Every question, answered with the actual math

The Smart Money Pool is an AI-managed prediction market portfolio built on the Oracles — the statistically validated Polymarket Oracles whose entries precede the market. It is currently pre-launch: the strategy runs live and in public on our own capital, and the pool is not yet open for deposits.

#

Like an ETF: depositors will own units of the pool rather than individual positions. Each unit is a proportional share of the whole book, so every participant holds the same diversified portfolio at the same price.

#

The planned fee is 10% of profits only, above a high-water mark — no management fee, no entry fee. If your units make no profit there is no fee, and the high-water mark means the same gain is never charged twice.

#

Not yet. The strategy, the validated Oracles and the live portfolio engine are already running in public on our own book, but the pool is not open for external capital. The waitlist gets first access when it opens.

#

An Oracle is what we call every wallet worth watching on Polymarket — the Oracles whose behaviour carries signal. The validated Oracles are the rare few whose entries statistically precede the market's price move, confirmed out-of-sample; the wider radar tracks a thousand more candidates, classified by behaviour as automated, hybrid or human.

#

Yes — without us revealing the edge. The Oracle identities are the alpha, so they're sealed behind codenames. Instead: every position is published and timestamped at entry, before resolution, so the record can't be curated after the fact; and the sealed roster plus the full positions file are SHA-256 hash-committed publicly. When we later reveal, the hashes prove nothing was changed retroactively. The aggregate data files the site renders from are free to download.

#

Mostly, no — you can buy a trading bot, which means buying software and inheriting all the work: hosting it, configuring it, picking its markets, and finding out too late whether it was any good. What this reader actually wants is exposure to the outcome. That is what The Smart Money Pool is built to be: when it opens, one deposit owns units of one risk-managed book run on statistically validated signals — no code, no infrastructure, no market selection, no monitoring. It is pre-launch; the strategy currently runs on our own capital, in public, while the record accumulates.

#

Usually not, for three structural reasons: they copy wallets that were never statistically validated (leaderboards are luck plus survivorship), they buy favourites — our backwards scoring measured entries priced 0.80–0.90 at −7.1% EV per dollar — and they act late without pricing that lag in. Fixing those three is the entire design here: validation with multiple-testing correction and holdouts, entry-price discipline published on every win, and execution costs charged against every quoted edge before it is believed. Whether that is enough is exactly what the live public book is for.

#

Demand four things and most of the market disappears: a live forward record, not backtest screenshots; entry prices on every published win, because favourite-buying at 0.95 fakes a beautiful win rate; tamper-evidence — a record timestamped at entry and cryptographically committed, so it can't be curated after the fact; and a published methodology with its limitations, because real research states where its power runs out. This site is our answer to that checklist: every number renders from downloadable aggregate data files, the record is hash-committed and timestamped at entry, and the statistics — including what isn't proven yet — are on the research page. You don't need to check any of it. But you can, and that's the point.

#

A backtest asks "what would have worked on the past?" — and with enough tries, something always did; that is curve-fitting, and it is why backtest-only products blow up. Validated means the pattern held on data it was never fitted to: our Oracles are confirmed on untouched chronological holdouts, corrected for 1,003 simultaneous tests, re-verified on complete unbiased position sets (12 of 12 survived), and then made to trade forward, in public — the only test that can't be gamed.

#

Because with 596,000+ wallets scored, extreme performers are guaranteed to exist by chance alone — so luck is our null hypothesis, not an afterthought. Every candidate faces a binomial test against chance, a Benjamini–Hochberg correction across all 1,003 simultaneous tests, and an untouched chronological holdout. Only wallets that survive all three are called validated.

#

It is the guard against finding patterns just because you looked many times. Order all p-values ascending and reject only:

With m = 1,003 wallets tested, an uncorrected 5% threshold would hand us ~50 false Oracles. BH controls the expected share of false discoveries across the whole family — the survivors are significant after accounting for how hard we searched.

#

A chronological split: the model scores each wallet on an earlier training window, then the claim must hold on a later window the model never saw. A pattern that only exists in-sample is memorisation; a pattern that survives out-of-sample is a property of the wallet. Every validated Oracle's lead rate is quoted from the holdout, not the training data.

#

After BH adjustment, the roster's p-values run from about 10⁻¹¹ down to 4.9 × 10⁻¹³¹. For scale: the weakest survivor's result would occur by chance roughly once in a hundred billion repeats of the experiment — and that is the weakest.

