On quantification of human prediction behavior

Abstract

Autoregressive (AR) model and differential model for human prediction behavior of time series resulting from a discrete linear dynamic system were compared in terms of Akaike's information criterion (AIC). Comparison of the minimum AIC values revealed that the AR model was the better fitting model than the differential model. This result suggests that the AR information is a more practical means for the subject to predict future states than using the prior knowledge of the system dynamics, and that the subjects tend to make predictions based on only two or three preceding states of the time series. Next, the frequency characteristics of human prediction were compared with tracking behavior which involves motor constraints as well as human prediction. While amplitude ratio decreases and phase lag increases with frequency in tracking behavior, only the phase lag varies with frequency in prediction behavior. This is a characteristic feature of human prediction behavior.