Here I'm trying to write something partly based on Cont's first model in the previous post. I plan to skip the Laplace transform and go for Monte Carlo simulation. Also, I'm trying to abandon the assumption of unified order sizes. To implement that, I need to shift from a Markov chain which is supported by discrete spaces, onto some other stochastic process that is estimatable. Moreover, although I actually considered supervised learning for this problem, I gave it up at last. This is because my model is inherently designed for high frequency trading and thus training for several minutes each time would be intolerable.