We use two different purely empirical approaches to estimate the non-linear transfer function between a reference rock and a sedimentary site from recordings of weak ground motions and site-condition proxies (SCPs). The modulus of the linear transfer function is first computed from weak motions and then modulated with a correction to consider non-linear soil behavior. Afterwards, a minimum-phase assumption is used to derive a complex transfer function, which then allows to recover the prediction of acceleration time series a(t) at the sedimentary site. The first approach uses the non-linear to linear site response ratio RSRNL-L proposed in Derras et al. (2020). This ratio considers both the amplitude changes and the frequency shift between non-linear (strong motion) and linear soil (weak motion) behavior. The data-driven model to predict RSRNL-L is based on various SCPs and ground Motion Intensity Measures (GMIMs). Five SCPs were considered: Vs30 (time-averaged shear-wave velocity in the top 30 m), B30 (vertical velocity gradient in the top 30 m), Vsmin (the minimum shear wave velocity), f0 (the fundamental resonance frequency, picked on H/V ratio) and A0HV (the corresponding peak amplitude), while only one GMIM loading parameter was considered (estimated PGA at KUMA site). In the second approach (Castro-Cruz et al. 2020), the non-linear transfer function is obtained from the linear transfer function (using weak motions recorded at both sites) applying a loading-dependent frequency shift (fsp). The fsp is predicted also using all weaker motion recordings available and the value of the predicted PGA at the sedimentary site. We apply these two approaches to the material provided for step 3 of the blind prediction exercise organized by the ESG6 Conference. We find that the discrepancy between the two methods was mainly concentrated on the 0.8 to 2 Hz bandwidth for energy content criteria, which is around the linear predominant frequency of the transfer function.