A prediction device includes: a normal prediction unit configured to predict first prediction values for target ranges of respective models, based on respective pieces of spatio-temporal observation data including observation values for respective observation locations and respective observation times observed for a given time period and the respective models for predicting variations of the pieces of spatio-temporal observation data for the respective target ranges defined for at least one of an observation location or a period; a residual extraction unit configured to extract, based on residuals between the respective pieces of spatio-temporal observation data being observed and the models corresponding to the respective pieces of spatio-temporal observation data based on the pieces of spatio-temporal observation data and the models, the respective pieces of spatio-temporal observation data with residuals greater than or equal to a threshold value, as spatio-temporal residual data; a residual prediction unit configured to predict, based on the spatio-temporal residual data, second prediction values for the target ranges corresponding to the spatio-temporal residual data in accordance with a predetermined prediction technique; and a synthesis unit configured to output prediction values obtained by synthesizing the first prediction values and the second prediction values.