnity $10.2bn up front for protection against up to $20bn of further adverse development on old long-tail reserves. That single transaction is a monument to AIG's reserve dishonesty and, at the same time, to the value of float. Berkshire happily took $10.2bn to invest for years against claims payable slowly.

GE's legacy long-term-care book needed a reserve strengthening of roughly $15bn, announced in 2018 (statutory, phased over years). Its long-tail promises were priced in the 1990s on assumptions (lapse rates, care costs, interest rates) that all broke the wrong way.

Inflation is the universal enemy of stale reserves. Claims settle at future prices, and the reserves were set at past ones. Today's live versions are US "social inflation" (jury awards, litigation funding) and wage-linked awards in Indian motor third-party.

4.10 The red-flag panel

Run these ten checks on any insurer in fifteen minutes. The first four are the brief's core, so know them cold.

Common mistakes & how experts think differently

1. Valuing a life insurer on statutory P/E. The rookie sees a life insurer at 90× reported earnings and calls it absurd. The expert knows new-business strain suppresses statutory profit in exact proportion to how fast the company is writing profitable business, so they reach for P/EV, RoEV, and VNB growth instead. It works the other way too. A life insurer whose statutory profits are suddenly gushing may simply have stopped growing.
2. Reading EV as book value. EV already capitalises the future profits of the in-force book. It's a DCF output wearing a balance-sheet costume. So comparing P/EV with another sector's P/B double-counts, and so does adding "EV + a multiple of earnings". The only thing a P/EV premium can buy is future new business.
3. Treating the combined ratio as sufficient. CR without float is half an equation. A 102% long-tail motor-TP book can beat a 99% crop book economically, as the two-scenario float arithmetic showed. Experts always attach the float, its size, tail and investment posture, before judging the CR.
4. Taking the reported CR at face value. That number is management's estimate, graded by management's later estimates. Experts read three years of reserve development first, and only then do they believe the CR or restate it. If accident-year and calendar-year CRs diverge, trust the accident year.
5. Applauding APE growth without the mix table. Distribution engines sell what the season rewards: ULIPs in bull markets, guarantees when rates spike. Experts multiply each product's APE by its margin before applauding. Back in the KPI panel, the mini-computation showed identical APE producing 21% or 26% blended margins on mix alone.
6. Confusing economic and operating variance in the EV walk. A bull market lifts EV through economic variance, and that's weather. Management didn't earn it. Experts grade management only on VNB, operating variances, and assumption honesty, and they treat recurring positive "assumption changes" as a flag.
7. Assuming insurance float is automatically Buffett float. Buffett's float is valuable because it's long-duration, growing, and generated at negative cost by disciplined underwriting. Short-tail float at a 110% CR is just expensive borrowed money. The framework only travels with all three adjectives attached.
8. Porting bank intuitions onto solvency. A bank at regulatory-minimum capital is in crisis. An insurer at 160% solvency is merely constrained. Its liabilities can't run, since policyholders have no depositor-style exit. But growth stalls and guarantees get riskier. Direction and speed of drift matter more than distance to the line. Four quarters of steady decline toward 150% tell you more than the level itself does.
9. Extrapolating one year of claims experience. One benign monsoon, one COVID spike, one cat-free hurricane season: single-year loss ratios are weather. Experts normalise, using 5-year average loss ratios by segment and staying cycle-aware, the same way Phase 3 taught you to normalise cyclical earnings.

Worked examples

Worked example 1: Vanya Life: the full life KPI panel from disclosures

Synthetic company; numbers designed to be computed by hand. Here's what Vanya Life's FY26 disclosure gives you. Individual annualised regular new-business premium is ₹4,600 cr and single premium ₹4,000 cr. VNB is ₹1,375 cr and opening EV ₹18,000 cr. Its 13th-month persistency is 84%, down from 88% two years ago. ASM is ₹7,440 cr against RSM of ₹4,650 cr. The EV walk runs unwind ₹1,530 cr, operating variances +₹70 cr, assumption changes −₹100 cr, economic variance +₹450 cr, dividends paid ₹250 cr. By APE, the product mix is ULIP 44% (up from 30% two years ago), par 16%, non-par 22%, protection 12%, annuity 6%.

Step 1. APE. 4,600 + 0.10 × 4,000 = ₹5,000 cr.

Step 2. VNB margin. 1,375 ÷ 5,000 = 27.5%. That's above the 25 to 30% "strong" bar, but check the trend against mix before you applaud (Step 6 does that).

Step 3. EV operating profit and RoEV. Operating profit = unwind + VNB + operating variances + assumption changes = 1,530 + 1,375 + 70 − 100 = ₹2,875 cr. RoEV = 2,875 ÷ 18,000 = 16.0%, which is healthy. Look at the −₹100 cr assumption change and read the EV report's note on it. Here it's an expense-assumption true-up (honest housekeeping) and no persistency downgrade.

Step 4. Closing EV. 18,000 + 2,875 + 450 − 250 = ₹21,075 cr. Markets rallied, so the +450 economic variance is inside closing EV but outside RoEV. It's rightly excluded from the skill read.

Step 5. Solvency. 7,440 ÷ 4,650 = 160%. That's above the 150% floor, but thin for a grower. If the non-par push continues, RSM rises. Model two more years of 20% RSM growth with retained earnings adding only ₹600 cr/yr to ASM, and the ratio pierces 150 within ~2 years. Flag 4 is live.

Step 6. Mix and the margin's quality. ULIP has gone from 30% to 44% of APE, yet the blended margin held at 27.5%. How? Only if someone raised the protection and non-par margins (through pricing or through assumptions) or VNB benefited from assumption choices. So cross-check flag 9. Were there positive "economic assumption" tweaks inside VNB? Now add persistency sliding from 88% to 84%. That's flag 2 territory, below the 85% healthy bar and still falling.

Put together, the panel reads like this: a growth engine tilting toward mis-sale-prone, low-value products, a headline VNB margin being defended optically, and a capital buffer thinning. Verdict: the 16% RoEV is real today, but three of the four brief red flags are lit amber. This is a "premium multiple not deserved" profile. If the market pays 2.5× EV here, the implied new-business multiple is pre-paying for a franchise whose distribution quality is visibly decaying.

Worked example 2: Meghdoot General vs Kaveri General: same combined ratio, opposite businesses

Synthetic pair. Both report CR = 102.5% on NEP of ₹7,600 cr, so on the all-earned basis each has an underwriting loss of ₹190 cr. Keep it simple and use −₹190 cr for both. Net worth is ₹5,000 cr each, and tax is 25%.

So the combined ratios match. One business compounds above its cost of equity and the other destroys value. Now stress Meghdoot's reserve honesty. Suppose its triangle shows accident-year ultimates re-estimating upward 6% per year. Its true accident-year CR ≈ 102.5 + ~4 pts ≈ 106.5%, so the underwriting loss ≈ ₹494 cr and the cost of float ≈ 4.1%. That's still cheaper than borrowing. But ROE falls to ~11% [(1,224 − 494) × 0.75 ÷ 5,000], and the trajectory is wrong, because under-reserving and growing means each correction is bigger than the last. The lesson stacks. Start from the CR and attach the float. Then verify the reserves, and only then judge.

