Wireless indoor localization technology can utilize Wi-Fi, RFID, Bluetooth and other signals to localize the target, and thus enables many intelligent location-based applications. In the existing wireless localization systems, Wi-Fi localization system is promising due to the ubiquitous deployment of Wi-Fi devices. Especially, some existing works show that we can realize admired submeter-level localization accuracy with channel state information derived from the physical layer. However, since the performance of CSI localization system depends on many factors such as the location of the test point, the layout of the Wi-Fi devices, it is not easy to find such a suitable layout to provide satisfying localization accuracy. Moreover, as some other works show that the deployment cost of CSI localization system cannot be ignored, it is necessary to predict the performance for the upcoming Wi-Fi localization system. To this end, this paper develops a prediction model to evaluate the performance of CSI localization systems under diverse scenarios. In the model, we take the spatial attributes of the antenna and the device, and the floor plan as the main considerations. First, we define the error differential function between a pair of antennas based on the propagation model of Wi-Fi signals. Considering the overall localization area, we can generate the error differential matrix and the corresponding heat map that can be utilized to predict localization performance of one pair of antennas. Second, we design multi-antenna fusion and multi-device fusion methods to extend the error differential function, thereby constructing a general prediction model for CSI localization systems. Third, we propose to integrate the heat map and the floor plan to provide a customized prediction solution for the given scenario. In addition to the theoretical model, we conduct extensive real experiments to verify the proposed prediction model under two different scenarios. The experimental results show that the localization performance is consistent with our theoretical model, and we can utilize the model to optimize the localization accuracy by 32%-37%.