Calculating a trend line and storing deltas off of that trend line might be mildly better for some series. Having an AR model and storing errors of the model might work even better, although the obvious (and common) problem is which AR model of the many possible ones you go for.
[1] https://en.m.wikipedia.org/wiki/Linear_predictive_coding
Edit: It seems that if you detrend the data, the "delta off the last delta", if I understand it correctly, is an MA(1) model. I think. Maybe. My math is wobbly.
Other aspects that touch on that is that we are usually not dealing with large amount of data, so there isn't enough room for a general algorithm to play its strengths.
For various reasons, we'll rarely see the raw data higher than 1 - 2 KB for each segment.
For this case that might mean convert the data into deltas, then write those deltas in a byte-oriented variable length integer fashion. The general purpose compressor can then pick up most of what you have left on the table by not being super careful with the balancing of the chosen integer format, and it might find some other useful patterns in the data without you having to worry about those.
Theoretically a single compressor crafted purely for the exact characteristics of the data at hand would be able to do a better job. But in practice the skill and time required to beat the hybrid approach tends to be a showstopper.