A new method for the estimation of simulated models is presented. It exploits a nonparametric sieve regression estimated through OLS to find the parameters of a simulation model producing statistics that are close to the ones obtained in real-world data. The simulation model is run for several values of the parameters, statistics are computed on each run, and the function linking the generated statistics and the associated parameters is estimated nonparametrically. Estimates of the parameters are then obtained through the previous nonparametric estimate using the real-world statistics as explanatory variables. At odds with simulated minimum-distance techniques (e.g., indirect inference and simulated method of moments), our framework does not involve any objective function, and no optimization algorithm is required. The full asymptotic theory of the estimator is explicitly and rigorously characterized, including the order of the bias, confidence intervals and hypotheses tests. The approach is evaluated through a small simulation study and the estimation of an agent-based computational model in which the evolutionary dynamics of the financial market are driven by agents with heterogeneous beliefs.