The clustering illusion thus stems from a form of over-generalization: We expect the correct proportion of heads and tails or hits and misses to be present not only globally in a long sequence, but also locally in each of its parts.
We over-generalise from very small samples.
