A residual plot indicates a good model fit when the points are randomly scattered around the horizontal zero line with no discernible pattern (no curve, no trend, no widening or narrowing spread) across the full range of predicted values, since this indicates the residuals behave like independent, homoscedastic noise consistent with the regression assumptions. A plot that trends from positive to negative as predicted values increase signals an unmodeled systematic relationship (model misspecification), and a plot whose spread visibly grows with the predicted value signals heteroscedasticity, both of which violate the assumptions underlying ordinary least squares regression. The plot lacking any such trend or funnel shape, with residuals evenly balanced around zero throughout, is therefore the one that reflects a good model fit.
Community Discussion