Quote · Economist Podcasts
Cockroaches 1, Modi 0: India’s remarkable protests
Where this was said
AI vs Humans vs Experts: Who Can Actually Trade the News?
At 17:59 · chapter starts 15:35
The Elm Wealth experiment didn't stop at human participants. The researchers also ran AI models through the same crystal-ball test and found the results were both reassuring and sobering. The machines correctly predicted market direction about 60% of the time — better than the 51% human average — but they made the same fundamental error: over-betting when uncertain and under-betting when confident. [1] — Joshua Roberts "AI models: 60% correct market direction: AI models tested in the same Elm Wealth experiment correctly predicted market direction 60% of the…" 17:00 Rosie Blau admits to finding it 'quite comforting' that the machines aren't brilliant at this. The real revelation comes when the five expert macro traders are introduced. Unlike the AI and lay participants, all five finished in profit, and on average they more than doubled their stake. [2] — Joshua Roberts "When Elm Wealth gave the same test to 5 expert macro traders, all five made money and on average more than doubled their stake. Their edge …" 17:37 Their edge wasn't dramatically better prediction — it was dramatically better bet sizing. On low-conviction days they chose not to trade at all. On high-conviction days they ramped up leverage aggressively. Roberts draws a broader lesson: the investing world obsesses over what to buy, but the more important question is how much to bet, calibrated to actual confidence. This, he argues, is possibly the most important question in investment, and the one we give the least thought to.
Markets don't move on events — they move on surprises versus expectations. If a great jobs report was already priced in, knowing it in advance is worthless. Participants in the Elm Wealth study called the market direction correctly only 51% of the time, barely above a coin flip.
AI models tested in the Elm Wealth experiment predicted market direction 60% of the time — better than humans — but made the same catastrophic mistake: they didn't calibrate bet sizes to their confidence. Even our supposed future overlords took too much risk when uncertain.
AI models tested in the same Elm Wealth experiment correctly predicted market direction 60% of the time, slightly better than humans, but still struggled with bet sizing.
All 5 expert macro traders recruited by Elm Wealth for the same experiment finished with profits and on average more than doubled their money.
When Elm Wealth gave the same test to 5 expert macro traders, all five made money and on average more than doubled their stake. Their edge wasn't better prediction — it was disciplined bet sizing. On uncertain days they didn't trade at all. On high-conviction days they went all in.
Investors obsess over which stock or bond to buy. But the Elm Wealth study reveals the more important question is how much: calibrating bet size to confidence level. Getting this wrong is what caused ordinary participants — and AI models — to go bust even with tomorrow's headlines.
