The applications of spectrum prediction span over a wide range of crucial fields in wireless networks, such as spectrum efficiency improvement, service quality enhancement, and network management. Despite such broad ranges of fundamental applications of spectrum prediction, existing methods are based on coarse measurement of power spectral density values. Few predictions target the actual binary tenancy of whether the spectrum slices are occupied or left unused, but their data resolution and prediction accuracy are far from satisfactory. To improve the accuracy of spectrum prediction, we propose the framework of Total Occupancy guided Prediction (TOP). It is a general prediction scheme that is flexible to incorporate an arbitrary algorithm into its framework with enhanced accuracy. Through characterizing the prediction of binary spectrum tenancy as data transmissions over the Binary Symmetric Channel (BSC), we analytically and numerically show that the two key assumptions justifying the superior performance of TOP are valid. To evaluate the accuracy of the TOP framework on spectrum tenancy from real world measurement, we set up a Software Defined Radio (SDR) testbed to measure LTE spectrum tenancy by decoding the Downlink Control Information (DCI) to gain high resolution usage at the same granularity with LTE scheduling. Armed with the high resolution data, we adapt the Multi-Layer Perceptron (MLP) algorithm into the TOP framework to validate its performance. The thorough experiments reveal that the TOP framework significantly improves MLP accuracy from 0.84 to 0.91, outperforming many state-of-the-art prediction schemes.