Statistical models cannot anticipate events outside their sampled historical range

Evidence

Key here is the issue of observational experience. If you haven’t been around for a third of the life of the universe (ten sigma), then how can you know what is possible over that time period? It’s the assumed distribution that tells you what is possible, not your experience. To return to the height example, just because your model estimates that an eight-foot-tall person does not exist, it doesn’t follow that she doesn’t actually exist and that you will not run into her. In Taleb’s example, all swans were white until Europeans went to Australia and found black swans. Their exhaustive, multiyear, multisite sample of all known swans had convinced Europeans that all swans were white—until they were not. Nothing in their prior sample, no matter how complete it was, could have told them that a black swan was coming. How, then, do you hedge against risks that are not in your sample? How can you know that which is unknowable until it happens? The answer is, basically, you can’t, and if you think you can, you are setting yourself up for a fall. — Mark Blyth, Austerity, 112

Categorization

• DEFINE: Clarifies the black-swan problem: the assumed probability distribution, not experience, determines what a model treats as possible.
• XREF: Directly invokes Taleb’s The Black Swan, which Blyth builds on for the austerity critique. — connects to: Historical patterns do not determine future outcomes — Hannah Arendt, The Origins of Totalitarianism, 33; Rare violent events resist prediction by their causal structure alone; Rare events resist prediction because their causes interact uniquely (+2 more)

Broader Connections

• Historical patterns do not determine future outcomes #corroborates (0.90) — Hannah Arendt, The Origins of Totalitarianism, 33
• Rare violent events resist prediction by their causal structure alone #expands (0.85)
• Rare events resist prediction because their causes interact uniquely #expands (0.85)
• Contingent events are non-random yet depend on aberrant chance conditions #expands (0.80) — John T. Jost and Jim Sidanius, Political Psychology, 4266