1. The event is a surprise (to the observer).
2. The event has a major effect.
3. After the first recorded instance of the event, it is rationalized by hindsight, as if it could have been expected; that is, the relevant data were available but unaccounted for in risk mitigation programs. The same is true for the personal perception by individuals.
Thanksgiving is a black swan for the turkey but not for the butcher. Taleb is trying to solve the problem of how not to be a turkey, and prediction is insufficient for that problem, because errors of prediction will let the worst events through anyway. Rather one should evaluate how much worse things will be as an event grows. If problems grow faster than linearly, trouble will strike eventually.
Taleb's model seems to miss the distinction between two properties of predictions: discrimination and calibration.
Discrimination is the correctness of the forecast of a single event. Did X happen?
Calibration is the closeness of fit between the predictions made and the distribution of outcomes. Given predictions X1, X2 ... Xn, how closely do the probabilities fit outcomes Y1, Y2 ... Yn?
Even if your calibration is very good, there are always outliers which will upset your model. You didn't discriminate them.
Comments
Taleb's definition is subjective. From wikipedia:
1. The event is a surprise (to the observer). 2. The event has a major effect. 3. After the first recorded instance of the event, it is rationalized by hindsight, as if it could have been expected; that is, the relevant data were available but unaccounted for in risk mitigation programs. The same is true for the personal perception by individuals.
Thanksgiving is a black swan for the turkey but not for the butcher. Taleb is trying to solve the problem of how not to be a turkey, and prediction is insufficient for that problem, because errors of prediction will let the worst events through anyway. Rather one should evaluate how much worse things will be as an event grows. If problems grow faster than linearly, trouble will strike eventually.
I thought I replied to this. I guess I didn't.
Taleb's model seems to miss the distinction between two properties of predictions: discrimination and calibration.
Discrimination is the correctness of the forecast of a single event. Did X happen?
Calibration is the closeness of fit between the predictions made and the distribution of outcomes. Given predictions X1, X2 ... Xn, how closely do the probabilities fit outcomes Y1, Y2 ... Yn?
Even if your calibration is very good, there are always outliers which will upset your model. You didn't discriminate them.