Home / Research
Research data temporarily unreachable — retrying
Live feed temporarily unreachable — retrying
Research

The research engine

Everything below exists so you never have to do any of it. Full-universe scanning, backwards scoring against resolved outcomes, out-of-sample validation — the work, shown.

Verification

Verifiable without revealing the edge

The Oracle identities are the edge, so we don't publish them. Here's how you verify us without them: every position is timestamped at entry, before resolution — the record can't be curated after the fact — and the sealed roster and full positions file are SHA-256 hash-committed below. When we later reveal, the hashes prove nothing was changed retroactively.

Roster commitment

SHA-256 of the sealed Oracle roster (full identities and statistics).

Positions commitment

SHA-256 of the full open-positions file and portfolio book (with market identities).

The gauntlet

810,747 wallets went in. 12 came out.

Every wallet on Polymarket enters the same statistical gauntlet. Almost none survive it.

wallets discoveredthe entire observable universe — everyone enters
scored against resolved outcomescut: too little history to score
deep-profiledcut: no durable expected value net of costs
validatedcut: no consistent lead on the market, or failed the untouched holdout

0.0015%survival rate — and the survivors are re-tested continuously

Statistical validation

The mathematics that survives peer review

Every claim on this site passes through the machinery below — with multiple-testing correction, chronological holdouts and cluster-adjusted inference. These are the actual formulas the engine runs.

p = 4.9×10−131
strongest BH-adjusted p-value in the roster
ρ = 0.633
Spearman rank persistence, train → holdout (p < 0.001)
d ≈ 0.29
Cohen's effect size; power analysis ⇒ n ≈ 92 clusters for 80% power at α = 0.05
Sharpe 0.54
cluster-level, day-clustered — small but real, before risk engineering
Benjamini–Hochberg FDR reject H0(i)  ⇔  p(i)im · α,  α = 0.05

Running 1,003 simultaneous lead-lag tests guarantees false positives by chance — BH controls the false-discovery rate across the whole family. Survivors carry adjusted p-values as low as 4.9×10⁻¹³¹.

Lead-lag hypothesis test H0: Pr(entry precedes move) = 0.5,  binomial(k, n)

For each wallet: of n resolved entries, how many came before the market's reprice, under minimum-move and minimum-lead thresholds (values proprietary)? Scored on a chronological train window, then re-scored on an untouched holdout.

Cluster-adjusted confidence interval CI95 = [+1.09%, +2.57%] per trade

Day-clustered bootstrap over 64 independent trade-clusters — trades on the same thesis count once, so the interval can't be inflated by correlated fills. It excludes zero.

Backwards payoff from ground truth payoff = sign(side) · (r − p) · size,  r ∈ {0, 1}

Every historical fill is joined to the market's resolved outcome r — no mark-to-market flattery, only what actually happened.

Lead-rate distribution — validated roster

Share of entries preceding the reprice, out-of-sample, per wallet.

Entry-price distribution — recent wins

Where the winning entries were priced. Left of 0.5 = mispricing calls.

Scoring & sizing

From signal to sized position

The techniques are named — that is the credibility. The calibration is the edge, and it stays private.

Expected value per dollar at fill EV/$ = w − pp  − execution-cost haircut

Win-probability w against fill price p. Every quoted edge is reduced by a measured execution-cost haircut so results are conservative by construction. The haircut's value is proprietary.

Fractional Kelly under concentration limits f* = w − p1 − p  →  scaled, then clamped

Textbook Kelly gives the growth-optimal fraction; the engine runs a proprietary fraction of it, scaled by signal conviction and market liquidity, clamped by hard per-market, per-category, per-Oracle and per-cluster concentration limits plus a cash floor. The coefficients and caps stay private.

Research

The research engine

The full methodology and wallet intelligence behind the pool — open for anyone who wants to go deep.

Research methodology

Every wallet enters a four-stage gauntlet: full-universe harvesting, backward scoring against resolved outcomes (net of an adverse-fill haircut), lead-lag testing on an out-of-sample holdout window, and Benjamini-Hochberg correction for multiple testing. Only survivors are validated.

StageCountWhat happens
DiscoveredEvery wallet surfaced from full-universe market harvesting
ScoredScored for expected value per dollar, net of adverse-fill haircut
Deep-profiledFull trade-history reconstruction and lead-lag testing
ValidatedBenjamini-Hochberg corrected and holdout-confirmed
Leaderboards & wallet intelligence

EV per dollar — top wallets

Ranked by expected value per dollar, net of adverse-fill haircut.

Category leaders

The strongest validated wallet per market category.

Smart-money pods

Wallet pairs that repeatedly enter the same markets early, together.

What is The Smart Money Pool?

The Smart Money Pool is an AI-managed prediction market portfolio built on validated Polymarket Oracles. Its research engine scans the entire wallet universe, validates the rare Oracles whose entries consistently precede the market — confirmed out-of-sample — and condenses their signals into one diversified, risk-capped portfolio. Everything on this page renders from two public data files, and every wallet can be verified independently on the Polygon blockchain. The live coverage has two layers: the wide radar (market-wide smart-money activity across the tracked candidates — not pool positions) and the validated roster whose trades the engine follows closely.

[X]ORACLE TERMINAL v1.0
You have activated retro mode. The janitor nods approvingly and keeps mopping.
Nothing about the data changed — it was already real.
Score: 10 points.