We describe a model for estimating past and current infections as well as future deaths due to the ongoing COVID-19 pandemic. The model does not use confirmed case numbers and is based instead on recorded numbers of deaths and on the age-specific population distribution. A regularized deconvolution technique is used to infer past infections from recorded deaths. Forecasting is based on a compartmental SIR-type model, combined with a probability distribution for the time from infection to death. The effect of non-pharmaceutical interventions (NPIs) is modelled empirically, based on recent trends in the death rate. The model can also be used to study counterfactual scenarios based on hypothetical NPI policies.