First-order evolving fuzzy systems (EFSs) typically self-organise and self-evolve their structure and parameters from non-stationary data streams on a sample-by-sample basis to approximate the changing data patterns in real time. However, processing individual data samples sequentially is computation-ally inefficient for large-scale data streams and makes EFSs more vulnerable to temporal fluctuations in data patterns, causing their prediction performances to deteriorate. To overcome this limitation , a novel chunk-wise learning evolving autonomous fuzzy system (CLEAF) is proposed in this paper. CLEAF identifies the antecedent parameters of fuzzy IF-THEN rules as prototypes from data streams via chunk-wise online clustering, and updates the associated consequent parameters using a modified chunk-wise fuzzily weighted recursive least squares algorithm that treats every data chunk as a single unit for calculation. Thanks to the novel chunk-wise learning mechanism, CLEAF attains greater computational efficiency and achieves better approximation with less fuzzy rules. Furthermore, to better approximate nonlinear, complex problems and further enhance the prediction performance , both the input and target output variables are leveraged in data density calculation for fuzzy rule identification, enabling CLEAF to more effectively capture the underlying data patterns and their overlaps. Numerical experiments on popular benchmark datasets demonstrate the superior computational efficiency and high-level prediction accuracy of CLEAF, outperforming state-of-the-art approaches.