ACGLArch Capital Group Ltd.

ACGL — Arch Capital Group Ltd. — dossier, 2026-09-22

Slot names-insurance-1. Model-drafted unattended by the cloud research routine (Opus 5), following .claude/skills/stock-sourcing-pipeline/SKILL.md. Nothing here is human-verified. Sourced facts carry the document they were read from; anything from background knowledge is marked [background]; anything from a search-result summary is marked [search-summary]. Private: this directory never reaches dist/.

Access this run

§1 — The finding: the release is short-tail, the long tail has already turned

The 2026-09-22 cluster pass ended with this instruction: "Its S3 gate-minimum read should start with Arch's Q2 10-Q reserve roll-forward, because the ex-cat ex-development gap identified in §1 is the thing a name pass can either confirm or kill." It is confirmed, and the 10-Q gives a split the cluster pass did not have — prior-year development by tail length.

Note 5 of the Q2 2026 10-Q, table headed "summarizes (favorable) and adverse net PYD by segment and tail length", $ millions, negative = favourable:

Six months: 2026 short (437) / long 48 / total (389); 2025 short (355) / long 5 / total (350).

Read it in the order that matters:

1. The headline improved. Net favourable PYD went $153m → $169m in the quarter, $350m → $389m over the half. On that number alone, reserving looks fine and slightly better than a year ago.
2. The whole improvement, and more, came from short-tail lines. Short-tail releases grew from $152m to $201m (+32%) in the quarter and $355m to $437m (+23%) over the half.
3. The long tail flipped. It went from a $1m credit in Q2 2025 to a $32m charge in Q2 2026, and from $5m to $48m adverse over the half. In Arch's own words, at the same locator:

"Long-tailed lines included $ 26 million of adverse development in casualty, primarily from the 2022 and 2023 underwriting years." — 10-Q Q2 2026, Note 5, paragraph headed 2026 Second Quarter, reinsurance segment

and in insurance: "Long-tailed lines primarily included adverse development in programs business, primarily from the 2021 to 2023 accident years." (same note, same paragraph block).

That is the cycle showing up in the accounts before it shows up in the combined ratio. The releases are coming out of 2024 and 2025 property and property-catastrophe underwriting years — a short-tail book being released a year or two after it was written, which is precisely the source that a softening property market stops replenishing. The charges are coming out of 2021-2023 casualty, which is the industry's known problem and which has years left to develop.

The mortgage segment is the third leg and it is fading on its own schedule: $64m released in Q2 2025, $45m in Q2 2026. The cluster pass already flagged the mortgage loss ratio going −1.2% to +6.5%.

Why this matters more than the headline. Arch's own release puts the combined ratio excluding catastrophes and prior-year development at 82.5% against 80.9% a year earlier. That 1.6-point deterioration is the underlying business. The reported result is held up by the $169m release. If the short-tail source dries up while the long-tail charge keeps growing, the reported number moves toward the underlying one with no catastrophe required.

§2 — The second finding: a debt-funded buyback into a softening market

From the Q2 2026 10-Q, three facts that belong in one sentence:

• Senior notes $2,729m at 2025-12-31 → $4,286m at 2026-06-30 (Note 9). The cash flow statement shows $1,977m of proceeds from borrowings in H1 2026 against zero in H1 2025, less $398m of repayments.
• Buybacks of $1,949m of cash in H1 2026 (Note 3: "Arch Capital repurchased 20.7 million common shares … with an aggregate purchase price of approximately $ 1.9 billion"), against $359.7m in H1 2025 — a 5.3x step-up. Average price paid ≈ $92 against a book value per common share of $68.04, i.e. ~1.35x book.
• Equity fell. $24,206m (2025-12-31) → $24,030m (2026-06-30), despite $2,104m of net income.

The Board added $3.0bn of authorisation on 2026-04-19; $2.2bn remained at 2026-06-30. Shares outstanding 359.0m → 340.9m, −5.0% in six months.

None of this is imprudent on its own — Arch is over-capitalised by any conventional measure and the cluster pass correctly reads the capital return as a cycle call ("the disciplined actor"). But the financing of it is new. A company that genuinely believes premium rates are unattractive returns capital out of earnings; a company that borrows $1.98bn to return $1.95bn is expressing a view on its own share price as well, at 1.35-1.41x book, while its underlying combined ratio deteriorates 1.6 points and its long-tail reserves turn adverse. Those two views can both be right. They are not the same view, and the cluster file recorded only the first.

§3 — S2 kill test

(XBRL companyconcept, CIK 947484 — Revenues, NetIncomeLoss, PremiumsEarnedNet, StockholdersEquity, LiabilityForClaimsAndClaimsAdjustmentExpense, WeightedAverageNumberOfDilutedSharesOutstanding.)

Share count over five years, the stage's loudest trap check: 406.7m (2020-12-31) → 359.0m (2025-12-31) → 340.9m (2026-06-30). No dilution anywhere; the 2023 equity jump came from retained earnings, not issuance.

