It is difficult enough to infer CNVs reliably with good WGS data so the algorithm can try to build a background on a large number of windows. You can see this on the fact that there are dozens if not hundreds of different CNV tools out there. I would not trust any results from ATAC-seq/ChIP-seq on this regard. Coverage in these is punctual since they assay like 1-5% of the genome in many cases, coverage is highly uneven both within and outside of peaks, and their purpose is simply not to call CNVs.

This is from the supplementary text (page-15) of TCGA ATAC-seq paper by Corces et al. Science 2018

"ATAC-seq data analysis – Inferring copy number amplification To infer DNA copy number amplifications from ATAC-seq data, we first tiled the genome into 2-Mbp windows using “tile” of genomic ranges for chromosome sizes in R. These window positions were then filtered against regions with known artefactual mapping issues using the ENCODE blacklist with the “setdiff” function in R. Then, the number of insertions within each filtered window was determined using “countOverlaps”. Next, the insertions per bp was determined within each filtered 2-Mbp window. Then, the percent GC content was computed for each filtered 2-Mbp window using the hg38 BSgenome in R. To estimate if a region is amplified, for each window we took the 100 nearest neighbors based on GC content and computed the average log2(fold change). If this was above 1.5 we considered this region as a candidate for amplification. This window size best captured smaller known amplifications, but added more false positives compared to 10-Mbp windows."