The Historical model replays the 1928–2025 annual record, and how it samples those years changes what risk survives. One sampler covers it, with four presets along a single dial, ranging from fully contiguous to fully shuffled.
- Block bootstrap (default). Splices random multi-year runs of the record together. Keeps sequence-of-returns risk (bad years arriving in clumps) while generating many distinct paths, and breaks the single-period regime tailwind. Average block length is the dial: 5–10 years ≈ a business/rate cycle; toward 1 it approaches IID, toward your horizon the contiguous lens.
- Random restart. One continuous run from a random start year, wrapping at the end of the record. Real sequences, more of them than the fixed windows, but each path is still a single contiguous regime.
- Contiguous windows. Replays every overlapping full-length window. This is the classic historical lens: coarse (few paths), and measurably the optimistic outlier: every full window carries the 1980–2020 falling-rate bond tailwind.
- Non-chronological (IID). Draws each year independently. This destroys sequence-of-returns risk and flatters success rates. It is a contrast baseline, never a planning default.
Whichever preset you pick, sequence risk is a thing the run contains; it is not a thing the run reports. What the plan makes of it, age by age, is confidence by age.
The blurbs above make falsifiable claims, and the test suite pins them structurally. For example, tests verify that contiguous replay is byte-identical to the original lens and that IID kills autocorrelation. If the engine drifts from the description, a test fails.