Why every model here comes with a ledger of nulls

Sep 11, 2026

Most published portfolios are the survivors of a search nobody shows you. A few dozen variants were tried, the best one was named, and the name is what you see. That is how a backtest wins.

The models on this site were built the other way round. Before each test ran, the hypothesis, the cells, the metrics and the pass bars were written down and committed. Then the test ran once. A miss was recorded as a miss, priced, and left in the log. A near-miss was refused rather than nudged over the line. When three rounds in a row failed, a stop rule closed the program.

Every model page carries that ledger: the rounds, the variants that failed, the one that passed, and the caveats stated before the results. The point is not that the models are guaranteed to work. It is that you can see exactly how much they were allowed to fail.

Here is what that looks like for the first model published this way. Its worst decline over the test window was -30.5% against -36.0% for a plain 60/40, and its worst year, 2022, is the first thing its page tells you. Every figure in this post is read from the model’s record file; hover one to see where it comes from.