Every signal you are testing against was written down the moment it was made, long before you ran anything. Move your thresholds around as much as you like. The thing underneath them does not move.
That is the entire difference between this and the backtest charts you have seen elsewhere, where the history and the rules were both adjustable and only one of them was mentioned.
Why nobody believes backtests
None of these require dishonesty. That is what makes them so common. A backtest handed to you by the company selling the product is a claim about the past made by the party with the most interest in that claim looking good, and here is how it usually goes wrong.
Try enough combinations of settings against the same stretch of history and one of them will look excellent by chance alone. Publish that one, mention none of the others, and you have a chart that is entirely true and completely worthless.
A test that begins in early 2020 and ends in late 2021 will make almost anything look clever. Nobody has to lie about the dates. They just have to pick them, and then not mention that they picked them.
The subtle one. A test accidentally uses information that would not have been available at the time, so it makes decisions no live system could have made. It is easy to do by mistake and impossible to spot from the outside.
The version that did well gets written up. The eleven that did not are never mentioned, because nothing forced anyone to mention them. This is the one that on-chain recording actually solves.
What is different here
The last one. Every signal the engine has ever produced was written to the chain when it was produced, with the time attached. We cannot add one now that we wish we had made, we cannot remove one that turned out badly, and we cannot nudge a confidence score to look better than it was.
So a test run here is over the complete history, including the stretches nobody would choose to show you. That does not make the other three problems disappear. You can still hunt for a flattering set of rules, and you can still pick a kind stretch of dates. What you cannot do is quietly bury the runs that went badly, and that is the one that made every other backtest in this industry unreadable.
It also means anyone can check your working. A published strategy's history is not a screenshot in a marketing page, it is a thing on a public chain that a stranger can go and read for themselves.
And because the whole record is there, you can run the same rules across one stretch after another instead of the one that suits them. A set of rules that only works in a single four month window is a set of rules that found that window, and moving between periods is the quickest way to catch that happening.
Which recorded signals your rules would have acted on, and which they would have let pass. Change a threshold and the same signals sort differently, which is the useful part: you are learning about your rules, not about the market.
rules confidence ≥ 0.70, both engines agree
period the full record, 2,318 signals
markets BTC, ETH, SOL
would have acted on 412
would have passed on 1,906
average confidence 0.79
every signal above is on-chain and timestamped
you can verify any of them without asking usThere is no profit figure in that output, and there will not be one. What you did with a signal depended on your position size, your exchange, your fees and your timing, none of which we know and none of which belong to us.