Investments for transportation infrastructure change the relative advantage that each region has through changes in accessibility to material/service markets and labour markets. For estimating the rate of return of transportation investments, we need a comprehensive model to determine how specialized transportation policies for enhancing the performance of transportation networks impact regional economies. The basic requirement to achieve this is to build an interregional model consistent with a computable economic general equilibrium model. The spatial computable general equilibrium model is an extension of CGE models that include interregional trade models, but empirical studies on the estimation procedures for behavioural parameters included in the interregional trade model have not been fully discussed so far. Therefore in this paper, we examine two approaches for estimating inter-regional trade coefficients that are used in the spatial computable general equilibrium model developed by the author. The first approach is based on the gravity model, in which a conventional statistical procedure is applied to identify parameters of trade coefficients. The second approach may be called a model-free approach and uses neural network models to model the complexity caused by differences between intra-regional and inter-regional trade coefficients. We show that traditional statistical approaches are insufficient to capture intra-regional trade coefficients because large variations exist in intra-regional commodity flows among regions and among commodity types. To overcome this weakness, we propose neural spatial interaction models and show that the neural network model is more flexible and outperforms the traditional method in capturing inter-regional trade coefficients. JEL Classification: R11, R13, R15