I have scrolled through the data sheet of this device, and first things first: never expect accurate values from a device that costs like 5 euro. looking at you equation, I think you are calculating the angle by effectively integrating the gyro value. Since integrating a sensor value is not an accurate estimation (because noise), the angle measured from the accelerometer could compensate for this deviation. This means that if the device is hovering statically, the angle saturates to the value of the accelerometer angle. The accelerometer calculates the angle using the gravitational acceleration (which it measures). However, the accelerometer also measures any other force acting on the system and that causes your accuracy problem.
The solution to an issue like that is hard, increasing the weight of the accelerometer improves estimation at constant velocities. decreasing it prevents measurement offsets when accelerating.
However, the sensor seems to have a couple tricks for you to use (DPM and noise filtering), so if you havent already, you might to check these.