Abstract. This work presents a data-driven nonintrusive model reduction approach for largescale time-dependent systems with linear state dependence. Traditionally, model reduction is performed in an intrusive projection-based framework, where the operators of the full model are required either explicitly in an assembled form or implicitly through a routine that returns the action of the operators on a vector. Our nonintrusive approach constructs reduced models directly from trajectories of the inputs and outputs of the full model, without requiring the full-model operators. These trajectories are generated by running a simulation of the full model; our method then infers frequency-response data from these simulated time-domain trajectories and uses the data-driven Loewner framework to derive a reduced model. Only a single time-domain simulation is required to derive a reduced model with the new data-driven nonintrusive approach. We demonstrate our model reduction method on several benchmark examples and a finite element model of a cantilever beam; our approach recovers the classical Loewner reduced models and, for these problems, yields high-quality reduced models despite treating the full model as a black box.Key words. data-driven model reduction, nonintrusive model reduction, projection-based reduced models, Loewner framework, black-box models, dynamical systems, partial differential equations AMS subject classifications. 65M22, 65N22 DOI. 10.1137/16M10947501. Introduction. Projection-based model reduction derives low-cost reduced models with low-dimensional reduced states that approximate the high-dimensional solutions of a large-scale system of equations [2,10,47]. Approximating full-model solutions with reduced solutions can reduce the runtime by orders of magnitude; however, the applicability and scope of model reduction is often limited because of the intrusive nature of reduction algorithms. Deriving a reduced model with, e.g., proper orthogonal decomposition [11,49], balanced truncation [35,36], the reduced basis method [15,19,21,47], and projection-based interpolatory model reduction [2,3] is intrusive in the sense that the operators of the full model are required either in an assembled form or through a routine that provides the action of the operators on a given vector. In many situations, however, the full model is given as a black box that computes solutions of the full model without providing the full-model operators. We introduce here a data-driven nonintrusive model reduction approach that constructs a reduced model from the solutions of the full model alone, without requiring the full-model operators.We consider here time-dependent full models with linear time-invariant (LTI) operators. In our setting, the full models map an input onto an output (quantity of