Much of the discussion here including the linked article fail to make an important distinction between domains. Prediction can be done quite effectively on thin tailed processes. A lot of the counter examples listed in the comments here are physical systems which are thin tailed. I see aircraft autopilot, collision detection, ozone depletion. These are all well understood physical phenomenon in which large deviations do not occur — your car doesn’t get teleported elsewhere in the middle of avoiding a collision. If a large deviation did occur, say a meteor striking between your car and the object it is attempting to avoid, the collision avoidance system would almost certainly fail. These events occur so infrequently that the system can just assume they won’t and boast a high success rate.

Meanwhile the examples from the linked article are fat tailed processes. Recessions, GDP, interest rates, exchange rates. These are all subject to large discontinuous jumps. Anyone doing a 5 year rate prediction in July 2019 would have been required to predict the pandemic in order to accurately forecast. This is a single example but predictions in this domain are regularly blown out by being teleported to a completely different world. Unlike the thin tailed domain these events happen frequently enough that they’re the only thing that matters for the forecast.

Knowing which class your generating process belongs to is critical to understanding whether forecasting will be effective or not. I’ll take collision detection and leave economic forecasts at the door any day.