Estimating causal relations of dynamic models from real-life data

In System Dynamics method, one tries to use data as much as possible, since model construction time and subjectivity can be reduced by the data analysis. In this research, our focus is the use of data analysis in (1) determining the polarity of causal effects and (2) discovering stock variables in a model. For determining the polarity of causal effects, we propose an algorithm, discoverpolarity, which is tested with seven data-sets. Then, the results are compared with Spearman's correlation analysis. The results show that discoverpolarity outperforms correlation analysis and is capable of obtaining useful and meaningful results when the input variables are properly selected and data-set comprises enough representative points in the causal domain. However, when the data only consists of all perfectly correlated data points, discoverpolarity may return multiple possible polarities instead of a unique solution. In addition, the modeler must determine the proper threshold values used in the algorithm. In further research, we plan to make discoverpolarity more robust to the input parameters. After enough tests with synthetic data, the algorithm must be tested with real data before it can be used in real-life modeling. Finally, the mathematical forms of the causal formulations can be estimated in further research, by extending the proposed algorithm. For the second thesis purpose, discovering stock variables, curvefitting algorithm is created and simulation-generated 'synthetic' data is analyzed in this algorithm to be able to evaluate the validity of the results obtained. The method is applied to three cases. We conclude that only in certain conditions, the algorithm may discover correct stock variables. In further research, we aim to categorize the monotonic relations where algorithm can find the correct stocks. In addition, we plan to focus on extending the curvefitting algorithm so that it can also analyze cases with multiple cause variables.