How the numbers are made

1.1Components and weights

Every component is clamped to 0–1, so no single input can push the score past its weight.

raw    = 30·pnl + 20·roi + 25·winRate + 10·sample + 15·recency    0–100
score  = raw × (1 − 0.3 · maxDrawdownPct / 100)                   drawdown penalty, §1.3
score  = score × (0.5 + 0.5 · sample)                             confidence gate
score  = round(clamp(score, 0, 100), 1)                           

The confidence gate reuses the sample component: a wallet keeps 53% of its score at 3 closed positions, 82% at 50 and 97.5% at 150. A three-trade lucky streak cannot read as sharp, however good the three trades were.

1.2Bayesian win rate

source apps/api/src/analytics/winRate.ts

Raw win rate is wins ÷ (wins + losses) over closed positions. A position is a win when its realised PnL is above zero and a loss when below; break-even positions are ignored. Small samples are shrunk toward 50% with a Beta(α = 10, β = 10) prior — as if every wallet started with ten wins and ten losses already on the books.

wrAdj = (wins + α) / (wins + losses + α + β) × 100    α = β = 10

The perfect three-for-three record ranks below the 70% record over 200 positions: three results are mostly prior.

1.3Drawdown penalty

source apps/api/src/analytics/drawdown.ts

Max drawdown is measured on the wallet’s cumulative-PnL curve from Polymarket’s user-pnl API, one point per UTC day for up to 730 days: the largest fall from a running peak, as a percentage of the wallet’s scale. Polymarket exposes cumulative PnL, not account equity, so the scale is the larger of the running peak, the largest absolute PnL anywhere on the curve, and $1,000. Without the floor a $5 early peak followed by a −$45 dip would read as a 1,000% drawdown; with it, 5%. Without the absolute-PnL term a wallet that peaked at +$17k, dipped to −$460k and grew to +$4.4M would read as 100%; with it, about 11%.

dd      = max over t of (peak_t − pnl_t) / max(peak_t, max|pnl|, $1,000) × 100    capped at 100
factor  = 1 − 0.3 · dd / 100                                                      

dd   0% → × 1.00      dd  25% → × 0.925      dd  50% → × 0.85      dd 100% → × 0.70

1.4Recency decay

Recency is e^−d/30, where d is days since the wallet’s last trade — a half-life of about 21 days (30 · ln 2 = 20.8). The component carries 15 points, so a wallet that stops trading gives back up to 15 raw points over a few months even if nothing else changes.

The traders-rank job recomputes every SharpScore from stored stats every 10 minutes, so decay applies without waiting for a wallet’s next full refresh.

1.5Worked example

Two wallets with the same PnL and ROI, both last trading two days ago, to isolate the effect of record size and drawdown: A won 140 of 200 closed positions with a 20% max drawdown; B won 3 of 3 with none.

Values are computed with the production function; intermediate points are rounded for display.

1.6Inputs

• PnL and volume — Polymarket’s leaderboard for that wallet (all-time), taken as reported.
• Wins, losses, sample — the wallet’s closed positions, newest first, up to 500 (10 pages of 50). See Limitations.
• Drawdown — the daily cumulative-PnL curve from the user-pnl API.
• Last trade — the wallet’s most recent trade on Polymarket, or on Sharpside’s own tape if newer.

1.7Versioning

The formula carries a version, currently v1, stored with every score and returned by the API as smartScoreVersion. Any change to a weight or curve bumps the version; the traders-rank job then re-scores every wallet from stored stats, so the leaderboard moves to the new formula in one pass. This page describes v1.

2.1Five signals

points_i  = weight_i × share_i      rounded to 0.1; 0 when not triggered
score     = min(100, Σ points_i)    rounded to 0.1

• The size signal starts at half its weight at Z = 3 and reaches full weight at Z = 6. A wallet that has traded only this market earns the full 20 for focus; five markets earn 4.
• The market-life condition keeps same-day sports and esports markets — where every position opens hours before the end — from tripping the timing signal. When a market’s start date is unknown, only the 72-hour rule applies.
• Missing inputs (age unknown, too few holders for a Z-score) leave that signal at zero points.
• Thresholds and weights are configuration (INSIDER_*, weights re-normalised to sum to 100). Each flag stores every signal’s weight as maxPoints, so a breakdown always shows the weights it was scored with.

2.2Z-score

source apps/api/src/analytics/zscore.ts

Z = (x − μ) / σ

x   = the holder's position value in USD  (shares × current price of that outcome)
μ   = mean position value over all visible holders of the market, both outcomes
σ   = population standard deviation over the same set
      undefined when fewer than 3 holders, or σ = 0

Polymarket’s holders endpoint returns at most 500 holders per outcome, so μ and σ are over the visible set. Large holders inflate σ, which makes this a conservative Z: a position has to stand well clear of other large positions, not just of the small ones.

2.3Worked example

A wallet that is 4 days old holds a position 4.5σ above the market mean, has traded 2 markets in total, entered at an average of 12¢, and opened 18 hours before the scheduled end of a market that ran for 30 days.

2.4Scanning and listing

Every 5 minutes the whales-insiders job scans up to 40 open markets: scans requested from market pages first, then markets with whale trades in the last 6 hours, then the largest markets by 24-hour volume. For each market it reads up to 500 holders per outcome, computes every holder’s Z-score, and evaluates the signals for the 15 largest holders with a position worth ≥ $10,000 (WHALE_POSITION_USD; positions under INSIDER_MIN_POSITION_USD = $5,000 or without a known entry price are never scored).

Positions are re-scored on every scan, so a listed flag’s numbers move while the position is open. A flag that falls below 50 leaves the list but keeps its record.

2.5Outcome tracking

source apps/api/src/jobs/resolutions.ts

Every 15 minutes the resolutions job re-checks markets that are past their end date or have dropped out of the synced set for 30 minutes. When Polymarket reports a market closed with a winner, every unsettled flag on it is settled: resolvedCorrect = true if the flagged outcome won, false if it lost. Until then the field is null and the flag reads open. The Insiders screen shows Correct / Wrong and can filter to settled flags, so the record of the flags themselves can be checked rather than taken on trust.

2.6What a flag is not

5.1Update cadence

“Hot” wallets are the top 500 by rank, favourited wallets and wallets that traded in the last 24 hours. The API caches public responses for 15 seconds and Pulse polls every 15 seconds. Every list carries asOf, the last successful run of the job behind it; the status strip switches from LIVE to DELAYED when data is more than 2 minutes old.

5.2Warm-up after deploy

On a fresh deploy the worker runs market sync, trader seed, tape ingest and aggregation once, in that order, then backfills prices if fewer than 5,000 price ticks exist. Because windows are built from Sharpside’s own bars, the 30m and 1h windows are complete after an hour, while 6h, 12h and 24h volume and trader counts fill in over the first day. The first full SharpScore pass over ~3,000 seeded wallets takes 2–3 hours; until wallets are scored, Smart Money Flow and sharp counts read near zero.

5.3Stored trades

The tape runs at roughly 1,500–2,000 trades a minute. Every trade feeds the bars, but only trades ≥ STORE_TRADE_MIN_USD = $100 are stored individually, for 14 days — about 7% of trades and 95% of volume. The Activity feed therefore starts at $100.