A numerical modeling study was performed to investigate fluid recovery following hydraulic stimulation in low matrix permeability formations. A simulator was used, CFRAC, that implicitly couples fluid flow with the stresses induced by fracture deformation in two-dimensional discrete fracture networks. An unstructured mesh was created around the fractures to simulate leakoff and flow in the matrix. Four simulations were performed in which fluid was injected, the wells were shut in, and then fluid was produced back to the surface. The baseline simulation contained a single, linear fracture propagating away from the wellbore. The other three simulations used a stochastically generated network of natural fractures and assumed that as hydraulic fractures propagated through the formation, they terminated when they intersected natural fractures. The termination process created complex, branching fracture networks. The simulations showed that fracture network complexity reduced fluid recovery because the natural fractures, which were not perpendicular to the minimum principal stress, closed at an elevated fluid pressure and created barriers for flow between the wellbore and the open, fluid-filled fractures away from the well. However, if the transmissivity of closed fractures was too low, the fracture network was inhibited from becoming complex, and fluid trapping was not as severe. In the two complex fracture network simulations with lower closed fracture transmissivity, the shut-in pressure transient showed abrupt changes in slope, which were caused by episodic growth of the fracture network due to leakoff of fluid into natural fractures, rapid propagation of opening along the natural fractures, and subsequent initiation of new hydrualic fractures. In practical applications, the observation of abrupt changes in slope during shut-in could be taken as evidence that episodic fracture propagation is occurring, which would imply a complex and branching fracture network. In one of the three complex fracture network simulations, it was assumed that closed fractures had relatively high transmissivity, and abrupt changes in the slope of the shut-in transient were not present, even though a complex fracture network developed during stimulation.