To alleviate the reliability requirement of NAND flash memory due to the increased capacity, the importance of error management has been brought up. The current error recovery flow can cause a high latency because it performs error recovery techniques sequentially regardless of the memory status. In this paper, we propose a machine learning based error recovery flow prediction method that can select the appropriate start point of error recovery which results in minimum latency with successful decoding. In addition to finding the optimal starting point that achieves minimal latency, we carefully consider input features that can be obtained during the reading process and without additional overhead. By simulation, we show that the proposed prediction method can achieve highly improved latency performance compared to the conventional scheme.