This paper proposes a new identification method for continuous-time SISO linear time-invariant models, which does not require the prior knowledge about the system order. First, an appropriately filtered input/output signal is projected onto a finite dimensional signal subspace. Then, based on the projected data, the system order is determined through a nuclear norm minimization which takes account of both model simplicity and output prediction accuracy. Numerical examples are given to demonstrate the effectiveness of the proposed method.