Effect of the complexity in Model 3 on mean ITD- prediction errors, averaged across participants, in HRTFs, and front (F) and rear (R) HARTFs. In comparison to i) Model 1, Model 3 was evaluated with ii) γ_2n = γ_3n = 0, iii) γ_3n = 0, and iv) all coefficients, each optimised based on feature subsets that were scaled using polynomial-regression weights (P = 4). Error bars represent standard deviations.