Autoguiding follows the typical Control Theory (usually taught in Electrical Engineering) feedback loop, albeit in very crude form.
In a feedback loop, the output (in this case, the guide star's deviation from its nominal centroid) is compared with a target (in this case, zero change in centroid location, unless you are doing some advanced thing to predict any mount error).
This comparison results in an error signal (in our case, a vector of two error signals in the RA and declination axes). The error signal is amplified, what is termed loop gain (in our case, it corresponds roughly to the the two Aggressive settings in the PHD2 parameters).
One of the objectives in Control Theory is to make the output follow what you target as the desired output (in our case, the ideal fixed guide star location in the guide plate) in the presence of external influence. In our case, the primary external influence is the mount's gear accuracy.
If you apply too low a loop gain (in our case, the aggressiveness values), you cannot bring the output error down sufficiently. Apply too much loop gain, and the feedback system can oscillate uncontrollably.
There are potentially other influences too than just imprecise gears, especially with legacy German mounts with the need for counterweights.
When the gears driving the mount does not have enough contact surface area, as with legacy German mounts, there is insufficient torque to move the instrumentation, and because of that, you need to use counter-weights.
The imbalance will cause an extra bias to the output, which the loop gain has to countermeasure.
Typically, this will exhibit itself as a constant error (as you saw): that is when the loop gain is just sufficient to cancel out the bias.
When there is a constant error at the output, you need to apply a larger loop gain (aggressiveness) to counter the bias. The problem is that too much loop gain will cause overcorrections that result in oscillations (your guide graph swings back and forth like a pendulum).
The usual way is to make sure your axes (I am assuming you do not have a fork type mount or a Harmonic Drive[tm] mount) are balanced before fiddling with the aggressiveness numbers, leaving just a slight imbalance to keep the gears fully meshed (problem with the imbalance is that the direction of the imbalanced that you need will change after you perform a Meridian Flip).
With the legacy German style mounts, unless all instrumentation are perfectly vertically over the declination axis, you can suffer from that is sometimes called the "third axis imbalance" (do a Google). Even an off axis camera that goes not protrude vertically up or down can cause this "third axis" problem. This is why you see proper Newtonian camera arrangements have the cameras pointed straight up into the sky or straight down toward the ground. Experience users do not mount their camera sideways on a Newtonian (not a problem with Harmonic Drive mounts).
Read up on balancing and particularly "third axis" balance. Then try to balance your instrumentation on the mount. If there is still some error in the baseline of the guide graph, try increasing the aggressiveness (loop gain), but monitor carefully that the graph does not go into oscillations (if it oscillates, back off the aggressiveness).
Good luck,
Chen
EDIT: I have seen people claim that a constant error like what you have seen is perfectly OK. It is not. If it is a bias in balance, said balance will change over the course of the night. As a result, the offset from the star centroid will move, causing the guide star to drift on your main camera's plate (it will almost seem like a differential flexure between the guide camera and your main camera, but is really caused by autoguiding error).
