MFFU growth study — which account type to buy for the next 18 months

1 · Headline verdict — best stable arm per campaign

Ranked on net_h12 median (every start whose 12 months close inside the window — the only like-for-like sample in the artifact). Lottery cells are excluded from ranking per the artifact's own rule: “treat any cell with funded_breach_rate >= 0.95 as a LOTTERY cell: its campaign net is a draw over funded start days, not a measurement of the arm. Rank on cells below that line, using EACH STRESS'S OWN rate — lottery status is per CELL, not per arm, and a worse book kills more accounts, so an arm can be stable at base and a lottery under slippage (and, less often, the reverse). A PAIRED base-vs-slippage delta additionally requires BOTH sides stable.” — 30 of 92 arms are excluded on the base book set, and a different 35 are excluded on the slippage book set (8 demoted, 3 admitted — see §2.5), which is why each card shows both: a same-arm re-read (the only honest paired number) and the slippage pick, re-picked from its own pool. The slippage_1tick counterpart of the same arm sits inside each card: it is not a footnote, it is half the decision.

Cross-campaign ordering — as computed by the artifact

best STABLE arm per campaign (funded_breach_rate < 0.95), h12 median net, this build's numbers. The head-to-head reads below are the artifact's own cross_campaign_ordering field, re-picked per stress (the best stable arm under slippage is not always the best stable arm on base), not arithmetic done on this page. The page independently recomputes the same picks and the build fails if they disagree.

2 · The decision reads

2.1 · 150K vs 50K within each family, and Rapid vs Pro at each size

The head-to-head table sits in the verdict section above, because pro150 vs rapid150 collapsing to a tie under slippage is a headline-level fact, not a supporting one. What follows is the arm-level detail behind it. Every comparison pairs two campaigns' best stable arms — different method arms on each side — so it is a package comparison (size, contract cap, split, price and the sweep that goes with them), never an isolated account-size effect, and every gap must clear the ±2% floor before it means anything.

Cross-campaign ordering by best stable arm is the same under base and under slippage: base ranks 1. MFFU Pro 150K · 2. MFFU Rapid 150K · 3. MFFU Rapid 50K · 4. MFFU Pro 50K; slippage_1tick ranks 1. MFFU Pro 150K · 2. MFFU Rapid 150K · 3. MFFU Rapid 50K · 4. MFFU Pro 50K. Inside MFFU Pro 50K the winning payout policy is buffer_first on base and greedy under slippage — it FLIPS. Inside MFFU Pro 150K the winning payout policy is streak_break on base and greedy under slippage — it FLIPS. The artifact is explicit that campaign-level slippage is not a scalar haircut on the base result, so read the slippage cells rather than discounting the base ones (see caveats).

2.2 · The same picks at 6, 12 and 18 months

The head-to-head above is a twelve-month read, because that is the artifact's only like-for-like sample. Here are the same two picks per campaign at all three horizons. The 6-month column is the honest short-commitment number — it is ramp-dominated (buy an eval, pass it, get funded, then earn to payout eligibility, all before any cash exists), so it is structurally far below the 12-month figure and the ratio between them carries no meaning. The 18-month column is greyed because it is not a usable statistic (n=5 — early-2025 starts only, indicative): it survives only a handful of start days and the artifact reports it for completeness only.

2.3 · The top-3 band per campaign — never a single winner

The artifact publishes a band, not a pick: THE >= 0.95 RULE IS NOT SUFFICIENT ON ITS OWN. Stable cells invert too, by ~1-2% (see cells_better_under_slippage rows with funded_breach_rate < 0.95): the same conveyor resampling that produces the large lottery swings has a small-amplitude version everywhere, because a one-day shift in a pass date resamples funded start days on ANY arm. Consequence for the report: differences smaller than a few percent between arms — under any stress — are NOT decision-grade. Show a top-3 band (see stable_top3), never a single winner separated from second place by ~2%.

