He kept trying different cars (US edition, probably 1.2). But, like I say, it was on a frozen game. So I'd say your 100 races are not going to be any more accurate. You'd need to take samples from as many different genuine games as possible. With different garage contents, different license test results, different non-[size=+1]GT2[/size] memory card contents, etc. And different game history; widely different game days, wins, win/loss ratios, etc. (But most especially, different game days).
After all, you appear to have demonstrated that something about game state determines the colour choice for each prize. That is, if you change something unknown, you will get a different colour choice if that prize is won. It doesn't seem impossible that something unknown changes the weighting of the probability for each prize. Perhaps something about the race you ran affects it; the time taken, the fastest lap, the winning margin, etc.
Prize awarding in [size=+1]GT2[/size] is pretty simple. The potential prizes are all put in a list in memory in a certain location, and there's a count of the prizes to choose from somewhere else. I actually wrote GameShark codes to alter those in various ways.
See also LeGeNd-1's suggestion re: Seattle.