Experimental prediction 1. (A) Schematic of the task for the case of a gaussian belief. The distributions of $yt+1$ under the prior belief $π(0)$ and the current belief $π(t)$ are shown by black and red curves, respectively. Two possible observations with equal absolute prediction error $δ$ but opposite sign bias $s$ are indicated by dashed lines. The two observations are equally probable under $π(t)$ but not under $π(0)$⁠. $SBF$ is computed as the ratio between the red and black dots for a given observation, whereas $SSh$ is a function of the weighted sum of the two. This phenomenon is the basis of our experimental prediction. (B) The average surprise values $S¯Sh(θ^=1,δ,s=±1,σC=0.5)$ and $S¯BF(θ^=1,δ,s=±1,σC=0.5)$ over 20 subjects (each with 500 observations) are shown for two different learning algorithms (Nas12$*$ and pf20). The mean $S¯BF$ is higher for negative sign bias (marked in blue) than for positive sign bias (marked in orange). The opposite is observed for the mean $S¯Sh$⁠. This effect increases with increasing values of prediction error $δ$⁠. The shaded area corresponds to the standard error of the mean. The experimental task is the same as the gaussian task we used in the previous section, with $σ=0.5$ and $pc=0.1$ (see section 4 for details).