I would first check out your couplers as the cause of backlash. The belt system shouldn't have any backlash if it's tight enough. I can't tell from the picture if you're using solid couplers or not. The best ones are the solid couplers with split collars that tighten around the shaft, not the ones with a single grub screw that presses against the shaft.
The original couplers were rubbish, Im now onto the solid split collar type that tighten around the shaft, pretty sure they're now doing the job.
I tried a new method today to measure my backlash. I have a light objet dsp which I can use to move my y axis using a command within the software. To measure the actual distance I am using my dial indicator pressed up against the gantry. Then I use the command to move the gantry in the y positive direction say 1mm to take the slack then again 1mm in the +Y direction to move the gantry 1mm. Then I reversed the 1mm movement and bring it back the other way. If the dial indicator tells me it moved 0.9mm I have 0.1mm backlash. This is what I measure about between 0.05mm and 0.1mm.
The next thing I did was move it just 0.1mm back and forth this indicated a movement of 0.05mm. Which does not seem like much.
So now im coming to the conclusion it may be as the canadianavenger suggested forward momentum of the carriage weight. But even on a very slow setting the typical backlash traits still exist?
I decided to run a engraving test using the Y-axis to raster the image. I chose a square box and spread the raster lines apart using a high stepover value of 0.4mm. Amazingly at a very slow speed the box looked more like a diamond, suggesting the carriage was losing steps.