2026 to date, and the turn. Q2 2026 against Q2 2025, from the results 8-K:

Survives or dies: survives, easily and with room. $24.0bn of equity, $1.1bn of cash, $65m of interest paid in a half against $2.1bn of net income, no maturity wall disclosed, a $2.2bn unexercised buyback authorisation and a reserve position the auditor treats as the single critical estimate rather than a solvency question. Nothing in this pass puts the company's existence in question. What it does put in question is the level of earnings: gross premium flat, net premium falling faster than gross because Arch is buying more reinsurance or ceding more, earned premium falling faster still, the underlying combined ratio deteriorating, and the released reserves that close the gap coming out of a short-tail book that is itself shrinking.

§4 — S3 gate: the three biggest risks management itself discloses

All three are taken from Arch's own ordered risk-factor summary (10-K FY2025, Item 1A, "RISK FACTORS SUMMARY", bullets 1-4 under Risks Relating to Our Industry, Business and Operations) — i.e. management's own ranking, not this pass's.

1. Competition. "We operate in a highly competitive environment, and we may not be able to compete successfully in our industry." (Item 1A summary, bullet 1)
2. Cyclicality — placed second, above catastrophes. "The insurance and reinsurance industry is highly cyclical, and we may at times experience periods characterized by excess underwriting capacity and unfavorable premium rates." (Item 1A summary, bullet 2) This is the cluster thesis stated by the company, in the company's own priority order, and it is the reason the S0 score of 2/4 is read as "cyclical business" rather than "bad company".
3. Catastrophe volatility. "Claims for natural catastrophic events could cause large losses and substantial volatility in our results of operations and could have a material adverse effect on our financial position and results of operations." (Item 1A summary, bullet 4)

Reserving appears further down the same summary ("Underwriting risks and reserving for losses are based on probabilities and related modeling, which are subject to inherent uncertainties.") and is the auditor's sole critical audit matter: "Valuation of Reserve for Losses and Loss Adjustment Expenses", on a $33.5bn balance. PwC, PCAOB ID 238.

Red-flag sweep. Run against the EDGAR filing index for CIK 947484 back to 2010 rather than via efts.sec.gov full-text: no NT 10-K or NT 10-Q, no UPLOAD or CORRESP (no SEC comment-letter correspondence at all), no 10-K/A and no 10-Q/A. Nine 8-K/A filings, which is ordinary. This is the cleanest filing index of the three names in this slot.

§5 — S5 ownership, and a defect that invalidates a row in every 13F table in this repo

13F holders at 2026-06-30 (threespread.py holders ACGL), top five: BlackRock 30,220,703 sh / $2,933.2m (−1,133,501 QoQ); State Street 16,251,212 / $1,577.3m (+490); BAMCO Inc (Baron) 15,872,450 / $1,540.6m (−837,309); Invesco 4,947,084 (+887,704); Royal Bank of Canada 3,495,371 (−36,673).

⚠️ The quarter is incompletely ingested — 609 filers against 898 for 2026-03-31. Prior holders with no filing in the quarter include both Vanguard entities (45.2m and 14.7m shares), Artisan (22.7m), WCM (12.6m), Geode (10.6m), JPMorgan, Northern Trust, Morgan Stanley and Dimensional. No exit may be read from any of those absences. This is the fourth consecutive slot in which the warning has fired.

13D/G: all 13G, no 13D; no activist and no strategic block.

Proxy (DEF 14A, 2026-03-24, as of 2026-03-09): "All directors and executive officers (17 persons) (22)" hold 11,618,680 common shares = 3.3% of 356,272,841 outstanding; every individual officer is below 1%. Director Brian S. Posner holds 3,000 Series F and 5,000 Series G preferred shares — the only preferred held by any insider.

⚠️⚠️ The Sixth Street row is not a holding. It is Schwab's book, counted twice.

3spread reports Sixth Street Partners Management Company, L.P. as a NEW holder of exactly 2,177,713 ACGL shares ($211.4m) — byte-identical to Charles Schwab Investment Management's line in the same table. The same pair appears in this slot at KNSL (199,715 shares each, $65.9m) and AX (751,934 each, $73.2m), and appeared at QCOM in slot names-ai-capex-2b (22,310,326 each, $4,122.7m), where the run recorded it as an unexplained duplication and did not carry it as a fact. That instinct was right. Here is the cause, traced to the source:

• Sixth Street's 13F-HR for 2026-06-30, accession 0001752724-26-000051, contains an information-table file named 13F_0000884546_20260630.xml — 2,049,306 bytes. 0000884546 is Charles Schwab Investment Management's CIK. The filer agent attached Schwab's holdings table to Sixth Street's submission.
• Sixth Street's own cover page in that same accession declares <tableEntryTotal>9</tableEntryTotal> and <tableValueTotal>579182557</tableValueTotal> — 9 holdings, $579m — against a table holding thousands of rows.
• Sixth Street filed a 13F-HR/A, amendmentType RESTATEMENT, the same day (2026-08-14, accession 0001752724-26-000055) with the correct 4,662-byte table 13F_0001812095_20260630.xml. Its nine lines are: Atlanta Braves Holdings, Caris Life Sciences (×2), Essent Group, MasterBrand, PG&E, Sixth Street Specialty Lending and Sprinklr (×2). No ACGL, no KNSL, no AX, no QCOM.
• 3spread has ingested the original and not the restatement.

