The turkey problem
Past data can encode maximal blindness to ruin.
The turkey is fed every day for 1,000 days. Every day, the data confirms the farmer's benevolence. The turkey's confidence in the farmer's intentions is highest right before Thanksgiving.
Taleb's metaphor for the problem of induction: long stretches of confirming data don't actually tell you much about the underlying probability distribution. The data is consistent with both "this will continue forever" and "this is about to flip catastrophically." The turkey can't tell the difference.
For operators, the warning is to be suspicious of long winning streaks. The same data that proves a strategy is robust would also be observed by a strategy about to fail catastrophically. Backtesting based purely on historical data is the turkey's confidence.
Examples in the wild
Banks that had decades of low loan losses going into 2008 thought their credit models were proven. They were turkeys.
The 'low-volatility' strategies that worked beautifully for years before 2020 were turkey strategies. The volatility regime changed and the strategies crashed.
Relationships that have looked stable for years aren't guaranteed to continue. The very stability that produces complacency is what conceals the brewing problems.
The turkey problem is one of the mental models we apply through real cases inside the Pareto MBA — a part-time program for professionals who want to think clearly about business.