PREVIEW. Considering the various utility tokens/cryptocurrency markets as "medium size markets". Where decisions of various participants are still significant strategically in deciding payoffs for actions of many other players. Many buyers and sellers are large enough relative the market to affect prices; meanwhile most pitches are technology pitches, basically. Different ones. And ones where features get added over time. So a systems would need need to predict the evolving value of many games with varying rules, each with unknown numbers of rounds and imprecise payoffs, considering many pitches are perfect substitutes for each other, allowing value to vary widely based on minor consumer preferences. Even if most are highly correlated most of the time, which is also partly for the same reason and partly because publicity and resources of the same developers and buyers and sellers are distributed over many competing pitches. (Value of a game is the average best possible payoff.)

One approach is that players don't know which game out of an infinity they are really playing. Players see payoffs, and games with some rules, different ones for different players, and make decisions based on that, but actually all play some one game with similar by continuous variation but not necessarily those same rules, and with payoffs that are the same only up to certain size random variations. Which affects the resources players have in later rounds then. Which changes which actions they have to choose from. Some players know better which games they are playing, and by their actual accumulated rewards and decisions signal that to other players.

Another approach is more abstract, but actually easier to match up predictions neural nets can easily perform.

There are several ways to relate this to mean reversion and pair trading strategies exist.