Presented herein is an updated model for empirical prediction of 5 %-damped elastic response spectra in the period range 0–10 s, peak ground acceleration and velocity, based on a global dataset of digital acceleration records. The predictive model features saturation of the shaking parameters with both magnitude $$M_{W}$$ and distance $$R_{ RUP }$$ , magnitude-dependent distance attenuation, alternative parameterisations of the amplification effects due to local site conditions (based either on ground types or $$V_{S,30}$$ ) and corrective terms for style-of-faulting. The calibration dataset comprises more than $$1{,}880\times 2$$ orthogonal horizontal accelerometer records with $$R_{ RUP }< 150$$ km from 98 global earthquakes with $$4.5\le M_{W}\le 7.9$$ . The processing technique applied to the acceleration data optimises the reliability of the predictions at long periods, as required by displacement-based design techniques. Developed independently of the recent NGA-West2 and RESORCE-based models, the new predictive tool effectively contributes to capturing the epistemic uncertainties associated with the prediction of seismic shaking levels for engineering applications.