We present a dynamical theory of a multi-agent market game, the so-called Minority Game (MG), based on crowds and anticrowds. The time-averaged version of the dynamical equations provides a quantitatively accurate, yet intuitively simple, explanation for the variation of the standard deviation ('volatility') in MG-like games. We demonstrate this for the basic MG, and the MG with stochastic strategies. The time-dependent equations themselves reproduce the essential dynamics of the MG.Agent-based games have great potential application in the study of fluctuations in financial markets. Challet and Zhang's Minority Game (MG) [1,2] offers possibly the simplest example and has been the subject of much research [2]. The MG comprises an odd number of agents N choosing repeatedly between option 0 (e.g. buy) and option 1 (e.g. sell). The winners are those in the minority group, e.g. sellers win if there is an excess of buyers. The outcome at each timestep represents the winning decision, 0 or 1. A common bit-string of the m most recent outcomes is made available to the agents at each timestep [3]. The agents randomly pick s strategies at the beginning of the game, with repetitions allowedeach strategy is a bit-string of length 2 m which predicts the next outcome for each of the 2 m possible histories. Agents reward successful strategies with a (virtual) point. At each turn of the basic MG, the agent uses her most successful strategy, i.e. the one with the most virtual points. Here we develop a dynamical theory for MG-like games based on the formation of crowds and anticrowds.The number of agents holding a particular combination of strategies can be written as a D × D × . . . (s terms) dimensional matrix Ω, where D is the total number of available strategies. For s = 2, this is simply a D × D matrix where the entry (i, j) represents the number of agents who picked strategy i and then j. The strategy labels are given by the decimal representation of the strategy plus unity, for example the strategy 0101 for m = 2 has strategy label 5+1=6. Ω is fixed at the beginning of the game ('quenched disorder') and can represent either the full strategy space or the reduced strategy space [1], depending on the choice of D. Σ is another time-independent matrix, containing all the strategies in the required space in their binary form: Σ r,h+1 describes the prediction of strategy r given the history h (where h is the decimal corresponding to the m-bit binary history string).We introduce a vector n(t): this contains the number of agents using each strategy at time t, in order of increasing strategy label. The vector S(t) contains the virtual score for each strategy at time t in order of increasing strategy label. The vector R(t) lists the strategy label in order of best-to-worst virtual points score at time t; if any strategies are tied in points then the strategy with the lower-value label is listed first. The vector ρ(t) shows the rank of the strategy listed in order of increasing strategy label at time t. Hence R(t) and ρ(t) can...