#ensembl

“GRCh37/hg19 and GRCh38 are genome builds rather than annotations, which describe where features are in a given genome build.

The actual sequences you'll get from NCBI/UCSC/Ensembl will be identical, but their annotations will be different 

and (importantly) updated at different frequencies.

NCBI's annotation is the "refseq" dataset (the "refGene" track in UCSC), which is essentially a subset of the UCSC and Ensembl annotations. 

•UCSC's annotations are kind of a mess. 

You'll find genes with the same ID on multiple strand and multiple chromosomes, which makes them a bit useless.

•Ensembl's annotations typically contain more features than UCSC (so a bit more noise), 

but they're otherwise much better put together (e.g., you'll never find a gene ID on different strand or different chromosomes) and their IDs are typically easier to map to other things (e.g., gene names, GO and pathway memberships). 

Ensembl also updates its annotation fairly often and versions everything nicely, so it's quite convenient to report what version you used in a paper (reproducibility is always a good thing).

Given the choice, use the Ensembl annotation.