One quick thought is to simply go back to the supply and demand curves. If you can predict the inelasticity of the supply and demand of a certain stock, you've got the first order taylor approximation of its new equilibrium price after a certain point.
The tricky part is that markets tend to get more elastic with time. If the price of gold hits $1500 for 2-3 hours, you'll start clearing out the market makers' inventories. If it hits $1500 for 2-3 months, the market will broaden as people start melting down their coins and jewelry. So the market might be REALLY inelastic for 2-3 seconds after the trade and then the slopes of the supply/demand curves might get cut in half after a minute and then maybe cut more after an hour and more after a day.
So one (admittedly very naive) model is to try and guess the slope of the supply and demand curve over the period you're interested in. There's no real control case for "What happens if buyer A didn't buy this instrument", so this calls for a little more thinking than just doing some linear regression. If you have the applicable data, this might be an interesting case for using a guided learning process that examines the relationship between market-takers' orders and price moves over the period you want to examine. The volume-weighted average of predicted price move for a transaction to stock traded for that transaction under similar market conditions (volume, volatility, news, etc) might give you some idea of what impact a trade will have on the market going forward.
***These are just the thoughts of a fixed income financial programmer with a CS Theory & Algorithms background. I am not a quant or financial engineer and defer to anyone with any sort of experience in the field of quantitative market research.***