#

It is dropped from the roster — the validation is continuous, not a one-off ceremony. In the most recent full re-test on complete, unbiased position sets of 862–1,788 resolved trades per wallet, 12 of 12 survived.

#

Honestly: at the portfolio level, barely — and we say so. The measured effect size is Cohen's d ≈ 0.29; a power analysis says we need roughly n ≈ 92 independent trade-clusters for 80% power at α = 0.05, and the backtest has 64. That is exactly why the strategy runs live and in public: the forward record is accumulating the remaining power in real time. The per-wallet lead-lag results are a different story — with p-values to 10⁻¹³¹, power is not the issue there.

#

Three mechanical defences: chronological holdouts (never random splits, which leak the future into the past), multiple-testing correction across every comparison we run, and day-clustered inference so correlated trades can't masquerade as independent evidence. Plus the strongest defence of all — the strategy trades forward in public, where overfit edges die on contact.

#

A correlation is a pattern in the data you have; a validated edge is a pattern that predicts data you don't have yet. The bridge between them is out-of-sample confirmation: cross-wallet persistence from training to holdout runs at Spearman ρ = 0.633 (p < 0.001) — wallets that led the market in one period kept leading it in the next, which is what makes the signal tradeable rather than historical trivia.

#

Every historical fill is joined to the market's actual resolved outcome:

where r is the ground-truth resolution and p the fill price. No mark-to-market flattery, no open-position optimism — a wallet's score is what its closed decisions actually earned.

#

Because win rate ignores price, and price is the product. Expected value per dollar at the fill is:

where w is the outcome's true win probability and p the price paid. A wallet that wins 40% of the time buying at 0.25 is printing money; a wallet that wins 85% of the time buying at 0.90 is losing it. EV/$ ranks wallets by the only thing that compounds: value per dollar risked.

#

This is the most counterintuitive and most important number on this site: in our backwards scoring, entries priced 0.80–0.90 measured −7.1% EV/$ — favourites lose money at the prices people pay for them. Winning 85 times out of 100 at 0.88 collects 85 × $0.12 = $10.20 per hundred dollars of exposure, while the 15 losses cost 15 × $0.88 = $13.20. The scoreboard says you're brilliant; the ledger says you're down. It's why every win on this site is published with its entry price — and why we'd rather back an Oracle that's "wrong" 60% of the time at 0.20 than one that's "right" 85% of the time at 0.88.

#

Every quoted edge is reduced by a measured execution-cost haircut on entries and exits, because anyone acting on a signal fills later than its source, into the move it caused. The haircut prices that disadvantage into every score before it is believed — quoted edges are conservative by construction, not best-case. The haircut's value is part of the calibration and stays private.

#

Because in a binary market the entry price is the implied probability you paid, and profit is the gap between truth and price — not the frequency of being right. A win entered at 0.14 was a genuine minority call on a mispriced market; a win entered at 0.95 collected pennies for real risk. Entry price is published on every win precisely so favourite-buying cannot masquerade as skill.

#

Loss censoring is a subtle bias where winning trades survive in full history but losing trades only appear in a current snapshot — so any naive join inflates win rates. We caught it in our own pipeline during re-validation, traced it, and removed 96 affected fields from the scoring tables, then re-ran the entire validation on unbiased position sets (862–1,788 resolved trades per wallet). The roster survived 12 of 12. We publish the incident because a pipeline that can catch its own biases is worth more than one that claims it never had any.

#

That the wallet's entry timestamp systematically precedes the market's subsequent reprice. For each resolved position we ask: after this wallet entered, did the market subsequently move in the entry's direction past a minimum threshold, with the entry meaningfully preceding it? The wallet's lead rate is the fraction of entries for which the answer is yes.

#

With deliberately coarse thresholds so noise can't qualify: a reprice only counts if the move clears a minimum size, and the entry must precede it by a minimum lead — both thresholds are proprietary. Each wallet's lead count then faces a binomial test against H₀: lead fraction = 0.5, corrected by BH, confirmed on holdout. The strongest validated Oracle leads on 96.2% of holdout entries across 210 trades.

