Findings · 2026-11-12
Our best result passed its test. Then we tried to break it.
Out of 100 trading rules we tested this month, one passed. This is what happened when we tested it again somewhere it had never been seen.
The claim. “A strategy that passes a strict backtest is ready to trade.”
What we found. It passed on one exchange and one year. On other exchanges and other years it was nothing.
The idea
Exchanges run programs that hand new tokens to people for free, for holding or staking something else. When such a token starts trading, many of those holders sell whatever the price. The rule: bet against the token when it lists, hold Bitcoin against it so that a falling market is not what you are measuring, and close after 14 days.
It passed
We locked the rule and the bar before computing a single return. On Binance program tokens since 2025, 59 listings, the rule returned +21% on average and cleared the bar comfortably. It was the only rule of 100 to do so.
Most people would stop here and start trading it. We wrote down what worried us instead: almost all of the result came from one year, 2025, when small coins fell against Bitcoin nearly all year. Before 2025 the same rule was much weaker, inside what luck alone produces.
The second test
If selling by free-token holders is the real cause, the trade should work on other exchanges that run the same kind of program. We locked a second test on Bybit and OKX program tokens, on listings the first test had never seen.
It failed. Across 23 listings the typical trade still won, but three trades lost between 137% and 180% each when the token kept climbing, and those three erased the rest. Before 2025 the average was -0.8%. From 2025 on it was strongly positive again.
So the pattern is tied to a period, not to a cause. In 2025 betting against almost any small coin worked. The rule had found the year, not an edge.
What we did
We did not add a stop-loss or a filter to rescue it. Changing a rule after seeing where it lost is how a test turns into a story. The rule now runs as a forward trial with its rule frozen, and the result will be published on the Proving Ground whichever way it goes.
This is the same bar we hold signal channels to. A good month, or a good year, is not evidence. The question is whether it survives somewhere it has not been seen.
How it was tested
- First test: 95 Binance program tokens, 2023 to 2026, rule locked 11 November 2026; passing period 2025 onward.
- Second test: Bybit Launchpool and OKX Jumpstart tokens not in the first set, rule locked 12 November 2026 before any price was fetched.
- Costs: 0.214% for both legs plus real funding. Uncertainty measured across weeks.
Research on past data, not financial advice. Every test we run, failures included, is on the Proving Ground.