Worked example 3: Pricing the franchise: HDFC Life-style vs LIC-style P/EV (illustrative, ~FY24 scale, pull current filings)

Start with H. The market pre-pays ~27 years of current-scale VNB. That's defensible only if VNB compounds in the mid-teens at stable margins for a decade-plus while persistency and solvency stay clean. Run the reverse-DCF instinct from Phase 3. If VNB grows 15% for 10 years then fades, and you discount at ~12%, the franchise is worth roughly 20 to 30× today's VNB. So ~27× prices in continued excellence, no accidents.

Now L. Its price is below runoff value, because the market is haircutting the EV's own assumptions (persistency, expense overruns, par-fund politics, state-directed capital use). Neither number is a verdict. Each is an implied story, and you now know how to interrogate it. The gap between 2.98× and 0.85× comes down to RoEV, margin, mix, and governance, which are the variables your KPI panel measures.

Worked example 4: Summit Casualty (synthetic): a US P&C insurer, and why a 96% combined ratio can still be flattered ($)

Synthetic US commercial-lines insurer; US GAAP, so both ratios sit on earned premium (no Indian written/earned split). All figures $m. Summit's FY26 shows net earned premium of $10,000, net incurred claims of $7,000 and underwriting expenses (commissions + opex) of $2,600. It holds float of $14,000 against long-tail commercial-auto and liability reserves, net worth of $6,000, and $20,000 of investments yielding 4.0%. US federal tax is 21%.

Step 1. The reported (calendar-year) combined ratio. Loss ratio = 7,000 ÷ 10,000 = 70.0%, and expense ratio = 2,600 ÷ 10,000 = 26.0%. So CR = 96.0%, and the underwriting profit = 10,000 − 7,000 − 2,600 = +$400m. On the surface that's a Progressive-grade print, the ~96 rule made flesh.

Step 2. Attach the float, finish the ROE. Investment income = 20,000 × 4.0% = $800m. Pre-tax = 400 + 800 = $1,200m, and PAT = 1,200 × (1 − 0.21) = $948m, so ROE = 948 ÷ 6,000 = 15.8%. Cost of float = −400 ÷ 14,000 = −2.9%. With an underwriting profit, Summit is paid to hold $14,000m of investable money. That's Buffett's negative-cost float, in dollars.

Step 3. Grade the print's honesty. Summit's claims line includes $300m of favourable prior-year development, old reserves that proved redundant and were released into this year. Strip it out. Current accident-year incurred = 7,000 + 300 = $7,300m, the accident-year loss ratio is 73.0%, and the accident-year CR = 99.0%. Underwriting profit is only +$100m, and the cost of float is −0.7%. So 3 of the reported 4 points of underwriting margin were borrowed from the past. This year didn't earn them. That's flag 6, in dollars.

Lesson. Read even a disciplined-looking sub-100 print twice. Start from the CR, then attach the float (size, tail, cost), then strip prior-year development, and only then judge the accident-year economics. A 96% that's really 99% on this year's business is still fine, given cheap durable float. But you only know that after the triangle.

Worked example 5: Ganga Life (synthetic): the VNB-margin mix bridge (₹)

Synthetic Indian private life insurer; numbers built to decompose by hand. All figures ₹ cr. Ganga's APE grows 27% year on year and its VNB rises. Its deck leads with "record new-business value". Yet the blended VNB margin falls. A mix bridge shows why.

Step 1. The headline. Blended margin Y1 = 655 ÷ 2,200 = 29.8%, and Y2 = 692 ÷ 2,800 = 24.7%. VNB grew only from 655 to 692 (+5.6%) while APE grew +27.3%, and the blended margin fell −5.06 pts. "Record VNB" masks stagnating value created per rupee sold.

Step 2. Decompose the −5.06 pts (shift-share). Split the change in two. A mix effect lets the weights move and holds each product's margin at Y1. A rate effect lets the margins move, at Y2 weights:

• Mix effect = Σ (w₂ − w₁) × m₁ = −5.49 pts. ULIP's share exploded from 22.7% to 42.9% of APE and dragged the blend down.
• Rate effect = Σ w₂ × (m₂ − m₁) = +0.43 pts. Only protection re-priced, from 65% to 68%, and that added a little back.
• Total = −5.49 + 0.43 = −5.06 pts, which ties.

Lesson. Almost the entire margin decline is mix. Headline APE growth was bought with low-value ULIPs, which is flag 1. And the +0.43 pt "defence" from a higher protection margin is real only if it came from actual re-pricing. If it came from softened assumptions inside VNB, it's flag 9 wearing flag 1's clothes. Run the bridge before crediting or blaming any VNB-margin move.

Practice set

Problems ramp from guided to independent to timed. Work each one before you read the solution.

P1 (guided). An insurer writes ₹1,800 cr of annualised regular first-year premium and ₹5,500 cr of single premium (mostly annuities). Compute APE. Why does the convention take only 10% of single premium?

Solution. APE = 1,800 + 0.10 × 5,500 = 1,800 + 550 = ₹2,350 cr. Single premium arrives once, while regular premium recurs for the policy's paying term. Weighting single at 10% roughly equates one-shot money with the annualised value of a recurring stream, so volume is comparable across insurers with different mixes. It's only a convention, though. VNB values each contract properly, so it outranks APE.

P2 (guided). From P1's insurer: VNB ₹564 cr. Compute VNB margin and grade it against the threshold. Its APE is 40% annuity, 35% ULIP, 25% par. Does the margin make sense, and what would you check?

Solution. 564 ÷ 2,350 = 24.0%, just under the 25 to 30% "strong" bar and respectable. But look at the mix. Heavy annuity at ~25% margin, heavy ULIP at ~10% and par at ~15% imply a blended margin of roughly 0.40×25 + 0.35×10 + 0.25×15 ≈ 17%. A reported 24% on that mix means one of two things. Either the product margins beat typical levels, which is possible through pricing or expense efficiency, or there's assumption generosity inside VNB. Check the EV report's sensitivity table (VNB at −50bps rates, −10% persistency) and whether assumption changes contributed. Never accept a margin that beats its own mix without an explanation.

P3 (guided). Here's a life insurer's cohort data. Premium due at month 13 from last year's regular-premium cohort is ₹2,400 cr, and ₹1,992 cr was renewed. Its 61st-month persistency is 41%. Compute 13th-month persistency and diagnose.

