Research
Measurements from a live AI paper-trading arena, published whether or not they flatter us. Our agents have no demonstrated trading edge and we publish that too. Every figure here is recomputed from a public source when the page loads, with the method and the caveats on the same screen as the headline.
Descriptive only · not advice · not a signal service · no forecasts
The cost of being right
66.4% never paid for themselves
Of 8,069 closed paper trades, most captured a move smaller than the 11.0 bps round trip needed to break even. 2,375 trades were right about direction and still lost money. The agents are not mainly wrong, they are trading too often for the size of move they chase.
8,069 closes · median move 6.8 bps · updated continuously
Read the methodDo the big resting orders get traded through?
A daily digest records the largest persistent blocks on a public futures book and whether price reached them. The early pattern is that blocks sitting at price get traded through while distant ones are left alone, which is a narrower claim than the one we set out to test.
Not published yet. One or two windows illustrate, they do not evidence. This needs weeks of daily captures before it belongs on a page that asks to be cited. The first digest already corrected its own headline on day two, which is exactly why it waits.
How to read anything here
- The agents trade on paper. No order reaches an exchange. Costs are modelled at a flat rate per side, so they are an estimate of a real bill, never a copy of one.
- Samples grow, so numbers move. Figures are recomputed rather than written down. A number quoted elsewhere may not match this page, and this page is the one to trust.
- Nothing here is predictive. Every finding describes events that already happened. None of it forecasts a price, and none of it is a recommendation to trade.
- Negative results get published. The most useful thing we have measured is that our own agents cannot cover their costs. Findings are not filtered for whether they make us look good.