Using machine learning to measure financial risk in China

Appendices

Appendix A Additional figures and tables LDA and search

The below table shows the first 30 words that were identified as being most similar to risk and finance by applying the word2vec algorithm. Judgement has been then applied to select the most relevant words for our final search query.

Table A.1 Expanded search term list

Term	Similarity with finance	Term	Similarity with risk
financing	0.43	risks	0.64
proposed	0.39	leverage	0.51
govern	0.38	problem	0.47
jiwei	0.38	conditions	0.43
prime	0.38	standards	0.43
banking	0.37	nature	0.43
lending	0.36	opaque	0.43
commerce	0.36	given	0.42
securitiesfinance	0.35	painful	0.41
tsipras	0.35	backdrop	0.41
mckinsey	0.35	unstable	0.40
jointly	0.34	extremely	0.40
tighten	0.34	technical	0.40
drafting	0.34	stable	0.40
regulate	0.34	situation	0.39
trustee	0.33	danger	0.39
chiefly	0.32	vasudevan	0.39
factual	0.32	solvency	0.39
nagano	0.32	proponent	0.39
broaden	0.32	regulation	0.39
jams	0.32	defaults	0.38
padoan	0.31	governance	0.38
amplified	0.31	quality	0.38
guidelines	0.31	unwinding	0.38
poorest	0.31	deterioration	0.38
benefitting	0.31	causes	0.38
peer	0.31	avoid	0.38
designed	0.30	vulnerabilities	0.38
shinzo	0.30	volatility	0.38
calls	0.30	risky	0.37

Figure A.1: Other identified topics

Source: Authors’ calculations.

Notes: The size of the words in the word clouds reflect the number of occurrences of that word in the topic considered. As common with this procedure, some topics end up being a collection of common words that do not form a logically coherent group. These are indicated with unk. 1” and unk. 2” which is short for unknown/unidentified.

Figure A.2: Indices obtained with the LDA procedure.

Source: Authors’ calculations.

Figure A.3: Indices obtained with the LDA procedure.

Source: Authors’ calculations.

Figure A.4: Proximity of the topics identified with the LDA procedure.

Source: Authors’ calculations.

Appendix B Additional results: Robustness

Table B.1. and B.2. present results from regression analysis run to identify the topics used to construct two additional financial risk indicators. More precisely, both tables present 6 different models which differ for the indicators used as regressors. Table B.1 presents regression analysis where the dependent variable is the China Dow Jones index, while Table B.2 presents regression analysis where the dependent variable is the CISS index. Starting from Model 1, where the regressors are the topics used to construct our core index in Section 4.1., we try in turn different model specifications by changing the regressors guided by their pairwise correlation with the dependent variables - with the aim of finding the model that maximizes the adjusted R-squared of the regressions. In both cases the model that delivers the highest adjusted Rsquared is Model 6. In the case of the Dow Jones China the model explains 0.14% of the variation of the dependent variable, in the case of the CISS the explained variance is almost double 0.35.

Table B.1.: Regression analysis results with China Dow Jones index as dependent variable.

(1) (2) (3) (4) (5) (6)

VARIABLES DowJones DowJones DowJones DowJones DowJones DowJones

FinancialMarkets -40.92** -32.05** -34.45** -34.83** -36.80** -33.35**

(17.94) (15.57) (15.03) (15.49) (15.03) (15.14)

Banking 25.54

(19.74)

FX 6.898

(14.18)

RealEstate -41.02

(37.23)

CorpProf -62.07** -27.53 -38.46 -39.03 -46.84*

(29.60) (31.31) (25.51) (26.11) (25.90)

CorpInv 57.27** 72.40** 82.40*** 82.19*** 112.1*** 86.47***

(26.81) (30.68) (25.79) (25.92) (31.28) (27.97)

GrowthOutlook -9.293 -31.31**

(15.39) (13.99)

LeisANDHosp -162.5*** -163.9*** -165.0*** -205.0*** -222.1***

(51.32) (51.18) (52.29) (56.65) (56.93)

Energy -92.48** -95.62** -96.18** -113.3*** -112.7***

(42.13) (41.74) (42.17) (42.90) (42.72)

IndANdManu 4.600

(42.69)

ForeignAffairs 48.96* 71.78**

(29.47) (32.38)

Constant 0.544* 1.507*** 1.452*** 1.445*** 1.183*** 1.185***

(0.290) (0.344) (0.331) (0.337) (0.367) (0.363)

Observations 213 213 213 213 213 213

Adjusted R-squared 0.056 0.126 0.129 0.124 0.136 0.143

Standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1

Table B.2.: Regression analysis results with China CISS index as dependent variable.

	(1)	(2)	(3)	(4)	(5)	(6)
VARIABLES	CISS	CISS	CISS	CISS	CISS	CISS
FinancialMarkets	-11.22	-13.57	-12.07	-9.304	-10.82
	(16.64)	(15.90)	(15.78)	(15.69)	(15.50)
Banking	32.50*	29.32*	32.17*	29.46*	38.05**	35.47**
	(18.20)	(17.40)	(17.02)	(16.92)	(17.07)	(16.79)
FX	-2.065	-14.92	-13.88	-10.32	-3.117
	(13.18)	(12.93)	(12.85)	(12.85)	(13.03)
RealEstate	-44.55	-61.15*	-58.36*	-78.32**	-72.54**	-75.74**
	(35.79)	(34.38)	(34.18)	(35.18)	(34.81)	(34.10)
CorpProf	89.78***	24.31
	(27.17)	(29.99)
CorpInv	96.02***	148.3***	159.4***	185.0***	138.1***	132.3***
	(26.10)	(27.66)	(23.99)	(26.74)	(32.74)	(32.05)
GrowthOutlook		65.46***	71.58***	85.63***	80.98***	77.79***
		(15.04)	(13.00)	(14.52)	(14.46)	(13.73)
IndANdManu				-88.13**	-87.14**	-94.28**
				(42.03)	(41.49)	(40.51)
ArtandCulture					134.8**	141.3***
					(55.76)	(53.97)
Constant	-1.986***	-2.419***	-2.401***	-2.236***	-2.234***	-2.206***
	(0.263)	(0.270)	(0.269)	(0.278)	(0.274)	(0.268)
Observations	192	192	192	192	192	192
Adjusted R-squared	0.249	0.316	0.317	0.329	0.346	0.350

Standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1

Table B.3.: Correlation across different financial risk indices and selected variables

	Core index	PCA based	Dow Jones Based	CISS based	CISS	Dow Jones China
Core index	1	0.15	0.78	0.85	0.39	-0.17
PCA based index	0.15	1	0.49	0.18	-0.09	-0.12
Dow Jones based index	0.78	0.49	1	0.83	0.40	-0.28
CISS based index	0.85	0.18	0.83	1	0.51	-0.18
CISS	0.39	-0.09	0.4	0.51	1	-0.24
Dow Jones China	-0.17	-0.12	-0.28	-0.18	-0.24	1

Figure B.1: Impulse-response functions of macro-financial variables to shocks in the overall financial risk indicator. 6 lags instead of 12

Source: Authors’ calculations.

Notes: IRFs report percentage changes for all variables excluding EMBI Global spreads and the 7-day repo rate which are reported in bps. Dotted lines report the 68% credibility interval. Retail sales is used instead of industrial production for China as an alternative activity variable.

Figure B.2: Impulse-response functions of macro-financial variables to shocks in the overall financial risk indicator. 8 lags instead of 12

Source: Authors’ calculations.

Notes: IRFs report percentage changes for all variables excluding EMBI Global spreads and the 7-day repo rate which are reported in bps. Dotted lines report the 68% credibility interval. Retail sales is used instead of industrial production for China as an alternative activity variable.

Figure B3: Impulse-response functions of macro-financial variables to shocks in the overall financial risk indicator. Alternative activity variable for China.

Source: Authors’ calculations.

Notes: IRFs report percentage changes for all variables excluding EMBI Global spreads and the 7-day repo rate which are reported in bps. Dotted lines report the 68% credibility interval. Retail sales is used instead of industrial production for China as an alternative activity variable.

Acknowledgements

The authors would like to thank an anonymous referee and participants at the 12th Annual International Conference on the Chinese Economy” organised by the Hong Kong Institute for Monetary and Financial Research of the Hong Kong Monetary Authority as well as participants at a seminar at the European Central Bank (ECB) for useful comments and suggestions. We are grateful to Martin Rasmussen for his excellent research assistance at an early stage of this project.

The paper was written while the author Apostolos Apostolou was at the European Central Bank. The analysis and conclusions expressed herein are those of the authors and should not be interpreted as those of the ECB or the IMF.

Alexander Al-Haschimi

European Central Bank, Frankfurt am Main, Germany; email: Alexander.Al-Haschimi@ecb.europa.eu

Apostolos Apostolou

International Monetary Fund, Washington D.C., United States; email: aapostolou@imf.org

Andres Azqueta-Gavaldon

Sensyne Health, Oxford, United Kingdom; email: Andres.Azqueta@gmail.com

Martino Ricci

European Central Bank, Frankfurt am Main, Germany; email: martino.ricci@ecb.europa.eu

Postal address 60640 Frankfurt am Main, Germany

Telephone +49 69 1344 0

Website www.ecb.europa.eu

All rights reserved. Any reproduction, publication and reprint in the form of a different publication, whether printed or produced electronically, in whole or in part, is permitted only with the explicit written authorisation of the ECB or the authors.

This paper can be downloaded without charge from www.ecb.europa.eu, from the Social Science Research Network electronic library or from RePEc: Research Papers in Economics. Information on all of the papers published in the ECB Working Paper Series can be found on the ECB’s website.

PDF ISBN 978-92-899-5509-6 ISSN 1725-2806 doi:10.2866/565243 QB-AR-23-004-EN-N

[1] Shadow banking refers to non-bank financial institutions providing bank-like financial services and in particular credit instruments.

[2] The measurement and impact of economic policy uncertainty in China has been studied in some detail; see, e.g., Davis et al. (2019), He et al. (2020), Huang and Luk (2020), Li and Wu (2020), Liu and Zhang (2020), and Sha et al. (2020), among others.

[3] According to Reporters Without Borders, an international non-profit organization which seeks to defend and promote freedom of information, China is ranked 175^th in over 180 countries in the world in 2022 for the freedom of its press.

[4] We use the Python implementation of the word2vec algorithm.

[5] Table A.1. in the Appendix shows the words obtained with the world2vec algorithm.

[6] See Figure A.4 in Appendix A for the proximity map.

[7] See also Hollo et al. (2012) for the methodology used to construct the index.

[8] The China-CISS defines systemic risk events as episodes when both the covariance and coextremeness across markets are jointly high as systemic stress events” and is based on 13 financial indicators of the equity market, the bond market, financial institutions and the foreign exchange market” (Ma et al., 2019).

[9] On the informativeness of the 7-day repo rate on China’s monetary policy stance see Kamber and Mohanty (2018) among others.

[10] The choice of the number of lags is rather standard considering the model frequency. Results are qualitatively and quantitatively similar if we change the number of lags to 6 or 8 (see Appendix B, Figure B1 and B2).

[11] This result seems to be confirmed also if we use alternative proxies of activity such as retail sales as shown in Appendix B Figure B3.

[12] Table B.3. in Appendix B presents a correlation matrix of the four indicators.