If you had 50 marbles, numbered 1 to 50, there would be a 10% chance of selecting a specific desired number with any 5 random selections from a set of 50. So 43% is only four times better than random guessing. Does the software know what the valid 50 numbers are, and pick the closest match? If so, the results are not impressive.
Whoa there... the number 50 is the size of their test sample, and nothing to do with the number of possible PINs, so your probability calculation is meaningless. In other words, their program is being asked to guess what the PIN is, and not "guess which one of these 50 known patterns/PINS" we've given you".
The way you should look at it is that each random PIN guess (having no accelerometer hints) would be right 1/10,000 of the time (ie, 0.0001). If they can guess the PIN 43% of the time with 5 guesses, then their success rate per guess is 0.43 / 5 or 0.086. So in fact their ability to guess a PIN is actually 0.086 / 0.0001 = 860 times better than chance, not four times better!