In timeseries forecasting, the phenomenon that is most difficult to identify, yet the most dangerous to handle is “self fulfilling prophecy”. Let me give you an example. Let’s say you own a coffee shop. I give you a forecast that underestimates the demand for a given day. You take that forecast and schedule your baristas accordingly. Let’s say, the demand was actually high that day. There was a long queue and the customers grew impatient and left. But the baristas were at full capacity so there were unrealized and unrecorded sells.
But if you think about your forecast’s performance: it is actually spot on. The numbers say so. Basically, your forecast influenced the underlying dynamics (small number of baristas -> low sells -> supports your underestimation) in such a way that they align with the forecast’s bias. A careless shop owner would think that the forecast is really good and keep subscribing to the forecast values.
Rob J Hyndman, a legend in timeseries forecasting in his own right, mentioned this point briefly in his talk When can we forecast and when should we give up. Essentially, he listed a set of Forecastability Factors where one of the factors were: “the forecast cannot effect the thing we are trying to forecast”. In hindsight, he might have indicated the stock price prediction problem, however, this point is strongly related to the self fulfilling prophecy (in my opinion).