Or maybe you could build an army of twitter bots, write some sockpuppet management software, and sell that... I hear the US is lagging behind its enemies in that regard.
It seems like a very tough problem to me.
You would probably be best to narrow the scope of the problem, like, "ok I want to predict the probability of an attack happening on just one train station."
Then you would need real time data of everyone coming and going. Real time data of metal detectors, fluid detectors, contents of garbage bags, pressure plates on floors and walls (for bags that are left behind).
There are a lot of vision, AI, GPS, demographic data problems that would all go into solving this problem. And there will be a ridiculous amount of false positives.
You probably would be better off training AI to identify needles in haystacks, or flakes of gold in a pan
And of course, the US government is an entity capable of absorbing investment losses -- but benefits greatly from potential advances in defense technology.