In analogy with Newton's law of gravity, the gravity law assumes that the number of individuals T ij that move between locations i and j per unit time is proportional to some power of the population of the source (m i ) and destination (n j ) locations, and decays with the distance r ij between them as ,where ! and ! are adjustable exponents and the deterrence function is chosen to fit the empirical data. Occasionally T ij is interpreted as the probability rate of individuals !of traveling from i to j, or an effective coupling between the two locations 24 . Despite its widespread use, the gravity law has notable limitations:i) We lack a rigorous derivation of (1). While entropy maximization 25 leads to (1) with ! = " = 1 , it fails to offer the functional form of f(r).ii) Lacking theoretical guidance, practitioners use a range of deterrence functions (power law or exponential) and up to nine parameters to fit the empirical data 5,7,8,11,14 .iii) As (1) requires previous traffic data to fit the parameters [ ], it is unable to predict mobility in regions where we lack systematic traffic data, areas of major interest in modeling of infectious diseases.iv)The gravity law has systematic predictive discrepancies. Indeed, in Fig. 1a we highlight two pairs of counties with similar origin and destination populations and comparable distance, so according to (1) the flux between them should be the same. Yet, the US census (see SI) documents an order of magnitude difference between the two fluxes: only 6 individuals commute between the two Alabama counties, while 44 in Utah.v) Equation (1) predicts that the number of commuters increases without limit as we increase the destination population n j , yet the number of commuters cannot exceed the source population m i , highlighting the gravity law's analytical inconsistency (see SI, Sect. 4).vi) Being deterministic, the gravity law cannot account for fluctuations in the number of travelers between two locations.Motivated by these known limitations, alternative approaches like the intervening opportunity model 26 or the random utility model 27 (SI, Sect. 7) have been proposed.While derived from first principles, these models continue to contain context specific tunable parameters, and their predictive power is at best comparable to the gravity law 28 .Here we introduce a modelling framework that relies on first principles and overcomes the problems (i) -(vi) of the gravity law. While commuting is a daily process, its source and destination is determined by job selection, a decision made over longer timescales. Using the natural partition of a country into counties (for which commuting data are collected), we assume that job selection consists of two steps ( Fig. 1 b, c):An individual seeks job offers from all counties, including his/her home county.The number of employment opportunities in each county is proportional to the resident population, n, assuming that there is one job opening for every n jobs individuals. We capture the benefits of a potential employment opportunity...