Solution. 1,992 ÷ 2,400 = 83%, below the 85% healthy bar. And with the 61st month at 41% (<50%), customers drop early and keep dropping. That's classic savings-product mis-selling through push distribution (the targets are >85% and >50%). The booked VNB assumed renewal streams that won't arrive. Future EV walks will show negative persistency variances, and RoEV will erode. Also check policy-count against premium persistency. If count-persistency is much lower, small-ticket customers, the most mis-sale-prone, are lapsing fastest.

P4 (guided). Build a mini-EV: ANW ₹2,200 cr. In-force expected after-tax distributable profits: ₹500 cr in year 1, ₹475 in year 2, ₹450 in year 3, then a level ₹430 cr perpetuity from year 4. RDR 12%. Deductions for cost of capital and non-hedgeable risks: ₹380 cr. Compute EV.

Solution. For years 1–3, PV = 500/1.12 + 475/1.2544 + 450/1.4049 = 446.4 + 378.6 + 320.3 = ₹1,145.3 cr in total. At end-year 3 the perpetuity is worth 430 ÷ 0.12 = ₹3,583.3 cr, and its PV = 3,583.3 ÷ 1.4049 = ₹2,550.6 cr today. Gross VIF = 1,145.3 + 2,550.6 = ₹3,695.9 cr, and net VIF = 3,695.9 − 380 = ₹3,315.9 cr after deductions. EV = 2,200 + 3,315.9 ≈ ₹5,516 cr. Real EVs come from policy-level actuarial models, and this is just the logic behind them.

P5 (independent). Take P4's insurer a year later. Unwind is ₹560 cr, VNB ₹610 cr, persistency variance −₹90 cr and expense variance +₹25 cr. Assumption changes add +₹140 cr (mortality assumption relaxed), economic variance is −₹210 cr (rates rose), and the dividend is ₹120 cr. Compute EV operating profit, RoEV, and closing EV, then write the two-sentence quality read.

Solution. EV operating profit = 560 + 610 − 90 + 25 + 140 = ₹1,245 cr, and RoEV = 1,245 ÷ 5,516 = 22.6%. Closing EV = 5,516 + 1,245 − 210 − 120 = ₹6,431 cr. Now the quality read. The printed RoEV is flattered, because ₹140 cr (11% of operating profit) came from relaxing an assumption while actual persistency ran ₹90 cr worse than assumed. Skill-based RoEV ex-assumption-changes is (1,245 − 140) ÷ 5,516 ≈ 20%. And the variances point the wrong way, with reality worse and assumptions softer. That's flag 9. The −210 economic variance is weather, so ignore it for skill.

P6 (independent). Solvency drill. ASM ₹5,580 cr, RSM ₹3,100 cr today. Under the business plan, RSM grows 18%/yr (guarantee-heavy growth) and retained earnings add ₹310 cr/yr to ASM. Compute today's ratio and the ratio after two years. In which year does it breach 150%, and what are management's three options?

Solution. Today: 5,580 ÷ 3,100 = 180%. Year 1: ASM 5,890, RSM 3,658 → 161%. Year 2: ASM 6,200, RSM 4,316 → 143.6%: breach in year 2. Management has three options. (1) Slow growth, or re-mix it away from capital-hungry guarantees, which cuts RSM growth. (2) Raise equity or subordinated capital, which raises ASM but dilutes. (3) Reinsure more. That cedes risk and RSM, but it cedes profit too. Solvency ratios die by trajectory, and the correction always costs either growth or shareholders.

P7 (independent). A general insurer discloses GWP of ₹14,000 cr, ceded premium of ₹3,200 cr and an increase in unearned premium reserve of ₹600 cr. Net incurred claims are ₹7,824 cr, net commissions ₹972 cr and opex ₹2,214 cr. Compute NWP, NEP, loss ratio, expense ratio (Indian convention), combined ratio, and the underwriting result. Then recompute CR on an all-earned basis.

Solution. NWP = 14,000 − 3,200 = ₹10,800 cr. NEP = 10,800 − 600 = ₹10,200 cr. Loss ratio = 7,824 ÷ 10,200 = 76.7%. Expense ratio = (972 + 2,214) ÷ 10,800 = 3,186 ÷ 10,800 = 29.5%. CR = 106.2%. Underwriting result = 10,200 − 7,824 − 3,186 = −₹810 cr. All-earned CR = 76.7% + 3,186/10,200 = 76.7% + 31.2% = 107.9%, which is 1.7 points worse. Fast-growing insurers look better under the written-premium convention, because unearned growth inflates the expense denominator. State the convention whenever you quote a CR.

P8 (independent). P7's insurer holds float of ₹16,000 cr and net worth of ₹6,000 cr, investing ₹22,000 cr at 7.3% in high-grade bonds. Tax 25%. Compute cost of float, ROE, and rule on the "106.2% CR is fine" claim. Then re-rule if ₹9,000 cr of the float sits against crop and group-health claims payable within 9 months.

Solution. Cost of float = 810 ÷ 16,000 = 5.1%. That's below the ~7% G-sec yield, so the float still adds value, but the margin of cheapness is thin. Investment income = 22,000 × 7.3% = ₹1,606 cr. Pre-tax = 1,606 − 810 = ₹796 cr, PAT = ₹597 cr, and ROE = 10.0%. That's around or below the cost of equity. Call it marginal. It doesn't earn the word "fine."

Now suppose ₹9,000 cr of the float is short-tail. The durable long-tail float is then only ₹7,000 cr, and the insurer is effectively running a recurring 106 CR on lines that generate no lasting float. Those lines are simply unprofitable products. The true sustainable ROE is dreadful: long-tail float 7,000 + equity 6,000 = 13,000 × 7.3% = 949, minus 810, ×0.75 → ₹104 cr, or ~1.7% ROE. Verdict: not fine. Fix the underwriting or exit the short-tail lines.

P9 (independent). Reserve honesty. A motor insurer's triangle: accident-year 2024 ultimates estimated ₹4,000 cr (end-24) → ₹4,240 cr (end-25) → ₹4,420 cr (end-26). It reported calendar-year CRs of 97% (2024), 99% (2025), 100% (2026) on NEP of ₹9,000 / ₹10,000 / ₹11,000 cr. Compute the adverse development recognised in 2025 and 2026, the points it added to those years' CRs, and the approximate true accident-year 2024 CR. What is the pattern's name and consequence?

Solution. Start with the development. In 2025 the insurer booked 4,240 − 4,000 = ₹240 cr of strengthening (+2.4 pts on ₹10,000 cr NEP). In 2026 it booked 4,420 − 4,240 = ₹180 cr (+1.6 pts on ₹11,000 cr). AY-2024's ultimate is ₹420 cr (10.5%) above the initial estimate. Restated on 2024's NEP, its true CR ≈ 97% + 420/9,000 ≈ 101.7%, against the 97% printed. It's a pattern called persistent adverse development. Those reported CRs were fiction, with each year's print borrowed from the next. And the calendar-year CRs, running 97 to 99 to 100, understate the deterioration, because they mix old strengthening with new-year optimism. So assume the current accident year is under-reserved too until it's proven otherwise. Restate your valuation on accident-year economics.

