Fade The Crowd
Archived · data through 2026-08-12

Research archive — Fade is no longer running live.

Live data ingestion, market resolution, signal computation and all trading bots were stopped on 2026-08-12. Everything on this site is historical data through 2026-08-12, preserved for research. Nothing here is a current signal, and nothing here is trading advice.

The contrarian “fade the crowd” thesis did not survive testing — read what we found and why we stopped.

Why Fade stopped

FadeTheCrowd was built on a simple idea: identify consistently losing wallets on Polymarket and take the other side. Between June and August 2026 we tested that idea properly, against a preserved dataset of every trade the system had ever recorded. It did not survive. Live operation stopped on 2026-08-12; the data below runs through 2026-08-12.

This page is kept public because a negative result that nobody publishes gets rediscovered by the next person at their own expense. None of it is trading advice.

What was tested against

Trades54,392,5432026-06-06 → 2026-08-10, validated row-for-row
Resolved positions51,912,474outcome labels with payout and cost
Markets1,800,929100% coverage of the trade corpus

What we found

Finding 1

The original signal measured its own filter

Defect

The live signal labelled a wallet "bad" using 24-hour net cashflow (sells minus buys), then measured whether those wallets were buying. Because accumulating a position IS negative cashflow, the filter selected wallets for having bought and then observed them buying. Resolution payouts are not trades, so a wallet that bought, won, and held to settlement stayed "bad" forever. The result was 33,045 BUY signals against 55 SELL — 99.8% one-sided.

Finding 2

The direction axis itself was broken

Root cause

Selecting no wallets at all — the entire population — also produced 99.35% BUY. The saturation was never about wallet selection: 83.3% of all trades in the feed are "buys", because in a prediction market a buy opens a position and a sell is an early exit most participants never take. 80.7% of wallet-position sequences never sell at all. Buy/sell is a position-opening indicator, not sentiment. The meaningful axis is which outcome token was bought (Up vs Down, Yes vs No), which is near 50/50.

Finding 3

Wallet quality carries no directional information

No support

Rebuilt properly — point-in-time wallet quality from 51.9M resolved outcomes, anchored so no result could leak backwards in time — the worst-decile and best-decile wallets showed a large apparent difference in aggregate. Compared inside the same market at the same moment, that difference collapsed to zero (+0.0008 ± 0.023). Seven pre-registered robustness variants agreed, with the sign flipping between them. Bad wallets are not systematically on the wrong side of a given market.

Finding 4

Order flow adds nothing to price

No support

Flow direction predicts outcomes on its own (AUC 0.83) but only because it echoes the price. Adding flow to a model that already knows the market price made out-of-sample predictions measurably WORSE in all four market families, with confidence intervals excluding zero. The market price is a well-calibrated probability (AUC 0.965); contemporaneous flow is a lagged restatement of it.

Finding 5

One real anomaly — too small to trade

Statistical only

In 5-minute crypto markets, favourites are genuinely underpriced by roughly half a cent to a cent near expiry. The effect is stable across every late week and all seven crypto assets, and is not a stale-price artifact. But it is smaller than the market's own 1¢ tick, it weakens or reverses when restricted to economically meaningful trade sizes, and the break-even friction (0.53–1.07¢) sits inside the plausible execution cost (0.50–1.71¢ one way). No market condition concentrates it.

How it was tested

  • Every hypothesis was pre-registered in code before results were computed, with the falsification criterion written down first.
  • Wallet quality used only outcomes from markets that had already ended at least a day before the observation, so no future information could leak backwards.
  • Evaluation was chronological — fit on June to mid-July, evaluated on mid-July to August. Never a random split.
  • Uncertainty was clustered by market, because overlapping observations in the same market are not independent. Correcting this changed conclusions.
  • Results are reported against the contemporaneous market baseline, not against a 50/50 null — the population itself is not balanced.

What is still here

The historical leaderboard, wallet pages and trade feed remain browsable as a static record of what the system observed. They are not updating. The signal strip, alerts and paid API have been switched off.