Capital market based warning indicators of bank runs

• We substantiate the uni- and multivariate capabilities of the LPPL for bank run prediction.
• The univariate discriminative powers of four LPPL parameters emerge up to 40 trading days prior to the default events.
• Our analysis on synthetic data prompts that LPPL structures in financial time series do not arise by chance.
• Previously published results on HSI data are reproduced with less LPPL parameters.
• A multivariate pattern recognition approach based on three LPPL parameters is successfully developed.

In this investigation, we examine the univariate as well as the multivariate capabilities of the log-periodic [super-exponential] power law (LPPL) for the prediction of bank runs. The research is built upon daily CDS spreads of 40 international banks for the period from June 2007 to March 2010, i.e. at the heart of the global financial crisis. For this time period, 20 of the financial institutions received federal bailouts and are labeled as defaults while the remaining institutions are categorized as non-defaults. The employed multivariate pattern recognition approach represents a modification of the CORA3 algorithm. The approach is found to be robust regardless of reasonable changes of its inputs. Despite the fact that distinct alarm indices for banks do not clearly demonstrate predictive capabilities of the LPPL, the synchronized alarm indices confirm the multivariate discriminative power of LPPL patterns in CDS spread developments acknowledged by bootstrap intervals with 70% confidence level.