P10 (timed, 12 minutes). Life insurer "Tunga Life" has EV of ₹26,000 cr, VNB of ₹1,690 cr on APE of ₹6,500 cr, a market cap of ₹65,000 cr, and an RoEV of 17%. Its 13th-month persistency is 87% and its solvency 196%. Its peer "Bhadra Life" has EV of ₹30,000 cr, VNB of ₹1,200 cr on APE of ₹6,000 cr, a market cap of ₹39,000 cr, and an RoEV of 11%. Its 13th-month persistency is 81% and its solvency 158%. Compute P/EV, VNB margin, and implied NB multiple for both. Then write two sentences: which is "cheaper," and is it?

Solution. Tunga's P/EV = 65,000/26,000 = 2.50×, its margin = 1,690/6,500 = 26.0%, and its implied NB multiple = (65,000 − 26,000)/1,690 = 23.1×. Bhadra's P/EV = 39,000/30,000 = 1.30×, its margin = 1,200/6,000 = 20.0%, and its implied NB multiple = (39,000 − 30,000)/1,200 = 7.5×. Bhadra looks cheaper on every multiple. But read its panel. RoEV of 11% is about the cost of equity, persistency is below the healthy bar, and solvency has drifted to within 8 points of the floor. Bhadra's machine adds little value and may need capital, so a near-1× EV is earned. It isn't a bargain. Tunga's 23× NB multiple pre-pays a decade-plus of strong VNB compounding. Expensive quality against cheap mediocrity is the sector's standing trade-off, and it's the analysis that decides, whatever the multiple says.

P11 (timed, 10 minutes). Red-flag sweep. Here's a general insurer's five-year record. GDPI CAGR is 24% (industry 13%), led by group health, which is now 46% of the book at a 104% loss ratio. The reported CR is steady at 99 to 100%. The reserve triangle shows cumulative favourable releases of 3 to 4 pts of NEP per year from motor-TP reserves. Equity capital-gains booked each year ≈ 60 to 90% of PBT. Solvency has gone 172% → 158% → 151% over three years. List every red flag with its number from the red-flag panel and give the one-line verdict.

Solution. Start with #7, growth bought via group health at a 104% loss ratio. Then #6, a reported CR that's flat only because of TP reserve releases, while the accident-year CR is ~103 to 104%. Next comes something #5-adjacent. Releasing 3 to 4 pts/yr from a long-tail line while growing fast is aggressive reserving until the triangle proves otherwise. Then #3, a CR ≈ 100% plus profit dependence on equity gains, which is underwriting masked by investment gains. And #4 is solvency drifting to 151%, one bad quarter from the floor. Verdict: a company buying market share with underpriced group business, propping the optics with reserve releases and market gains, and running out of capital. It's the full anatomy of the next New-India-style chronic underwriter, or worse. Avoid it, and mark the file for the "what breaks it" teardown drill.

P12 (guided). A US life insurer, Cypress Financial (synthetic), reports GAAP common equity of $32,000m, which includes accumulated other comprehensive income (AOCI) of −$6,000m from unrealized losses on its available-for-sale bond book after rates rose. Adjusted operating income is $4,940m, and market cap is $46,800m. It publishes no embedded value. Compute book value ex-AOCI, ROE on that base, and P/B ex-AOCI, and say why the ex-AOCI adjustment is made.

Solution. Book ex-AOCI = 32,000 − (−6,000) = $38,000m. Adding back the negative AOCI removes the bond mark-to-market, which reverses as bonds pull to par. ROE = 4,940 ÷ 38,000 = 13.0%, and P/B ex-AOCI = 46,800 ÷ 38,000 = 1.23×. So you're looking at 1.23× book against a 13% ROE, with a ~9 to 10% US cost of equity. That's the excess-return premium in GAAP clothing, the same story P/EV tells for an Indian insurer, one metric down. Why strip AOCI? Leaving the −$6,000m in would shrink the equity base and inflate both ROE and P/B on pure rate noise. Strip it, and the multiple reflects operating economics instead of wherever the yield curve sat on the reporting date. Don't compare this 1.23× directly with an Indian insurer's 2.5× EV, though. EV already capitalises future in-force profit, and book value doesn't.

P13 (guided). Yamuna Life (synthetic): embedded value ₹32,000 cr, VNB ₹1,600 cr on APE ₹6,000 cr, market cap ₹80,000 cr, RoEV 18%, cost of equity 12%. Compute P/EV, VNB margin, and the implied new-business multiple, then run the coherence test.

Solution. P/EV = 80,000 ÷ 32,000 = 2.50×. VNB margin = 1,600 ÷ 6,000 = 26.7%, which clears the 25 to 30% strong bar. Implied NB multiple = (80,000 − 32,000) ÷ 1,600 = 30.0×, so the market pre-pays ~30 years of current-scale VNB for the future franchise. Now run the coherence test from the valuation section. If RoEV merely equalled the cost of equity and new business added nothing, fair P/EV ≈ 1×. Every turn above that has to be earned, either by RoEV above the cost of equity or by a valuable VNB franchise. Yamuna has both. Its RoEV is 18% against a 12% cost of equity, a real 6-pt spread, so the in-force book compounds value on its own. And its VNB franchise runs a 26.7% margin and is growing. 2.50× is defensible only if that spread and mid-teens VNB growth persist for a decade-plus. Treat the 30× NB multiple as the reverse-DCF question from Phase 3, something to interrogate. It isn't a verdict.

P14 (independent). Tapti Life (synthetic) booked ₹900 cr of VNB on last year's cohort, of which ₹540 cr is the present value attributable to renewal-year profits, priced on an assumed 13th-month persistency of 85%. Actual 13th-month persistency lands at 76%. Give a first-order estimate of the VNB leakage (the operating variance) and its drag on RoEV if opening EV is ₹22,000 cr. Why is the true hit likely larger?

Solution. The surviving renewing base is 76 ÷ 85 = 89.4% of what was assumed, so the renewal-linked value scales to 540 × 0.894 = ₹482.8 cr. Leakage ≈ 540 − 482.8 = −₹57.2 cr of negative operating variance. RoEV drag = 57.2 ÷ 22,000 ≈ 0.26 pts off this year's operating return. Why is the real hit bigger? A full actuarial model compounds the shrunken surviving base across every later renewal (months 25, 37, 49…). So a year-1 persistency miss cascades through the whole future stream, well past year 1. That's the straight line from a mis-sold cohort to a future EV write-down. Also split premium- from count-persistency. If small policies lapse fastest, count-persistency is worse and the mis-selling read is sharper.

P15 (independent). Narmada Life (synthetic) holds ₹50,000 cr of non-par reserves guaranteeing policyholders an effective 6.2% a year for the next two decades. Its backing assets currently earn 6.9% (bonds bought when rates were high). Compute today's spread. Then rates fall. As bonds mature and roll over, the book's earned yield drifts to 5.0%. Compute the new spread and the swing. What single disclosure settles whether this is a crisis?

