This research aims to verify the hypothesis that lane traffic distribution initiates traffic regimes occurrence in freeways, and in particular to unveil the inter-dependence of lane vehicle allocation and congestion formation. The developed methodological framework consists of (i) a clustering approach to separate traffic regimes and (ii) a multi-regime modeling approach to identify the critical traffic variables that influence the traffic distribution in freeway lanes. The prediction of regime occurrence, through lane-related parameters, could be implemented in control algorithms, in order to ameliorate their efficiency by timely activating a policy. A neural network algorithm is invoked, in order to detect in an unbiased approach homogeneous groups of the prevailing traffic regimes during peak periods, and a denoised input space. Based on the optimized clusters of traffic states and peak periods and on the rationale that congestion propagates from left to right in the onset of peak period indicating a transition to another state, lane density distribution ratio (LDDR) and density of the right or left lane, depending on the state, are introduced as promising determinant response variables. LDDR appears to be site-independent and is intermittently occurring during congestion and free flow conditions. To efficiently forecast stream dynamics, their parameterization is effectuated with lane-related variables. Static and dynamic regression models are deployed for congested or saturated regimes, for free flow regimes and also for the entirety of traffic states. Static models are formed, pursuing a simple approach that would ensure feasibility of implementation. Statistical assessment indicates that the prediction models yield significant prediction accuracy.