The Wallet That Fooled Me Longest — Technical Teardown

1 · The ruler: beat-the-line z — not dollars, and not win rate

Dollars won measure bankroll and variance as much as skill. The honest ruler is whether the account won more often than the prices it paid implied. For resolved bets at entry prices p_i (each the market's implied probability), expected wins are Σ pi, and:

z = ( actual_wins − Σ pi ) / √( Σ pi(1 − pi) )

z is significance-weighted, so a single giant lucky ticket can't move it and a wall of favorites earns no credit. Reading: z ≈ 1 is roughly what luck alone throws off; ≈ 3 is the zone you'd trust as skill; below 1 is noise. The account's lifetime 1.45 sits in the gray zone — suggestive, unproven. Everything below is the attempt to resolve which side of that line it's really on.

And win rate is not a shortcut for this. Same account, same bets, split by the price band it was buying — the raw rate and the verdict point in opposite directions:

59% on favorites is a losing record (z −0.93). 24% on longshots is also losing (z −0.63). 58% on coin flips is the account's only positive band (z 1.95). The three rates rank 59 > 58 > 24; actual quality ranks 58 > 59 ≈ 24. A win rate quoted without the price it was bought at isn't a weak signal — it's no signal, which is why every table below pairs it with the line.

2 · The paper case for a sharp

The reason I flagged it. Every one of these is real, straight off the public tape — and every one is what a genuine directional bettor is supposed to look like:

Put together: a diligent, independent, diversified human compounding a self-made bankroll on late information. That is exactly what a real edge looks like from the outside — which is why it took the next four sections to see through it.

3 · Skill decayed every month while dollars went vertical

Grade the same account month by month against the line it was paying, and it slides the entire way down — through its biggest-dollar month and out the bottom. Note the second column: it bought near-coin-flip prices every month (implied ~54% throughout), which is the only reason its raw win rate happens to track its edge here — a coincidence of this book, not a principle. The column that matters is vs line:

Only March clears the luck floor (z ≈ 1 — what luck alone throws off). After that the edge decays to +2.7 and +2.2 points, statistically nothing, and inverts to −5.8 in July. Yet June added ~$173k at a skill score of 0.57. The account wasn't getting better as it earned; it was betting more into a run that was already cooling. Dollars up, skill down: those two only diverge when someone is sizing up into variance.

Here is that divergence — the machinery of a fading coin still printing six figures:

Read the two tables against each other: the edge over the line collapses +16.1 → +2.2 → −5.8 pts while money at risk goes $108k → $975k a month. May and June — the months that made the money — beat the line by 2.7 and 2.2 points. That is the entire edge behind $341k: ~two games per hundred above what the market priced, squarely inside what luck produces. The size wasn't compounded from winnings either; the account was fed capital, deposits climbing ~$600 → $128k by June. Mechanism: a two-point maybe × a million dollars a month of turnover. The ten biggest wins (mostly five-figure World-Cup over-unders) are 27% of gross winnings. Then July ran the identical size at −5.8 pts vs the line and surrendered $128k in ten days. Symmetric up-then-down at constant stakes is the fingerprint of variance, not edge.

4 · Profit by event — the diversification was camouflage

"Six sports" implies a fluke can't hide. Split the profit by event and it hides in plain sight:

Half the profit is one World Cup — a tournament that expires, taking its markets and its edge with it. And MLB is the tell: 65 bets at z −0.69, a category it lost to the closing line and only booked a small profit on through variance. (An earlier, cruder pass had graded its baseball "+33%" and called it a strength — a reconstruction bug, corrected here.) The book wasn't diversified insurance against a fluke; it was cover for one.

And the World Cup is worth one more beat, because it's both the biggest slice and the weakest evidence. 110 bets over a one-month tournament is far too small a sample to reject luck: at z 0.70 the one-sided p-value is ≈ 0.24 — you'd see a run this good roughly one time in four by chance alone. Yet that single tournament is ~half the P&L and the reason the account cleared the leaderboard's dollar threshold at all. The signal that surfaced this wallet as a "candidate" is, statistically, the flimsiest thing in the entire book. For intuition: at ~100 hands an hour, the whole 440-bet record is a single evening of blackjack, and the World-Cup "edge" is about an hour of it — nobody grades a card-counter on one hour at the table.

5 · The on-chain origin: $607 → $24k in five days

The account settles on a public chain, so its entire funding history is readable. It resolves the whole thing in one table. The opening deposit was $606.60; total deposits across the first six days were $620.64; every other dollar that built the balance came in from the exchange — i.e. winnings:

Three views of the same opening, all consistent. The on-chain balance hit $24k on March 19. The authoritative marked-P&L feed spiked to $17,767 that same day — then surrendered nearly all of it by March 24 and sat underwater, to −$8,824, through April. And cumulative realized profit didn't cross $24k until bet #47 in late April, because many early World-Cup tickets don't settle until summer. Every clock says the same thing: a fast, lucky opening run, largely given back, with the real dollars arriving only in the May–July ramp stacked on top of it. (This is why the daily curve looks flat until May at linear scale — the early spike and the April drawdown are real but tiny next to the six-figure summer; a log axis brings them back.)

6 · Sizing: it bet like a hot hand, not an edge

Here is what a $607 stack was doing in its first days — stakes as a share of the whole bankroll:

The opening three bets each committed essentially the entire account to one game; even with a cushion it routinely staked a quarter to a half of everything on a single result. A Kelly-sane fraction on a ~55%-edge bet is a couple of percent — this is 10–50× that. Staking 20–50% of the roll per bet drives risk of ruin toward certainty over enough trials even with a real edge; with a beat-line edge indistinguishable from zero, ruin isn't a risk, it's a schedule. Stakes cluster at round dollar figures ($1,000, $2,000, $2,500), so it sized in dollars, not contracts. Combine with §3: the median stake kept rising as the skill score fell — pressing a cooling hand. It only ever looked like conviction because, for a while, it kept winning.

7 · The strip test — and the forward test that ended it

The single move that settles it. Walk cumulative realized P&L from the first bet and cut it where it crosses the first $24k — that opening run is "the streak." Then grade the two halves:

Strip the 47 bets that made the first $24k and the remaining 397 bets — the four months of careful, independent, late-news betting that so impressed me — grade z 0.58: noise. (Tape z runs a touch under the authoritative 1.45; the ratio is the point, and it's method-independent — the opening run carries the whole score.) The record isn't a sharp with a hot start. It's a hot start with a sharp's paperwork stapled to the back.

And then the forward test rendered its own verdict. The account peaked at $341,538 on July 4 and gave back 38% — about $128k — in the ten days after, winning 44% on prices implying 50% — −5.8 points vs the line, z −0.92: its first genuinely negative stretch. Its first real losing stretch re-graded my top pick, by its own results, from "confirmed sharp" to "cooling."

Prefer the story to the spreadsheet? The plain-English version is the same teardown without the tables.