Discovery
The search is cheap.The judging is theproduct
A model can propose strategies faster than anyone can read them. That is only useful if something kills the bad ones without being asked nicely — so every candidate goes straight into the same backtester, and most of them do not survive the first run.
What the loop actually is
Propose, test, record, repeat
The model writes a candidate. It is run against real history with fees and slippage. The result — including the failures — is written to a runs table with what it scored and why it stopped. The next iteration reads that record, which is the only thing that makes the search a search rather than a random walk.
The first live run killed both of its own ports in minutes
Two screener strategies were ported into the loop and the backtester removed both almost immediately. That is the system working: the expensive failure mode is a search that produces plausible strategies nobody tested, and the cheap one is a strategy that dies before you have read it.
Nothing promotes itself
A candidate can reach "backtested" and no further. Paper and live are human decisions with scopes attached, and the runtime accepts a champion change only on restart — deliberately, so a strategy cannot swap itself in mid-session.
The ceiling is the data
With about two years of history, the number of independent out-of-sample windows is small, and no amount of search improves that. A discovery loop on thin data finds strategies that describe the data. The honest framing is that this generates candidates for you to judge, not conclusions.
What the loop will not do
- Promote a candidate past backtested on its own.
- Trade an account. A candidate is code and a score, not a position.
- Report a run that crashed as a strategy with no signals.
- Hide the failures. The runs table keeps what did not work, which is what stops the same idea being proposed every week.
- Claim an edge from a window it also chose. That is what the out-of-sample piece is for.
Questions about strategy search
Has it found anything that works?
It has produced candidates and killed most of them. The honest state is that dataset depth is the binding constraint, and a platform claiming a discovered money-printer on two years of crypto history would be telling you something about its marketing rather than its results.
What does a run cost?
Model calls plus compute, which is exactly why searches are one of the things the plan ladder rations. The cost is per run and visible.
Can I steer it?
Yes — the search is seeded with your constraints and the memory of previous runs. Steering it toward a market you understand is far more productive than letting it roam.
Watch a search fail honestly
The runs table keeps the failures, which is the part worth reading.