Fifteen trading ideas in, fourteen funerals

The strategy factory has now run three batches. Five ideas per batch, generated blind by an AI that never sees results before its ideas are locked. Then deterministic code tests everything and a pass bar decides.

Batch one: four died, one made it. Batch two: five for five, dead — and one of those would have passed batch one’s bar. It died purely because the bar had gone up, which is the design: the more you search, the more proof the next idea owes you. Batch three, a few days ago: five more funerals. The bar sits at 0.51 now, up from 0.42 when we started.

So the scoreboard reads fifteen ideas, fourteen dead, one survivor on probation. And I get why that looks like failure. It’s the opposite.

Most trading ideas don’t work. That’s not pessimism, it’s the base rate — and any factory claiming a better hit rate is either lying or not testing hard enough. What the graveyard buys us is real: every dead idea gets tagged with its cause of death, and the AI reads the whole graveyard before writing the next batch. “Fees eat this” gets learned once. The last batch also told us something bigger — the ideas kept dying of causes we’d already seen, which is the machine’s way of saying its vocabulary was too small. So we widened the vocabulary. That’s what the deaths are for.

One survivor, fully earned, still on probation. I’ll take it.

— Adam, Studio Amadeus