Solution. Today's spread = (6.9% − 6.2%) × 50,000 = +₹350 cr/yr, comfortably positive. At a 5.0% earned yield it becomes (5.0% − 6.2%) × 50,000 = −₹600 cr/yr. That's a negative spread (gyaku-zaya), and every policy is now a bond the insurer issued at an above-market coupon it can no longer earn. The swing is 350 − (−600) = ₹950 cr/yr of deterioration on this one book. You settle it with one disclosure, the asset-liability duration match (or hedging) in the EV report. If the guarantees are cash-flow matched with long bonds and forward-rate agreements, the asset yield is locked and the spread is protected. If the book is short-duration against 20-year promises, this is Japan's late-1990s trap forming (flag 8). Never applaud guaranteed-non-par growth in a high-rate year without seeing the hedge.

P16 (independent). Cascade Mutual (synthetic), a US personal-auto insurer, reports calendar-year combined ratios of 97%, 97%, 98% over three years on net earned premium of $8,000m, $9,000m, $10,000m. Its loss-development triangle shows favourable prior-year development (reserve releases) reducing each calendar year's incurred claims by $240m, $270m, and $150m respectively. Compute the accident-year CR for each year, and state what the trend tells you.

Solution. As points of NEP, the releases are 240 ÷ 8,000 = 3.0 pts, 270 ÷ 9,000 = 3.0 pts and 150 ÷ 10,000 = 1.5 pts. Releases reduce reported (calendar-year) incurred, so the accident-year CR is the reported CR plus those points: 100.0%, 100.0%, 99.5%. Those flattering 97–98% prints were bought with prior-year releases, and the current-year book is running right at breakeven. Watch the trend. The release contribution has shrunk from 3.0 to 3.0 to 1.5 pts. The cookie jar is emptying, and when it runs dry the reported CR snaps up 1.5 to 3 pts with no change in the underlying business (flag 6). Trust the accident-year line.

P17 (timed, 12 minutes). Godavari General (synthetic), an Indian multi-line insurer, reports GWP of ₹18,000 cr, reinsurance ceded of ₹3,600 cr and an increase in unearned premium reserve of ₹900 cr. Net incurred claims are ₹10,395 cr, net commissions ₹1,440 cr and opex ₹3,168 cr. It holds float of ₹20,000 cr against net worth of ₹8,000 cr, investing ₹28,000 cr at 7.1% in high-grade bonds. Tax is 25%, and G-secs yield ~7%. Compute NWP, NEP, loss ratio, expense ratio (Indian convention), combined ratio, the underwriting result, the all-earned CR, investment income, PAT, ROE, and the cost of float. Then rule on "the big float makes this fine," and adjust if the triangle shows ~2 pts/yr of adverse development.

Solution. NWP = 18,000 − 3,600 = ₹14,400 cr, and NEP = 14,400 − 900 = ₹13,500 cr. Loss ratio = 10,395 ÷ 13,500 = 77.0%. Expense ratio = (1,440 + 3,168) ÷ 14,400 = 4,608 ÷ 14,400 = 32.0%. So CR = 109.0%. Underwriting result = 13,500 − 10,395 − 4,608 = −₹1,503 cr. All-earned CR = 77.0% + 4,608 ÷ 13,500 = 77.0% + 34.1% = 111.1%. Investment income = 28,000 × 7.1% = ₹1,988 cr. Pre-tax = −1,503 + 1,988 = ₹485 cr, PAT = 485 × 0.75 = ₹363.75 cr, and ROE = 363.75 ÷ 8,000 = 4.55%. Cost of float = 1,503 ÷ 20,000 = 7.5%, above the ~7% sovereign yield. Unlike Meghdoot, with its cheap float, this insurer is "borrowing" at a higher rate than the government.

Verdict: not fine. A 109% CR overwhelms even a large float. Its 4.55% ROE is well below any plausible cost of equity, so the business destroys value, and it fails float tests 1 and 4. Add ~2 pts/yr of adverse development and the accident-year CR runs ~111%+, while the cost of float climbs past 8%. The trajectory worsens too, as growth compounds the under-reserving. It's a chronic-underwriter profile to avoid.

Applied mini-project: the insurance double workup

Task. Run both playbooks on real companies: one Indian life insurer, and one Indian + one US general insurer. Timebox: ~4.5 hours.

Leg 1. Life (HDFC Life or SBI Life; your choice). From the company's website, pull the latest annual Embedded Value disclosure/report (externally reviewed), the investor presentation (product mix, persistency, VNB walk), and the annual report. Take the current market cap from NSE/BSE or screener.in.

1. Build the KPI panel. Start with APE and 3-year APE CAGR. Then do VNB and VNB margin, with a 3-year trend. Lay out the full EV roll-forward for the latest year, separating unwind / VNB / operating variances / assumption changes / economic variances, and compute RoEV two ways: as printed, and excluding assumption changes. Add persistency at months 13 and 61, with the 3-year trend against the >85% / >50% bars. Add the solvency ratio's trend, its distance and its direction relative to 150%. Finish with a product-mix table by APE, giving each bucket's disclosed or estimated margin.
2. Compute P/EV and the implied new-business multiple. Then write three sentences on what the price implies and whether the panel supports it.
3. One page, Minto-style: verdict first, then the three numbers that carry it, then the two risks that could break it.

Leg 2. General (ICICI Lombard vs Progressive). Work from ICICI Lombard's annual report + investor presentation and Progressive's 10-K (SEC EDGAR, with their shareholder letters as a bonus):

1. For each: premium cascade (GWP/GDPI → NWP → NEP), loss ratio, expense ratio, combined ratio. State the convention you used, and put both companies on the same all-earned basis.
2. Decompose ROE into underwriting result and investment income. Estimate float and investment leverage, then compute the cost of float.
3. Reserve honesty. Using ICICI Lombard's claims-development table and Progressive's 10-K loss-development disclosure, has development been favourable or adverse over the last 3 years? Adjust your view of the reported CR to match.
4. Half page: which is the better business (not the cheaper stock), and what would change your mind?

Scoring rubric (0 = missing/wrong · 1 = computed but shallow · 2 = correct and interpreted):

Pass at ≥80% of available points (13/16), the programme's written-deliverable standard. For the gate itself, see the mastery check below.

