Whole-genome sequencing data allow detection of copy number variation (CNV) at high resolution. However, estimation based on read coverage along the genome suffers from bias due to GC content and other factors. Here, we develop an algorithm called BIC-seq2 that combines normalization of the data at the nucleotide level and Bayesian information criterion-based segmentation to detect both somatic and germline CNVs accurately. Analysis of simulation data showed that this method outperforms existing methods. We apply this algorithm to low coverage whole-genome sequencing data from peripheral blood of nearly a thousand patients across eleven cancer types in The Cancer Genome Atlas (TCGA) to identify cancer-predisposing CNV regions. We confirm known regions and discover new ones including those covering KMT2C, GOLPH3, ERBB2 and PLAG1 Analysis of colorectal cancer genomes in particular reveals novel recurrent CNVs including deletions at two chromatin-remodeling genes RERE and NPM2 This method will be useful to many researchers interested in profiling CNVs from whole-genome sequencing data.