As has been suggested already++, rand is as good a starting point as any for guessing a random number. You could probably try a best-fit polynomial, though it's doubtful it would perform any better at guessing the next random number than rand() would. It really depends on how random it is.

In the 90s, I would have said, "make an artificial neural network", and I guess nowadays, I would say the equivalent "machine learning"... but given that no one has broken the stock market with their hugely-above-average performing ANN/ML/AI systems, I doubt that'd have a much higher success rate than rand().

If it's a computer-generated random number (PRNG) to begin with, you could try to reverse engineer whatever PRNG is being used to by the "roulette" wheel, but with all the various manipulate-and-mod methods, with their huge numbers of possible state values, make that a daunting perspective. (If you know which implementation of their rand() function, you'd have a better chance of brute-forcing the state, especially if it's a smaller number of bits in their algorithm.)

If it's something more truly random (ie, a non-rigged physical process), then all bets are off, and rand() is as good as you're likely to get. If it is rigged, then you've got a better chance, and doing best fits, or other statisitical analysis, would be a good starting point.

Oh, I seee that Happy-the-monk mentioned rigged roulette wheels while I was writing this up, and hdb suggested some statistical packages to use to analyze data. That makes my post all-but-superfluous, but it will still be posted due to sunk costs. :-)