Reading & resources

• Warren Buffett's Berkshire Hathaway shareholder letters: the float education. Start with the 2014 letter's "Berkshire — Past, Present and Future" insurance section, then read the insurance discussion in any five letters across decades and watch the float table grow. berkshirehathaway.com/letters. [Free] [Beginner→Intermediate]
• Lawrence Cunningham, The Essays of Warren Buffett: the insurance/float essays collected thematically. [Paid] [Intermediate]
• IRDAI Annual Report and Handbook on Indian Insurance Statistics (irdai.gov.in): market shares, segment loss ratios, solvency data, penetration. It's the primary dataset for every India claim you've read here. [Free] [Intermediate]
• Company actuarial disclosures: HDFC Life / SBI Life / ICICI Prudential Life Embedded Value reports and investor presentations (each company's investor-relations page). Read one full EV report end to end, once. It's the life insurer's true annual report. [Free] [Intermediate]
• ICICI Lombard annual report (incl. the claims-development table) and quarterly investor presentations: the India GI panel in the wild. [Free] [Intermediate]
• Progressive Corporation annual reports & shareholder letters (progressive.com; 10-K on SEC EDGAR): the cleanest articulation of underwriting discipline (the 96 rule) and a model loss-development disclosure. [Free] [Intermediate]
• Aswath Damodaran, "Valuing Financial Service Firms" (free paper at stern.nyu.edu/~adamodar) + the financial-service-firm chapters of Investment Valuation: why book/EV-based lenses replace FCFF for insurers. [Free paper / Paid book] [Advanced]
• Society of Actuaries / Institute of Actuaries of India papers on Embedded Value (e.g., SOA's "Embedded Value: Practice and Theory"): the methodology beneath the life KPI panel, for when you want the machinery. [Free] [Advanced]
• Swiss Re Institute sigma reports (swissre.com/institute): world insurance data, protection-gap estimates, cycle commentary; the global context layer. [Free] [Intermediate]
• Insurance Information Institute (iii.org): US P&C industry data, combined-ratio history, cat losses. [Free] [Beginner]
• Deloitte IAS Plus: IFRS 17 resources (iasplus.com): when you meet a CSM in an AIA- or European-style report and want the mapping to EV concepts. [Free] [Advanced]
• Zerodha Varsity: Personal Finance module (insurance chapters): Indian product taxonomy from the buyer's side. It's useful for seeing how the products you now analyse are actually sold. [Free] [Beginner]
• Going deeper: a real US P&C ten-year loss-development triangle (e.g., Chubb, Travelers, or Progressive; the "development of incurred losses" schedule in the 10-K on SEC EDGAR) [Free] [Advanced]. Take one actual table, restate three accident years, and compute calendar-year vs accident-year combined ratios yourself. Reading reserve development only becomes reflex once you've graded a real reserver's guesses against how they later resolved.
• Going deeper: post-mortems of Japan's 1997–2001 life-insurer failures (Geneva Association working papers; Financial Services Agency of Japan histories; searchable academic surveys of gyaku-zaya) [Free] [Advanced]. This is the evidence behind the negative-spread lesson in the product taxonomy (Nissan Mutual, Chiyoda, Kyoei). Read it before you judge any Indian guaranteed-non-par build-out written near a rate peak.

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": "Life and general insurance end to end: APE, VNB margin and the mix bridge, the EV walk and RoEV, solvency against the floor, the premium cascade and the combined ratio on both conventions, float economics and the cost of float, prior-year development and the calendar-year versus accident-year distinction, P/EV and the implied new-business multiple, and the ten-flag panel",
  "use": "Insurance disclosure is voluminous, formulaic and scattered across documents nobody aggregates — which is the exact profile of work worth handing over. (1) TRIANGLE EXTRACTION: the claims-development tables in an Indian annual report and the ten-year incurred-and-paid development tables in a US 10-K are dense grids that a document tool reads reliably and a human transcribes badly; ask for them as tidy long-form rows (accident year, development year, cumulative incurred, cumulative paid) with a page cite each. (2) The EV-WALK reconstruction across five years: unwind, VNB, operating variances, assumption changes, economic variances, dividends — so RoEV is computed on the operating line and the flag-9 pattern (positive assumption changes year after year while operating variances run negative) becomes visible as a series rather than a sentence. (3) SEGMENT DECOMPOSITION for the mix bridge: product-wise APE and margin for life, line-wise GDPI and loss ratio for general, which in India means parsing the mandated public disclosure forms quarter by quarter. (4) The DEFINITION and convention audit: which combined ratio is this, expenses on written or earned, is 'health' retail or group, does 'persistency' use premium or policy counts. What it must not do is set an actuarial assumption. The tail factor, the loss-development pattern, the discount rate in an EV — those are the judgements the entire valuation rests on, and a model asked for one will supply a confident number with no basis whatever.",
  "tools": ["Document/PDF AI — for claims-development triangles, EV walks and segment tables out of annual reports, EV reports and 10-Ks", "Chat assistants — Claude, ChatGPT for convention audits (written-vs-earned, calendar-vs-accident-year, retail-vs-group) and definition diffs across quarters", "Coding assistants — for the chain-ladder engine, the tail-factor sensitivity and the break-even-combined-ratio surface below", "Spreadsheet/notebook tooling — pandas for the long-form triangle, so a development pattern is data rather than a re-typed grid"],
  "prompt": "From the attached annual report / 10-K for <INSURER>, extract the claims-development (loss-development) disclosure into long-form rows: accident_year, development_year, cumulative_incurred, cumulative_paid, currency, units — one row per cell, with the page number. Do NOT complete the lower triangle, do NOT project any ultimate, and do NOT compute development factors. Separately, extract the reported prior-year development (favourable or adverse) for each of the last five years with its sign and the exact wording used, and list every place the company changed a reserving assumption, quoting the sentence. Where a cell is not disclosed, write NOT DISCLOSED.",
  "verify": "Three verifications. First, the SHAPE: cumulative incurred must be non-decreasing along a development row only if the company is not releasing; a cell that falls is favourable development and is information, not an extraction error — but a cell that falls and then rises is usually an extraction error, so re-read it. Second, TIE THE TRIANGLE TO THE P&L: the latest diagonal plus IBNR should reconcile to the balance-sheet claims reserve, and this year's incurred claims should reconcile to the change in the triangle plus payments. Third, and this is the one that matters, RE-RUN THE ANALYSIS AT MORE THAN ONE TAIL FACTOR. The chain ladder is arithmetic once the pattern is fixed; the pattern's tail is an assumption nobody discloses, and the section below shows it is worth roughly nine points of combined ratio on a plausible book. Any single-point reserve estimate — yours or the company's — is a distribution reported as a number.",
  "diy": "The gate is unaided. You compute APE from regular and single premium, VNB margin, the EV operating profit and RoEV, solvency against the 150% floor, the loss and expense ratios on the stated convention and the combined ratio on both, the underwriting result, the cost of float, the float-inclusive sustainable ROE, the shift-share decomposition of a VNB-margin move, and P/EV with its implied new-business multiple — all with a calculator. The reason is that every one of those is a place where the reported number and the economic number differ, and the only way to know which you are looking at is to have built both."
}

Modern Data Analysis

By hand first. You built Vanya's panel from disclosures and ran Ganga's shift-share bridge to −5.49 points of mix and +0.43 of rate. You computed Meghdoot's combined ratio on both conventions and attached the float. And you stripped Summit's $300m of favourable development to find that three of its four points of underwriting margin were borrowed from the past. Keep doing those by hand. Each one is a place where the reported number and the economic number diverge, and you can't spot a divergence you've never constructed.

