Theories propose that human affective forecasting is an adaptive learning process guided by prediction errors. Although this learning process can be formally described by a Kalman filter, human forecasts are suggested to be biased and computationally suboptimal. We compared the accuracy of human affective forecasts to statistical forecasts made using a Kalman filter and explored the differences between these two processes. Participants (from the general population) repeatedly rated current levels of affect and forecasted levels of affect that they would experience 2-3 hr later (Study 1, n = 62), 1 min later (Study 2a, n = 91), and 1-2 hr later (Study 2b, n = 87), in daily life or in experimental settings. Results showed that compared to statistical forecasts, the participants' forecasts showed larger absolute errors in hour-long forecasting (dz = 0.42 and 0.30) but not in minute-long forecasting (dz = 0.17). Relative errors were also evaluated in each study, showing no differences in Studies 1 and 2b (hour-long forecasting in daily life) but more optimistic errors in participants' than statistical forecasts in Study 2a (minute-long forecasting in an experimental setting). Across the three studies, participants exhibited a strong tendency to project their current affective experience onto a new forecast, and this may explain human-specific forecasting errors. (PsycInfo Database Record (c) 2024 APA, all rights reserved).