ABSTRACT In the advent of large-scale surveys, individually modelling strong gravitational lenses and their counterpart time delays in order to precisely measure H0 will become computationally expensive and highly complex. A complimentary approach is to study the cumulative distribution function (CDF) of time delays where the global population of lenses is modelled along with H0. In this paper, we use a suite of hydrodynamical simulations to estimate the CDF of time delays from doubly imaged quasars for a realistic distribution of lenses. We find that the CDFs exhibit large amounts of halo–halo variance, regulated by the density profile inner slope and the total mass within 5 kpc. With the objective of fitting to data, we compress the CDFs using principal component analysis and fit a Gaussian processes regressor consisting of three physical features: the redshift of the lens, $z$L; the power-law index of the halo, α, and the mass within 5 kpc, plus four cosmological features. Assuming a flat Universe, we fit our model to 27 doubly imaged quasars finding $H_0=71^{+2}_{-3}$ km s−1 Mpc−1, $z_{\rm L}= 0.36_{-0.09}^{+0.2}$, $\alpha =-1.8_{-0.1}^{+0.1}$, log (M(< 5 kpc$)/M_\odot)=11.1_{-0.1}^{+0.1}$, $\Omega _{\rm M} = 0.3_{-0.04}^{+0.04}$, and $\Omega _{\rm \Lambda }=0.7_{-0.04}^{+0.04}$. We compare our estimates of $z$L and log (M(< 5 kpc)/M⊙) to the data and find that within the sensitivity of the data, they are not systematically biased. We generate mock CDFs and find with that the Vera Rubin Observatory (VRO) could measure σ/H0 to ${\lt}3{\rm {per \, cent}}$, limited by the precision of the model. If we are to exploit fully VRO, we require simulations that sample a larger proportion of the lens population, with a variety of feedback models, exploring all possible systematics.