Today's workflow. First the gate, then the triangle. Insurance is the second sector in the phase where generic machinery produces confident nonsense, and the refusals are specific:

• Refuse P/B on a life insurer. Book value ignores the in-force block, which is the whole asset. Use P/EV, and anchor it on the coherence test from the valuation section: RoEV equal to the cost of equity implies fair P/EV of about 1×.
• Refuse a combined ratio without a float attached. Float arithmetic is unambiguous here, and the quantitative lens below turns it into a number. In the Meghdoot and Kaveri pair you worked earlier, the same 102.5% is worth a 15.5% ROE on a long tail and 5.25% on a short one.
• Refuse a calendar-year CR as evidence of pricing until prior-year development is stripped. Stripping it is what moves Summit from 96.0% to 99.0%.
• Refuse to compare an Indian expense ratio with a US one without restating both to a common denominator. On identical facts, the written-versus-earned split moves Meghdoot's reported CR by 1.5 points.

Then comes the quantitative core, the one thing here that really needs code: the loss-development triangle. A triangle is a grid of cumulative incurred losses by accident year and development year. Its lower-right half is missing because it hasn't happened yet. The chain ladder fills it in. You compute volume-weighted age-to-age factors from the observed cells and apply them forward. Then you multiply by a tail factor for development beyond the oldest observed year, and read off the ultimates. Subtract what has been reported to date, and what's left is IBNR. It's twenty lines of arithmetic and one assumption, and the assumption is the whole business.

Tools & sources (IN + US). Use numpy and pandas for the triangle and the factors, and a plain loop for the ladder. An insurance-specific library is unnecessary and hides the step you must understand. matplotlib handles the development-pattern and tail-sensitivity charts. For India, start with IRDAI's mandated public disclosures on every insurer's own website. These are the L-forms for life and NL-forms for general. They're filed quarterly and include the segment-wise premium and claims forms and the claims-ageing/development tables. Add the insurer's annual report and its actuarial disclosures, plus the separate Embedded Value report (methodology, assumptions, sensitivities and the EV walk) that listed Indian life insurers publish. IRDAI's Annual Report and Handbook on Indian Insurance Statistics give you line-wise industry GDPI, penetration and the loss-ratio history behind the market dynamics above. Its monthly business figures give segment-level share. And read the Motor Vehicles Act / MACT award context, which drives third-party reserve inflation.

For the US, use SEC EDGAR 10-Ks. Their ASU 2015-09 short-duration disclosures have carried ten years of incurred and paid loss development by line since 2016, the single richest free reserving dataset anywhere. NAIC statutory filings and Schedule P give the statutory triangle and its ten-year history. A.M. Best and the NAIC have industry combined ratios. FRED (DGS5, DGS10, AAA, BAA) gives you the yield the float actually earns, and the mark it carries when rates move.

# Chain ladder + the tail factor nobody discloses. Recomputed in-session — R3.
import numpy as np
NEP, EQUITY, TAX = 7600, 5000, 0.25                    # Meghdoot's scale, §4.6-§4.7
pat = np.array([.553, .795, .909, .970, 1.000])        # % of ultimate reported, by devt year
ult = np.array([3900, 4270, 4700, 5140, 5624.])        # AY5 ties to the module's 5,624
tri = np.full((5, 5), np.nan)
for i in range(5):
    for j in range(5 - i):
        tri[i, j] = round(ult[i] * pat[j])             # the observable upper triangle

ldf = np.array([sum(tri[i, j+1] for i in range(4-j)) / sum(tri[i, j] for i in range(4-j))
                for j in range(4)])                    # volume-weighted age-to-age factors

def ultimates(tail):
    out = []
    for i in range(5):
        k = 4 - i; v = tri[i, k]
        for j in range(k, 4):
            v *= ldf[j]
        out.append(v * tail)
    return np.array(out)

base = ultimates(1.00).sum()
for tail in (0.97, 1.00, 1.03, 1.06):
    d = ultimates(tail).sum() - base
    print(f"tail {tail:.2f}: ultimates {ultimates(tail).sum():9,.0f}  "
          f"delta {d:+7,.0f}  = {d/NEP*100:+6.2f} pts of CR  "
          f"= {d*(1-TAX)/EQUITY*100:+6.2f} pp of ROE")

# The break-even combined ratio: the CR at which ROE equals r, as a function of float.
cr_star = lambda r, F, y=0.072: 1 + (y*(F+EQUITY) - r*EQUITY/(1-TAX)) / NEP
print(f"long tail  (float 12,000): CR* at r=12% is {cr_star(.12, 12000):.2%}")
print(f"short tail (float  3,000): CR* at r=12% is {cr_star(.12, 3000):.2%}")

Verify. Prove the engine on figures you've already worked first. The chain ladder recovers the constructed ultimates exactly (3,900 / 4,270 / 4,700 / 5,140 / 5,624). That's the point of building the triangle from a known pattern. If the ladder can't recover an answer you planted, it won't find one you didn't plant. Meghdoot's cascade reproduces: NWP ₹8,000 cr, NEP ₹7,600 cr, loss ratio 74.0%, expense ratio 28.5%, combined ratio 102.5% on the Indian convention and 104.0% all-earned, investment income ₹1,224 cr. The float-inclusive ROE of 13.8% and the short-tail variant of 4.1% both reproduce too. So does worked example 2's stress, on its own basis. A 6% re-estimation of the ₹5,624 cr ultimate is ₹337 cr = 4.44 points of combined ratio (the "~4 pts" quoted there), giving an underwriting loss of ₹494 cr, a cost of float of 4.12% and an ROE of 10.95%. Summit reproduces as well: CR 96.0%, underwriting profit +$400m, ROE 15.8%, cost of float −2.9%, and on the accident-year basis CR 99.0% with underwriting profit +$100m. Ganga's shift-share reproduces to two decimals, with the blended margin going from 29.77% to 24.71%, a fall of 5.06 points, split −5.49 mix and +0.43 rate.

Quantitative lens

The triangle and the float turn three of the central claims above from assertions into quantities. Every figure here comes from a triangle built at Meghdoot's own scale and from Meghdoot's own float economics. They're properties of those stated figures. They don't measure any insurer.

When a general insurer first reports a combined ratio, 44.7% of that year's loss is still an estimate. The development pattern above reports 55.3% of an accident year's ultimate loss by the end of the year itself, 79.5% a year later, then 90.9% and 97.0%, and only then 100%. So the combined ratio in the results press release for the year just ended rests on a number that's fifty-five percent observation and forty-five percent opinion. That opinion keeps being revised for four more years. It's what the reserve-development section means by "reported profit is an estimate that gets revised for a decade." Across the five accident years, the ladder puts IBNR at ₹4,124 cr against ultimates of ₹23,634 cr. Right now, 17.5% of all the losses this book has ever incurred are a number someone chose. That's the sense in which reserve honesty is the first quality screen and not the fifth.

