Compression predictors store the difference between neighbouring cell values, rather than the values themselves. If your raster has a continuously-varying value across a field, you may end up with a smaller file size when enabling a predictor. If, however, there are sudden changes in value in your raster, a predictor would probably not help much with the file size. Predictors are called filter algorithms in the PNG specification, and describes how they work quite well.
If your raster does not contain floating point values, you will not get any effect from floating point prediction.
Data compression is always a trade-off between the relative values of:
- your time (both processing, and your time spent fiddling with options);
- storage cost;
- compatibility (there will still be some programs, somewhere, that choke on anything but uncompressed TIFFs).
Unless you're on a seriously storage-constrained system, the only time (wild hand-wavy generalization here) that file size matters these days is if you're transmitting data. 7Zipping a whole folder takes a whole lot less time, and ends up with a much smaller overall data size, than working with compressed rasters in your daily workflow.