Linear versus non‐linear modelling: A systemic analysis with empirical data on credit and investment

The purpose of this article is to shed light on the capabilities of linear versus non‐linear forms of modelling complex systems. A comparative analysis is realized, which shows the differences between two established methods — regression analysis (RA) and system dynamics (SD). To realize that methodological comparison, we need a benchmark in substantive terms. Therefore, we embed our analysis in a realistic scenario. A set of hypotheses related to the banking and investment sector is tested on the basis of rich empirical macro‐ and firm‐level data. We examine (1) whether the central bank's monetary policy affects the lending rate of commercial banks and (2) whether the lending rate of the commercial banks affects non‐financial firms' investment. A RA is conducted, to test our hypotheses on the assumption of linear relationships between the variables. To explore the non‐linear relationships between the variables, and to test the robustness of our regression results, we then model the system under investigation using the SD method(ology).1 The regression results show that the monetary policy rate has a positive and immediate effect on the lending rates and a negative but lagged effect on the bank loan supply. The long‐term lending rates are negatively associated with corporate investment, whilst the bank loan supply has no direct effect on corporate investment. Although the regression results indicate that monetary policy rate has no direct effect on corporate investment, the SD model shows that the monetary policy rate exerts a negative effect on corporate investment. Hence, the two types of modelling can come to different results, given the distinct implications of the exogenous perspective (i.e. the exogenous condition of predictor variables typically required in RA) and the endogenous perspective (i.e. the predictor variables are endogenized, as in SD). Comparing the outcomes of the two methodologies, we conclude that they complement each other. RA is easier to model. SD enhances the RA method by enabling a better understanding of non‐linear phenomena.