An innovative machine learning workflow to research China’s systemic financial crisis with SHAP value and Shapley regression

Financial Innovation

Table 10 RF-Shapley regression with China data (sorted by the size of share)

Feature	Sign^a	Share (SSCvalue)	Coefficient	Exponentiation	Change of \(\frac{{y}_{risk}}{1-{y}_{risk}}\)
YieldCurveSlope	+	0.1660***	0.0359	1.0366	Significantly increased by 3.66%
Credit	+	0.1398***	0.0075	1.0075	Significantly increased by 0.75%
NarrowMoney	+	0.1382***	0.0116	1.0117	Significantly increased by 1.17%
Credit-Dif	+	0.1106***	0.0082	1.0082	Significantly increased by 0.82%
CreditCostGrowth	+	0.1091***	0.0327	1.0332	Significantly increased by 3.32%
DidYield	−	0.1081***	0.0311	1.0316	Significantly decreased by 3.16%
CPI-Dif	−	0.0823***	0.0190	1.0192	Significantly decreased by 1.92%
GlobalNetExport	+	0.0611***	0.0134	1.0135	Significantly increased by 1.35%
FinancialConcentration	−	0.0432***	0.0091	1.0091	Significantly decreased by 0.91%
NetExport	−	0.0417***	0.0062	1.0062	Significantly decreased by 0.62%

***, **and * respectively means that the null hypothesis is rejected at the significance level of 1%, 5%, 10%
^a Since we don’t have risk target of China, it is impossible to determine sign for China with Logistic Regression. Therefore, we took the feature value as the independent variable, the corresponding SHAP value as the dependent variable, and then fitted the line to obtain the sign. Moreover, to be robust, we averaged the SHAP value calculated by RF and GBDT for features.