In order to monitor the physical change in real-world incrementally, minimizing energy consumption at sensor nodes to prolong the network lifetime has been the most important issue due to fact that a sensor has limited battery power. Thus, in sensor data monitoring, there has been much work for approximate data acquisition using prediction. However, traditional approximate data acquisition techniques using prediction incur frequent update of a prediction model when sensor reading changes dynamically. In this paper, we propose an efficient method for sensor data acquisition using delayed prediction in order to reduce the communication overhead with respect to the change of sensor readings. In our work, basically, a sensor reading is estimated by a prediction model. But, in contrast to traditional techniques, since a sensor compresses sensor readings for a period time when prediction was failed, our work prevents the frequent update of a prediction model.