The Asian Financial Crisis Pitfalls and Possibilities of Predictive Models
Prepared by: February 24, 2000 Arjay Jensen Chip Krotee Hilde Larssen John Mack Gerald Nessmann February 24, 2000
Mapping the Crisis
Objectives and Hypotheses Goals: Build predictive models using data up to the onset of the Asian crisis Analyze model performance out-of-sample during the crisis Expectations: ( - ) Models would likely not accurately predict crisis ( + ) Exercise would be useful in generating insight into alternative means of predicting such economic events Question: What might have worked better, and why?
Analysis of Pre-crisis Predictive Models In-sample
Performance of Predictive Models In-sample
Performance of Predictive Models Out-of-sample
Increasing Predictive Accuracy Method: Build models including out-of-sample data compare these models to models based on in-sample data use observations to construct a thought framework
Pre- vs. Post-Crisis Models
Increasing Predictive Accuracy Hypotheses: There exist variables which can and do accurately predict economic crisis Data underlying these variables are usually hard to obtain due to: government disclosure restrictions scale issues with data aggregation disincentive to disclose proprietary information (intellectual property) George Soros is not the typical investor. His models aren’t the typical models. He has data that you don’t have. His data consists of “key driver” variables, highly correlated with macro-economic events If you have an excellent model, don’t give it away!
Conclusions Market forecasts using standard methodologies generally failed to predict the Asian financial crisis Predictive models using “key driver” variables can have far better success at predicting such crises It is generally difficult to obtain these variables, but it may be possible to derive them indirectly