In econometrics, conventional statistical methods often struggle with modelling complex, non-linear relationships among economic variables. The Wilkie Investment Model and its derivatives, such as the SUPA model, have limitations due to their assumptions of normal distribution and stationarity. Recent development in machine learning methodologies has produced more robust models such as the Graph Neural Network (GNN). In this paper, we propose a GNN based Economic Graph Lottery Ticket (EGLT) algorithm, a new method for economic forecasting based on GNN framework. The EGLT approach generates an optimal adjacency matrix without the need for prior knowledge of existing economic relationships. The algorithm identifies key interdependencies among the economic variables through an iterative process. This paper shows that the EGLT algorithm improves forecasting accuracy when compared to the well-established cascading stochastic economic model, SUPA model, calibrated by using eight major Australian economic variables. The EGLT method reduces the Root Mean Square Error (RMSE) significantly, highlighting its potential for better economic predictions. The results of the paper demonstrate that EGLT approach is more accurate and data-driven for modelling and analysing the dynamics of major economic variables.