A fair coin can't be predicted. Flip it a thousand times, and no strategy beats 50%. This is basic probability, and it's correct — for the coin in isolation.
James Stein demonstrates that the same fair coin, embedded in a random walk within “an environment of some complexity,” becomes predictable at rates exceeding 50%. The result is Blackwell's Demon: an entity that exploits inhomogeneities within statistically homogeneous systems. The coin is still fair. The flips are still independent. But the context surrounding each flip contains information that the flip alone does not.
The demon doesn't violate probability theory. It shifts the unit of analysis. Instead of predicting the next coin flip (impossible by construction), it predicts which of several possible contexts currently holds — and different contexts correlate differently with outcomes. The coin is fair globally but embedded locally. The gap between the two scales is where the demon operates.
This is the structure of every edge-based trading strategy. A prediction market is roughly efficient — prices near 50% for binary outcomes mean the crowd is uncertain. But a trader with an external signal (a weather forecast, a price feed from another exchange) can identify specific windows where the market's 50% assessment is wrong. The market is the coin. The external data is the environment. The strategy is the demon.
The through-claim the paper gestures at but doesn't state: predictability isn't a property of the process. It's a property of the process-plus-context. A fair game becomes locally unfair whenever someone has access to context that the game's participants don't share. The coin doesn't change. The embedding does.
This inverts the usual framing. Randomness is typically presented as a property intrinsic to the process — the coin is fair, the walk is random, full stop. Blackwell's result says randomness is relational. The same process is random relative to one observer (who sees only the coin) and predictable relative to another (who sees the environment). Neither observer is wrong. They're measuring different things: the coin versus the coin-in-context.