SUMMARY Seismic and aseismic slip events result from episodic slips on faults and are often chaotic due to stress heterogeneity. Their predictability in nature is a widely open question. In this study, we forecast extreme events in a numerical model. The model, which consists of a single fault governed by rate-and-state friction, produces realistic sequences of slow events with a wide range of magnitudes and interevent times. The complex dynamics of this system arise from partial ruptures. As the system self-organizes, the state of the system is confined to a chaotic attractor of a relatively small dimension. We identify the instability regions within this attractor where large events initiate. These regions correspond to the particular stress distributions that are favourable for near complete ruptures of the fault. We show that large events can be forecasted in time and space based on the determination of these instability regions in a low-dimensional space and the knowledge of the current slip rate on the fault.