AFL — Aflac Incorporated

Aflac Incorporated (NYSE: AFL) — Deep-value report (REFRESH)

Date: 2026-09-03 · Price: … · Market cap: … · Shares: 501.34 mn (yfinance) / … mn at 06/30/2026 (10-Q) · Sector: Supplemental insurance (Financial Services) · Reporting currency: USD (Aflac Japan operates in JPY; consolidation and quotation are in USD — no price/reporting mismatch, but with a huge translation exposure, quantified below)

Regime: REFRESH on top of the reference thesis of 07/26/2026 (verdict then: WATCH, GBL 59.3%, price …). Trigger: 2 new filings (10-Q 08/07/2026 for Q2 2026, 8-K 08/06/2026) + one falsifier that remained unverifiable.

New sources consulted in this session (not in the old report): Q2 2026 10-Q afl-20260630.htm (accession 0001628280-26-054618), 8-K/EX-99.1 Q2 2026 earnings release, EX-99.4 CFO Max Brodén presentation (08/06/2026), DEF 14A filed 03/19/2026 (afl-20260319.htm), SEC XBRL companyfacts (AR/DAC/premium series 2022-2026), deep_data_pack.py 20260903, mc_dcf.py run 09/03/2026, web research brief 20260903.

What I carry over unchanged from the 07/26/2026 report (verified that the new filing doesn’t contradict it): the description of the business model and the Japanese distribution network (Dai-ichi Life, Japan Post, Daido Life, ~90% of banks); the 2022-2025 capital-allocation history (…/2.8/2.8/3.53 bn buybacks); the quotes from the FY2025 10-K about intensifying competition post-2001 deregulation and about digital-disruption risk; the segment structure (Japan ~53% of adjusted revenue, ~70% of segment pretax profit). What I re-derive from scratch: all Q2 2026 figures and the 06/30/2026 balance sheet, the owner-earnings bridge, the entire triangulated valuation, the Monte Carlo simulation, the Quality of Earnings chapter and the CEO chapter (both were missing as such from the old report).

Two substantive corrections to prior artifacts, both with consequences:

1. The 09/02/2026 Monte Carlo simulation was invalid. mc_dcf.py --oe requires TOTAL owner earnings, in millions; the previous run fed it 7.2, i.e. adjusted EPS per share interpreted as …lion for the whole company. The result — median intrinsic value …/share, MOS …, probability of undervaluation … — wasn’t a verdict, it was a unit error by a factor of ~500×. The correct run (09/03/2026, --oe … --sh …) gives a median intrinsic value of … and P(undervalued) …%. The … figure independently confirms the ~…/share base-case DCF from the July report, so the per-share model was correct; only its transcription into the simulation wasn’t.
2. The “company’s own 2026 adjusted EPS guidance of …-7.20” doesn’t exist. Aflac doesn’t publish EPS guidance. Company guidance is on operating ratios: Japan benefit ratio 60…, US net premium growth 3…, segment expense ratios. The …-7.20 figure in the reference report was analyst consensus, not guidance — and using it as “management itself says 2025 won’t repeat” was a stronger argument than the facts support. The argument stays valid on substance (see Red flags), but on a weaker source. yfinance today gives forward EPS consensus of …

Executive summary

The thesis, in three sentences. Aflac is a genuinely high-quality business — the undisputed leader in Japanese “third-sector” insurance, with a quasi-exclusive distribution network that’s hard to replicate, 16.6% ex-FX adjusted ROE and 43 consecutive years of dividend growth. But the core business is shrinking in volume terms: net earned premiums fell from … mn (2021) to … mn (2025) and annualized … mn in H1 2026, and EPS growth comes almost entirely from buybacks (… shares in 4 quarters) and margin expansion, not from business growth. At … the investor pays 15.5× TTM adjusted earnings against a five-year median of 12.3× — meaning 62% of the total return over the last four years came from multiple re-rating, not from earnings.

The return-decomposition arithmetic 2022→today (calculated in this session, not carried over): adjusted EPS grew from … (2022) to … (TTM Q3’25–Q2’26), i.e. …. The price grew from an annual median of … (2022) to … i.e. …. Check: 1.333 × 1.594 = 2.125 — exact. The multiple rose 9.7× → 15.5×, …. In log terms, earnings growth explains 38% of the return, multiple re-rating 62%. This isn’t a bubble accusation; it’s a finding that the dominant engine of the last four years isn’t repeatable without a further re-rating.

What the new filing (Q2 2026) moved, briefly. (a) Adjusted EPS … … reported but … ex-FX (…), with the average yen at 159.45 versus 144.60 — a 9.3% headwind; H1 2026 adjusted EPS … vs … … ex-FX. (b) US combined RBC rose to “slightly above 600%” from 570% at 12/31/2025 — the old thesis assumed RBC would fall quickly toward the 350… target and that this would slow buybacks; that falsifier broke in the positive direction. (c) Japan ESR 226% (240% with USP), down quarter-over-quarter due to subsidiary dividends. (d) The Japan benefit ratio is now guided to the upper end of the 60… band (worsening), and US net premium growth is “slightly below” the 3… band (a second consecutive downgrade: previously it was “the low end of the band”). (e) US pretax margin 20.9%, −160 bp, from higher group disability claims. (f) H1 2026 CFO … mn vs … mn, the first increase after four years of decline.

Normalized earning power remains …/share — unchanged from the July report, but now for different, better-documented reasons. The original argument (a … haircut for the Q3 2025 Japanese reserve release) is partly refuted: in Q2 2026, reserve remeasurement gains were … mn BELOW plan, and variable investment income was … mn (…/share) below long-term expectations — so the quarter under-earned, not over-earned. But new counter-arguments of similar size appear: a … mn expense-provision release in the US segment (+…/share, non-recurring), the Japan benefit ratio guided to the upper end, US growth below band, and — most important — adjusted book value per share ex-FX remeasurement fell 4.1%. The two sets cancel out; I keep … and explicitly state that I kept it, not that I didn’t check it.

Estimated value. Five triangulated models give a range of …–…/share, with a median MOS of ** …** versus the current price. Monte Carlo (20,000 scenarios on the base-case DCF) gives a median intrinsic value of … and P(undervalued) …% — but with a P10–P90 range of … to …, i.e. a width of 94 percentage points. The divergence between the two isn’t a contradiction: the base-case DCF is the only one of the five models that comes out positive, and the Monte Carlo simulates only that model, spreading its assumptions. EPV, historical multiples, the Graham Number and the bear DCF all give values below the price.

Verdict: WATCH — I don’t buy at … Versus the July report (…), the price has fallen 6.7% and the balance sheet has strengthened visibly (RBC >600%, unencumbered liquidity … versus a … minimum, adjusted leverage 21.8% in the target band). But none of the three structural problems has resolved: net premium keeps falling, 34.4% of book value is an LDTI accounting artifact tied to the discount rate, and capital returned over the last 12 months (… mn) exceeds adjusted earnings (… mn) by 27%, partly funded by net borrowing. The level at which this becomes interesting remains …–105 — the band where the price hits the EPV, the Graham Number and the historical multiple simultaneously.