Lesson #26A smoother dividend yield is not necessarily a more accurate one
Classify the cash before averaging it. Separate ordinary income from special, variable and non-cash distributions, then check the payment trend, currency and share basis.
Inspect evidence and consequence
What we observed
Why the first approach looked sensible. We initially used several complete years of dividend history to stop an exceptional payment from dominating a forward ranking. Extending that idea to a three-to-five-year average and removing unusually large observations sounds robust. But an average can dilute a special payment without actually identifying it. An outlier filter can also remove a genuine ordinary increase, while an old average can overstate income after a dividend cut. Averaging historical yields introduces another problem: their share-price denominators differ from the price an investor pays today.
Novo Nordisk shows the difference. Its annual reports list dividends of DKK 5.20, 6.20, 9.40, 11.40 and 11.70 per share for 2021-2025. The five-year mean is DKK 8.78, below the latest fiscal-year amount. At an explicitly hypothetical price of DKK 300, those cash amounts imply gross yields of 2.93% and 3.90%. Neither number is a live quote or a forecast; the example shows how smoothing a rising payout can lag. Novo also pays unequal interim and final instalments, so the smaller interim payment must not be treated as the entire year or the larger final payment rejected merely for being larger.
Our own Novo audit was not just an averaging problem. On October 16, 2026, the old published 0.90% net assumption differed from the corrected NOV.XETRA snapshot: 4.49% gross regular trailing yield and 3.28% after the model's 27% withholding assumption. Investigation also exposed period, duplicate-event and currency-basis differences. A nicer average cannot repair those inputs. These are dated model observations, not today's dividend yield or an investor's personal after-tax return.
Special, variable and non-cash payments answer different questions. Costco's December 2023 announcement identified a USD 15 special cash dividend. That cash belongs in actual payment history, but must not be repeated automatically in a future-income assumption. A spin-off distributing shares is not recurring cash; returns of capital and fund distributions also require their own treatment. Progressive's declared USD 13.50 annual payment alongside a USD 0.10 quarterly payment illustrates why two components can share a payment date without being duplicate events. CME explicitly separates quarterly dividends from its annual variable dividend: a distribution can recur without its amount being dependable ordinary income.
A special label is evidence, not a universal economic definition. Companies can use that label for distributions that recur, and data providers can leave types missing or classify them inconsistently. FTSE Russell even documents a context-specific rule for treating some repeatedly paid special cash distributions as ordinary; that is an index rule, not our automatic override. We do not relabel every repeated special as ordinary. Our current pipeline respects explicit special or variable labels unless an exact, dated issuer-backed payment review establishes a different ordinary component. Raw labels and amounts stay available. A questionable label can therefore make the recurring estimate conservative until a separate audit corrects it; it never pauses an asset or the daily ranking workflow.
What we do now. The headline regular-income estimate uses eligible ordinary cash over the trailing 12 months divided by one dated current price, with an annualized 24-month cash comparator using that same price. Both use consistent currency, split and listing-share units: GBP cash cannot be divided directly by a GBX price, and an ADR cannot be assumed to represent one underlying share. Payment identity is reconciled before summing; legitimate ordinary and special components on the same date must survive. Model withholding has its own provenance. Special and variable cash remains in actual history, rather than silently disappearing because it is excluded from recurring projections.
Trend and seasonality matter. For missing or ambiguous types, our automated rule compares earlier payments of the same cadence or season. It can estimate an ordinary component and a potentially special excess, with the method recorded. Similar repeated higher payments can establish a new current level instead of being capped indefinitely. Interim and final payments are compared with appropriate peers. The 12-month versus 24-month comparison helps reveal rising or declining payouts; it does not justify extrapolating growth forever. A cut, suspension or changed schedule can still make trailing income stale and needs dated issuer evidence. Sparse history uses a labelled ordinary estimate so publication continues, with periodic audits prioritizing material uncertainty.
The AI lesson. Shared bad inputs do not become reliable because several advisors agree with them. Payment normalization, classification and return arithmetic need deterministic rules, dated evidence and inspectable receipts. Current rankings and dividend-inclusive illustrations can then use corrected inputs while frozen forecasts keep their original assumptions. A constant-yield compounded illustration is a model convention, not a promise about future distributions.
What changed
Use payment-level recurring-income classification and dated current-price yields instead of treating a three-to-five-year average as a substitute for understanding the cash. Keep ordinary, special, variable and estimated components distinguishable; automate ambiguous cases without asset holds; audit source labels and material trend changes separately. Preserve original and sealed historical records.
Evidence scope
Lesson origin: Engine workflow
October 16, 2026 Batch 7 dividend-policy audit and full-cohort current-publication replay, plus issuer dividend disclosures and FTSE Russell's published classification example. The arithmetic example is illustrative; it is not a controlled comparison proving one forecasting policy universally optimal.
Supporting references
- Novo Nordisk: five-year dividend and share-price history (2025 annual report)
- Novo Nordisk: dividend policy and interim/final payments
- Costco: USD 15 special cash dividend announcement (December 2023)
- Progressive: annual and quarterly payments announced together (2025)
- CME Group: quarterly and annual variable dividend structure
- FTSE Russell: ordinary dividends, special dividends and capital repayments
Also relevant to: Technical Workflow & Processes; Asset & Market Insights. Search by these category names to find cross-topic lessons.
Limits
- — Trailing yield is backward-looking. Announced cuts, suspensions, seasonality, schedule shifts and incomplete coverage can make it a poor forward estimate.
- — Automatic anomaly and repeated-level rules are operating heuristics, not a guarantee or an industry-wide standard. Unknown types can overstate income; explicit special labels can understate recurring cash until issuer review.
- — Recurring variable distributions may be economically important even when excluded from the regular-income projection. Actual cash history must remain complete.
- — FX timing, withholding, fees, ADR ratios, splits and non-cash corporate actions affect comparability. A model net yield is not a personalized tax calculation.
Adoption and evidence
Learned during the October 2026 review of Batch 7 current rankings. The recurring-income policy replaced the earlier multi-year averaging assumption in lesson #14; original forecasts and historical ranking snapshots were preserved.
Incorporated since
Learned during
Batch 7
September 2026
The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.