Put five analysts in a room, ask each for a view, and all five say buy. It feels like strong evidence. It is only as strong as their independence, and analysts who read the same news, use the same models and talk to each other are not independent.
Statistics has a clean way to say how much a correlated panel is worth. If N analysts make errors with an average correlation of ρ, the panel is worth
N / (1 + (N − 1) ρ)
independent opinions. With five analysts and a correlation of 0.6, that is 5 / 3.4, or 1.47. Five voices, one and a half opinions.
Adding more voices barely helps
Move the first slider. At a correlation of 0.6, twenty analysts are worth 1.6 opinions. The ceiling is 1 / ρ, about 1.67, however many you add. Past a handful, a bigger panel buys almost nothing.
Why this matters more for AI
A panel of AI agents is often the same model with different instructions. Their errors can be far more correlated than any human team’s. Without a discount, a system that sizes positions by how strongly its agents agree will bet biggest exactly when its agents are most alike, which is when agreement means least.
The fix is simple to state: measure how correlated the agents’ errors really are, and let agreement count only for the independent opinions it represents.
What to ask of any AI research process
How many independent views does your panel actually contain? If nobody has measured it, the honest answer is “fewer than it looks”.
The arithmetic is textbook. Numbers here are illustrative. More ideas like this are on the maths page.