2.4 · Top 5 rankable arms per campaign — base-ranked, slippage adjacent

2.5 · What the slippage arm does to the grid

85 of 92 campaign cells are WORSE under slippage and 7 are BETTER. 4 of those involve a LOTTERY cell on at least one side (funded_breach_rate >= 0.95; see lottery_side) and invert by a large margin; 3 are stable on BOTH sides and invert by only ~1-2%. Neither is a benefit of slippage: the first is a resampling draw, the second is resampling NOISE. Across the 54 cells that are stable on BOTH sides — the only cells a paired base-vs-slippage read is legitimate on — the campaign-level median slippage delta is -45.4% (p25 -57.4%, p75 -25.1%) — around twice the ~22% book-level haircut, because a worse book also breaches more funded accounts and a breached account stops earning while its replacement pays fees to re-qualify. Note that this pool is smaller than the base-stable pool: the both-sides-stable requirement is what the 8 demoted cells below fall out of.

the conveyor only ever opens ~20 funded accounts per cohort, and WHICH start days those are is set by eval pass dates. A one-day shift in a pass date swaps one funded start day for another with a materially different life. On arms where nearly every funded account dies (funded_breach_rate ~1.0) that draw dominates the ~22% book haircut. This is engine.js's documented campaign caveat: 'per CAMPAIGN no per-account ordering survives ... never assert that one bound dominates a campaign result.'

the transform is strictly worse per account: over all 390 start days of fund_rapid150K_r2400c100 (the largest campaign inversion) slip books breach on the SAME days and pay strictly less gross. The campaign inversion does not come from the account model.

7 cells score higher under slippage — 4 of them carry a lottery cell on at least one side and invert by a large margin, and 3 are stable on both sides and invert by only ~1–2%. The second group is exactly why the ±2% floor exists: THE RESAMPLING EFFECT HAS TWO AMPLITUDES, and only the large one is excluded by a breach-rate rule. On arms where nearly every funded account dies (funded_breach_rate >= 0.95) the draw dominates the book haircut and campaign net swings by multiples — those cells are LOTTERY cells and must not be ranked. But STABLE cells resample too, and invert by ~1-2%. So a NOISE FLOOR of about +/-2% applies to every arm-vs-arm comparison in this study, under every stress. Report a top-3 band (slippage_check.stable_top3), never a single winner separated from second place by ~2%.

The filter working — cells that change lottery status between the stresses

Because a worse book kills more accounts, the ≥95.0% rule does not select the same set of cells on both book sets. 8 cells are rankable on base but demoted to lottery under slippage, and 3 go the other way and are admitted to the slippage ranking. This is why the slippage-side winner on this page is re-picked from its own pool rather than being “the base winner, re-read”.

2.6 · What does not survive: winners_half

Halving every winning trade takes every single one of the 92 arms negative — zero survivors. The least-bad arm in the whole grid under that stress is rapid50 eval $1,600 · cap 50funded $400 · cap 50daily_greedy at −$6,004 h12 median against $4,266 of fees. There is no MFFU account type that is a good buy if the method's winners come in at half size; the choice on this page is conditional on the edge surviving at roughly its measured strength.

3 · Full arm tables — method arm × policy arm × stress

Every arm, every stress. Rows are ordered by base h12 median descending inside each campaign, with each arm's stress rows directly under it. Each of the m6 / m12 / m18 boxes carries its own n because they are true elapsed-month horizons computed on different cohort sets: m6 n=263, m12 n=133, m18 n=5. The m18 column is greyed on purpose — the artifact reports it for completeness only and states it is not a usable statistic; it never appears in a headline on this page. The two "% +" columns are different samples: the % h12 + column is a near-binary sign test over the like-for-like h12 set, and % full + is the same statistic over the harsher full-horizon set that includes cohorts truncated to a month or two. Dispersion lives in the p25/p75 column, not in either percentage.

READ THE COHORT COUNTS. net_at.m6 / m12 / m18 are computed on DIFFERENT cohort sets by TRUE elapsed-month eligibility: n = 263 / 133 / 5. The m18 box survives only five start days and is reported for completeness only — it is not a usable statistic. Comparing m18 against m12 across arms is the marginal-median trap; net_h12 (identical to the m12 box) is the like-for-like ranking sample.