The tail factor is the single largest undisclosed judgement in a P&C insurer's accounts, and it's worth about nine points of combined ratio. Everything in the chain ladder is arithmetic except one input: how much development happens after the oldest observed year. Move the tail factor by three points in either direction and the total ultimates move by ±₹709 cr. That's ±9.33 points of combined ratio, and after tax on ₹5,000 cr of equity it's ±10.64 percentage points of ROE. Three points is a modest range for a long-tail motor-third-party or commercial-liability book, and no filing discloses it. Now read that against the thresholds above. With the same claims and the same premiums, the same book is a 93.2% combined ratio at a 0.97 tail and a 111.8% at a 1.03 tail. A six-point range in one unpublished assumption spans the entire distance from "Progressive-grade discipline" to "chronic underwriter, avoid".

So you can see why persistent favourable development, the fingerprint from the reserve section, matters so much. You can't audit the tail directly, so you audit it retrospectively, by watching whether this company's past estimates proved high or low, year after year. For the same reason, any single-point reserve number, the company's or yours, should be reported as a range with the tail stated.

"Meaningless without the float" understates it: the exchange rate is computable, and it is steep. For a one-year underwriter, the break-even combined ratio is the CR at which the float-inclusive ROE equals the required return r, or CR* = 1 + [y(F + E) − rE/(1 − t)] / NEP. Take Meghdoot's economics at a 12% required return. A long-tail float of ₹12,000 cr supports a break-even CR of 105.58%, while the short-tail ₹3,000 cr float supports only 97.05%. That's a gap of 8.53 points of combined ratio between two insurers with identical premiums, identical equity, identical investment yields and identical required returns. All that differs is how long they hold the money. That gap is also invariant to r, because r appears in both terms and cancels. It's a handy result. Float duration buys a fixed number of combined-ratio points, whatever your discount rate.

Two derivatives fall out of the same identity, and they're worth memorising as sensitivities rather than formulas. The first is d(CR*)/dy = (F + E)/NEP, which comes to 2.24 points of combined ratio per 100 bp of investment yield on this book. So a rate cycle re-prices the underwriting standard the whole sector can get away with. Insurers written in a low-rate decade have to underwrite better than insurers written in a high-rate one to earn the same return. The second is the reserve-error amplifier. With loss reserves at k times NEP, a δ percent reserve error is k × δ points of combined ratio. At a plausible k of 1.0, a 5% reserve error is 5.0 points of CR and 5.7 percentage points of ROE. At k = 2.0, an ordinary long-tail liability book, it's 10.0 points and 11.4 points of ROE. The reserve-to-premium ratio levers the reserve estimate into the return, and long-tail lines are long-tail because that ratio is high.

Now the limits, and at length, because a chain ladder is the easiest object in insurance analysis to over-believe. The triangle above was constructed from a known development pattern and then solved. That proves the engine and proves nothing about the world. A real triangle has noisy age-to-age factors. Here we used the volume-weighted average. It's one of several defensible estimators, alongside the simple average, the latest-n average and Bornhuetter–Ferguson with an a-priori loss ratio. They'll disagree on the same data, sometimes materially. This method also assumes the development pattern is stable across accident years. That's the very assumption that breaks in the situations you most want to detect. Think of a claims-inflation regime change, a shift in case-reserving philosophy, or a change in the mix between short- and long-tail lines within a segment. Or think of a legal or tribunal shift such as the wage-linked motor-third-party awards described in the India market section.

No standard error is computed here at all. Mack's method or a bootstrap would give one, and any serious reserve range needs it. The ±3-point tail band above is an illustrative sensitivity. It isn't a confidence interval. The break-even-CR identity holds the investment yield constant and ignores the duration risk on the bond book that funds it. So a rising-rate world both raises the yield and marks the existing portfolio down. It also treats float as free and permanent, which it is not for a shrinking book. And none of this touches the two things that actually kill insurers: a catastrophe with a fat tail that no triangle contains, and a reserving philosophy that changes at the same moment the numbers do. The arithmetic here tells you how much a given assumption is worth. It can't tell you whether the assumption is honest.

Do it in code. Hold the triangle as long-form data instead of a grid. Compute factors under at least two estimators and report the spread. Make the tail factor an explicit argument with no default, so it can never be assumed silently. Report reserves as a range with the tail stated and, where you can, a bootstrap standard error. And compute the break-even combined ratio beside every reported one, because the CR alone answers no question this sector asks. See G1/QM1.01, G1/QM1.03 and G2/DA1.05.

Where this goes next: galaxy cross-links

• G1 · QM1.01 Probability, Distributions & Estimation. A reserve is an estimate with a distribution, and here that sentence becomes operational. You get estimators, bias and standard errors, and you see why reporting a point when you have a distribution is the error the whole reserve-development story is about.
• G1 · QM1.03 Time Series, Monte Carlo & ML. Bootstrapping the chain ladder for a proper reserve range, simulating catastrophe tails the triangle can't contain, and modelling claims inflation as a process rather than a constant.
• G2 · DA1.05 Statistics in Code. The triangle as tidy long-form data, development factors under competing estimators, and sensitivity surfaces you can actually read. Together they turn the reserve "honesty meter" into a chart you re-run every quarter.
• G1 · FI1.02 Duration & Convexity. The float is a bond portfolio, so the same rate move that raises the break-even combined ratio marks the assets down, and you learn to compute both halves instead of one.
• G2 · CS2.06 Software Engineering & AI-Assisted Development. The sector gate as an assertion (no P/B on a life insurer, no combined ratio without a float, no calendar-year CR without prior-year development stripped), and a tail factor that must be passed explicitly rather than defaulted.

Teach it back & journal

Feynman prompt. Explain this to a smart 15-year-old: "How can an insurance company lose a little money on every policy it sells and still be one of the best businesses in the world? And why does a life insurance company's reported profit this year tell you almost nothing about how good a year it actually had?" Keep it to one page. Use the words float, reserve, and embedded value correctly. Give at least one worked number for the float side (a combined ratio and a cost of float) and one for the life side (why growing profitably can look like losing). End with the single sentence you'd use to decide whether any insurer is honest.

Journal reflection. Write three entries. (1) Float is leverage that made Buffett, and under-reserving is what nearly killed GEICO in 1976. In your own words, write down the conditions that separate "leverage that compounds" from "leverage that kills", and where else in your investing world that distinction applies. (2) Of the four core red flags (ULIP-tilt margin fade, persistency slide, investment-gains-masked CR, solvency drift), which would you most likely have missed a month ago? And what will you now check first on any insurer? (3) Pick one listed insurer (India or US) for your Phase 8 teardown bank, and note the one question you must answer about it before anything else.

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