What the test does and does not prove

How the test was run

• Out of sample: 22 weeks, 2026-05-10 to 2026-10-04, on trades the models had never seen: the first 20 weeks graded, then 2 newer weeks from 2026-09-27 on which nothing was chosen.
• Models retrained every 4 weeks on earlier data only.
• Positions opened before a model's training cut excluded.
• Market-making accounts excluded before the cut-offs and the books, as the live engine never scores them.
• Entry at the next minute bar after the trader's first fill.
• Broker bid/ask spread paid both ways.
• Exit with the trader or after 480 minutes.
• Results in bps per position after costs.
• The live engine's own inputs. Each position was scored by the models as tested, from the facts the live engine itself computes at the trader's first fill.
• Bars held as live holds them. The live engine keeps each market's bars where they were last set until they are refreshed; the test held them the same way, each block of weeks keeping the bars set on the week before it.

Each figure is the average result of one copied position after the broker's spread. The evidence page gives every market, with the weeks positive and how sure each figure is.

Copying only when both scores clear their bars, the test made +7.40 bps per copied position after costs, on the trades TROID can copy live: 20 of 22 weeks positive, t 5.19, over 20,888 sampled positions, 2026-05-10 to 2026-10-04.

• Chosen on the same weeks. The markets TROID trades, and copying only when both models agree, were chosen after the first 20 test weeks were seen. The models and their bars were set on earlier weeks only, but that choice was not, so the figure can overstate what live trading makes.
• Checked on newer weeks. On the 2 weeks after that (from 2026-09-27), on which nothing was chosen, the same rule made +10.39 bps per position, 2 of 2 weeks positive. That is a short check.
• Overnight financing is not charged. A broker charges a fee for holding a position overnight on some markets, and the figure leaves it out. On the newer weeks, charging it on BTCUSD and ZECUSD gives +9.05 bps per position instead of +10.39.
• A sample. The test counted the positions opened in one minute out of every 4 (1,592,942 of 6,219,853 positions), not every position.

How the bar is set

At the trader's first fill two models score the trade: the copy model, how likely the price is to move the trader's way before TROID exits (it reads 50 facts known at that moment, from the trader's own record on that market to recent price moves and volatility), and the add model, how likely the trader is to add to the position later, at a worse price (it reads the same kind of facts, known at the trader's first fill). Each market has its own bar for each score: the score only the top tenth of that market's trades reached in the latest tested week. On every market, TROID copies a trade only when both scores clear their bars (tier A, both models agree). Everything else is a pass, and TROID never takes the other side of a trader.

What is not established

• Every figure is a measurement on past trades. A result that held over these weeks has not been shown to hold over the next ones; the live record on the demo account is the only test of that.
• Below the usual bar of 2, and so possibly luck: tier A XAUUSD (t 1.93).
• Fewer weeks measured: BTCUSD (measured in 18 of the 22 weeks) and ZECUSD (measured in 10 of the 22 weeks).
• One week, the week of 2026-09-20, made 27% of everything the test made, at +28.05 bps per position; the other 21 weeks made +5.81 bps per position.
• The worst week of the whole test was the week of 2026-05-10, at -6.55 bps per position. Weeks like those can come again, and a run of them is what your hard stop is for.
• The test held one set of bars for 4 weeks at most. The live engine holds them until they are refreshed, and a longer hold has not been tested.
• The test took each position at the next minute's price after the trader's first fill. A live order can fill later, at a different price or not at all, and a broker with a wider spread takes more of any edge.