Real positions, in public.

Every other figure on this site comes from a backtest we chose. This is one wallet on Polymarket, read straight from the chain — real fills, real dollars, and verifiable by anyone without trusting us. The address is under the table.

How to read the record above

A handful of positions cannot overturn a study of 2,100 windows, and are not offered as if they could. The backtest below is the stronger evidence and it argues the other way — read it before the table above persuades you of anything.

On chain
Read from one wallet address through Polymarket's public data API — the address, never a display name, which its owner can change. Verify every row yourself at the profile linked under the table.
The account, not the client
Every fill this account makes is here, placed by hand or by a client. A row is a real trade under this name; it is not a claim that the model picked it.
Real dollars
Actual fills at actual sizes and prices — not a paper record, not returns per unit. This is the operator's own money.
Open, marked to market
A position still live shows its mark-to-market P&L and is labelled open. Nothing unresolved is shown as a win.
Sold before it resolved
Exiting a position realizes its cash, so it counts in P&L like any settled market — labelled sold, never won. The win rate stays a statement about markets that answered their own question.
A small sample
A few fills cannot say whether the model has an edge. The backtest below is the stronger evidence, and it argues it does not.

We publish the backtest that argues against us.

2,100 resolved windows, second-resolution data, chronological split, every statistic clustered on the window.

The model does not beat the market price.

Model minus market log score, cluster-bootstrapped. The interval sits entirely left of zero, so this is not noise around a tie.

-0.01890
Every model, same 4,181 test rows
Market price-0.5436
Cascade (all)-0.5433
Stack (all)-0.5445
Stack (own)-0.5695

Longer is better. The best fitted model ties the price it was fitted against, and the intervals overlap on all of them.

69.27%
model accuracy, out of sample
69.31%
market accuracy — the benchmark
0.94 bps
feed basis — the noise floor these windows are decided inside
+0.135 / -0.045
P&L per trade, first half against second — the sign flips

The binding constraint

Any exchange feed differs from the settlement stream by 0.94 bps. That caps confidence at 86% one bps from the strike, and 30% of windows settle inside two. Reading Chainlink directly removes this term.

Why it still loses

The market is not slow. It converges 4% toward the model and stops — and its disagreement predicts the next 12–60 seconds of BTC (t = 3.4–5.3). It is ahead, and no model fitted to past prices recovers that.