#

Long enough to act on. The edge-decay curve — how much of the eventual move remains capturable at each delay after the Oracle's entry — shows the median validated Oracle retaining most of its edge at the 6-hour mark (per-wallet retention runs from ~16% to ~97%, and is published per Oracle on the roster). Signals here decay in hours, not milliseconds.

#

Because a copier's whole disadvantage is time: you can only fill after the Oracle. If the edge died in seconds, following would be worthless. Since validated Oracles lead repricings by hours and retain measurable edge at the 6-hour mark, a follower acting well inside that window captures most of the move even after the execution-cost haircut. How quickly our engine acts is proprietary.

#

Fractional Kelly under hard caps:

Full Kelly is growth-optimal only if your edge estimate is perfect — ours never is, so the engine runs a proprietary fraction of Kelly, scales it by signal conviction and market liquidity, and clamps the result with hard concentration limits. Sizing is proportional to the book, in dollars. The coefficients are the calibration; they stay private.

#

They become one position, not several. Attribution inside that position is conviction-weighted, never stake-weighted — because stake-weighting breaks exactly when it matters: one whale's oversized fill would drag the combined position through the per-market limit on the back of one opinion. Three independent validated signals on one thesis is more evidence but still one thesis, and it is sized once.

#

Hard concentration limits, enforced mechanically on every allocation: per-market, per-category, per-Oracle and per-correlated-cluster caps, plus a cash floor. The specific values are proprietary — but their live utilisation is published on the portfolio page as a share of each cap, so you can watch the constraints bind without learning the calibration.

#

By clustering positions whose outcomes cannot be independent — same underlying event across market variants, mirrored YES/NO phrasings, and co-resolving families (the same match traded as winner, totals and spread). Clustered exposure counts against the correlated-cluster limit, so the book cannot quietly stack five versions of one opinion.

#

Positions close on resolution, on thesis invalidation, or when the engine's exit discipline triggers. The engine's judgement lives in which signals to take and how much to size them; the specific exit rules are part of the calibration and stay private.

#

Machines can't help being regular. The classifier fingerprints each active wallet's inter-arrival time distribution, multi-leg sweep structure, sizing regularity and around-the-clock activity — humans burst and sleep; automation metronomes. From the live tape, active Oracles classify as 46 automated, 141 hybrid and 165 human.

#

CV = σ/μ — the spread of a wallet's firing delay relative to its average. A low CV means the wallet fires on an essentially fixed clock. Our tightest fingerprints have coefficients of variation near zero: not a Oracle having a fast day, but a schedule — and a schedule is knowable in advance. Which Oracles run on clocks, and when they fire, is precisely the alpha, so those values stay private.

#

Because a perfect hit rate can be perfectly worthless. If a market recurs on a fast cycle and your predicted window is as wide as the cycle, you "hit" almost every time by geometry alone — high baseline, perfect hit rate, information ≈ nothing. So the guard is lift = hit − baseline, with a proprietary minimum: your window must beat what chance already gives you, by a lot. That single inequality cut the seemingly-perfect candidate list by an order of magnitude — the best proof on this site that we'd rather have a short honest list than a long impressive one.

#

A follower pair is two wallets where one consistently fires just after the other on the same markets, with the lag structure holding out-of-sample. If the relationship runs both directions, neither is following the other; they share an upstream trigger, so the pair is non-causal and excluded. What remains is a directional map: watch the source, and you are structurally ahead of the follower.

#

Six, in a pipeline. The discovery scanner enumerates every Polymarket market and harvests every wallet that ever traded. The backwards scorer joins each fill to ground-truth resolutions and prices every decision net of the haircut. The validation gauntlet runs lead-lag tests, BH correction and holdouts to name the Oracles. The intelligence engine fingerprints automation and follower structure on the live tape. The risk engine turns validated signals into one book under hard concentration limits. The portfolio engine maintains the live positions you see on this site — how it sources and times them is proprietary.

#

The scanner's live counts are on this site and climb with every sweep: at the time of writing, 810,000+ wallets discovered across 113,000+ enumerated markets, with 596,000+ scored backwards against resolved outcomes and 1,003 deep-profiled. The research page shows the live funnel.

#

Public sources end to end: Polymarket's public market data joined to on-chain resolutions. What we publish is the aggregate layer — codenames, statistics, dollar figures — because the identities and calibration are the edge. The exact JSON files the site renders from are downloadable, and the sealed layer is hash-committed so the aggregate numbers are provably honest.

#