Consequences for this repository, stated plainly so the next run does not re-derive them: (a) every "Sixth Street Partners Management Company" row in a threespread.py holders output is Schwab's position duplicated and must be dropped; (b) the shares reported total at the top of each table is overstated by that amount; (c) the QCOM scorecard's S5 note, which called it "read as duplication, not carried as a fact", is correct and can now name the cause; (d) this is not a 3spread bug — 3spread faithfully ingested a defective filing — so the fix is to prefer the latest 13F-HR/A for a given manager-quarter, not to distrust the vendor. threespread.py gained a mechanical duplicate-row warning in this commit so the next run is told rather than having to notice.

§6 — S8 valuation and sizing (proposals only — ⏳ OPEN (user))

Price $95.78 at 2026-09-21 (ROIC NASDAQ:ACGL). Shares 340.9m → market cap $32.6bn. Common equity $23,200m ($24,030m less $830m of preferred) → BVPS $68.04, P/B 1.41. ADV $349.7m over the 21 sessions to 2026-09-21. LTM net income $4,692m (FY2025 $4,399m less H1 2025 $1,811m plus H1 2026 $2,104m) → LTM P/E ≈ 7.0x.

Out-year model. The driver is not premium volume — it is the three-part combined ratio, so the model is built on it. Assumptions named so they can be disagreed with: net earned premium −7% in 2026 and −5% a year after; ex-cat ex-PYD combined ratio worsening 1.5 points a year from 82.5%; catastrophe load 6 points; favourable PYD fading from ~4 points to ~1.5 points by 2028 as the short-tail source shrinks; buybacks continuing at ~$2.7bn a year; tax ~14%.

Note what the model says: EPS is flat for three years while the combined ratio deteriorates 8.5 points, because the buyback shrinks the share count 23% over the period. Per-share earnings hide the cycle; book value per share does not, and it compounds ~11% a year even after the buyback drag.

Against $95.78, the base case is +10% over roughly 2.3 years, before a dividend of about 0.1%. That is the sixth name in the last nine in this rotation whose correct operating forecast produces little or no return at the traded price (MRVL, ARM, ASML, CBRS, QCOM, now ACGL), with APH the exception.

Proposed buy price and size: ⏳ OPEN (user). For what it is worth to that decision: ADV of $349.7m/day imposes no liquidity constraint at any size the owner is likely to run, and the shares trade at 1.04x the modelled 2028 book value, which is the least demanding entry arithmetic of the three names in this slot.

§7 — S9 monitoring

• Next earnings: late October 2026, DERIVED from filing cadence, not scheduled. Q3 results 8-Ks were filed 2025-10-27 and 2024-10-28; the 10-Q followed on 2025-11-06. ROIC.ai's calendar needs a paid plan and Equibles — which gave QCOM a real scheduled date last slot — did not connect this session.
• Filing feed: the Tier 0 EDGAR sweep already covers CIK 947484 (watch_edgar.py), so a 10-Q, 8-K, 13D/G or Form 4 cluster promotes the name automatically.
• Leading indicator between quarters: January renewal commentary from the European reinsurers and the retrocession market. Arch's own gross-flat/net-down-10% signature is a pricing statement; if 1/1 2027 property-cat rates fall again, the short-tail release source for 2027 shrinks with them.
• The specific trigger that converts this to a buy: the long-tailed PYD column returns to zero or better for two consecutive quarters while the ex-cat ex-PYD combined ratio stops rising — i.e. the casualty years stop deteriorating before the property release runs out — with the shares still inside 1.15x book. That combination would mean the cycle risk is priced and the reserve risk has peaked. Neither condition holds today.

§8 — Evidence that is thin, named

• No earnings-call transcript at all. Equibles did not connect; fool.com has no Q2 2026 ACGL transcript. The Q1 2026 fool.com transcript exists but WebFetch returns a model summary of it, not a locatable span, so it is [search-summary] and is quoted nowhere in this dossier or in relationships.yaml. Management commentary here is entirely from the 8-K release and the filings.
• No short interest. FINRA is not reachable from this environment and Equibles, the usual route, was down. Four previous name slots had this figure; this one does not.
• No analyst-coverage count. Standing gap; the cluster pass proposed folding it into the S6 search, which was not re-run for this name.
• No shelf read. No S-3 appears in the recent filing index, but absence in an index read back to 2025-06 is not proof that no effective shelf exists; ACGL is a WKSI and may have one outside the window read.
• Proxy compensation structure not extracted — ownership table only. This is the fourth name slot to leave it, and it is now the longest-standing unclosed item in the S3 gate-minimum table.
• Segment-level 2026 premium detail not read beyond the group table in the release; the cluster file's segment figures (reinsurance NPW −10.4%, insurance −5.1%) are carried from there.

Process support, not investment advice.