The financial and operational impacts of European SMEs’ use of trade credit as a substitute for bank credit

ABSTRACT

We study the impacts of the use of trade credit on SME financial performance and operational distress in a sample of 74,036 SMEs across 19 EU countries between 2006 and 2015. Under the premise that trade credit acts as a substitute for bank credit, our results show that supplying trade credit improves profitability, but we show little evidence that such an investment is more profitable for bank credit richer SMEs, although such firms did redistribute more bank fund through trade credit to their customers. For receivers, we show that the use of trade credit finance alleviates operational distress, especially for those SMEs facing liquidity constraints, such as those which have less access to bank credit or under credit tightened periods. This distress reduction effect is also reflected in their profitability indicators. However, the longer the average collection period and credit period, the less effective the trade credit effects respectively on improving SME profitability and reducing operational distress.

KEYWORDS:

• Trade credit
• Bank lending
• Credit constraints
• SME profitability
• SME distress
• European Union

JEL classifications:

• G21
• G30
• M41

Acknowledgement

We thank Prof. Chris Adcock and two anonymous referees for their valuable and insightful comments, which have hugely improved the paper.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

Note: please see variable definitions in Appendix 1.

Notes: *, ** and ***, respectively, represents significance at 10%, 5% and 1% level. Standard errors are reported in parentheses and they are clustered at firm-level. F.E. stands for fixed-effect estimator. The Adjusted R-squared for F.E. models account for variations captured by firm fixed-effects. Note 1 explains whether the firm-level variables are one-year lagged. Note 2 (Model 7) annotates that all dependent and independent variables are in 1-year difference, and Model 7 is estimated by OLS estimator with robust standard errors. Model 8 is estimated by two-step ‘System’ GMM estimator (Arellano and Bover 1995; Blundell and Bond 1998). The results of relevant diagnostic tests designed for endogeneity correction models (i.e. Model 8), e.g. Overidentification tests and serial correlation tests, are not presented for space reason, same in below tables.

Notes: *, ** and ***, respectively, represents significance at 10%, 5% and 1% level. Standard errors are reported in parentheses and they are clustered at firm-level. F.E. stands for fixed-effect estimator. The Adjusted R-squared for F.E. models account for variations captured by firm fixed-effects. Note 1 explains whether the firm-level variables are one-year lagged. Note 2 (Models 7 and 10) annotates that all dependent and independent variables are in 1-year difference, and these two models are estimated by OLS estimator with robust standard errors. Models 8 and 11 are estimated by two-step ‘System’ GMM estimator (Arellano and Bover 1995; Blundell and Bond 1998).

Notes: *, ** and ***, respectively, represents significance at 10%, 5% and 1% level. Standard errors are reported in parentheses and they are clustered at firm-level. F.E. stands for fixed-effect estimator. The Adjusted R-squared for F.E. models account for variations captured by firm fixed-effects. Note 1 explains whether the firm-level variables are one-year lagged. Note 2 (Mode 5) annotates that all dependent and independent variables are in 1-year difference, and the model is estimated by OLS estimator with robust standard errors. Model 6 is estimated by two-step ‘System’ GMM estimator (Arellano and Bover 1995; Blundell and Bond 1998).

Notes: *, ** and ***, respectively, represents significance at 10%, 5% and 1% level. Standard errors are reported in parentheses and they are clustered at firm-level. F.E. stands for fixed-effect estimator. The adjusted R-squared for F.E. models account for variations captured by firm fixed-effects.

Notes: *, ** and ***, respectively, represents significance at 10%, 5% and 1% level. Standard errors are reported in parentheses and they are clustered at firm-level. F.E. stands for fixed-effect estimator. The adjusted R-squared for F.E. models account for variations captured by firm fixed-effects.

Notes: *, ** and ***, respectively, represents significance at 10%, 5% and 1% level. Standard errors are reported in parentheses and they are clustered at firm-level. F.E. stands for fixed-effect estimator. The adjusted R-squared for F.E. models account for variations captured by firm fixed-effects. Section ‘notes’ explains individual restrictions as detailed.

Notes: *, ** and ***, respectively, represents significance at 10%, 5% and 1% level. For each model, the 1st, 2nd and 3rd line reports the corresponding result where ‘Altman Z-score Trinary form’, ‘Altman Z-score Binary form’ and ‘Altman Z-score Binary form without grey zone observations’ are used as dependent variables, respectively. Standard errors are not reported for space reason. RE OLOGIT stands for random-effect panel ordered-logit estimator, RE-PP stands for random-effect panel probit estimator. The calculation of ‘Goodness-of-fit’ values for all LDVM models is explained in footnote 41.

1 For a review of trade credit policy contract structure, please see Bazzana et al. (2019).

2 Our sample includes the following countries: Austria, Cyprus, Germany, Denmark, Estonia, Spain, France, UK, Greece, Croatia, Hungary, Ireland, Lithuania, Latvia, Malta, Netherlands, Poland, Portugal, and Slovenia. Iceland as a non-EU state is included along with the other 19 EU members in our sample. However Icelandic SMEs contribute only 0.17% to our sample’s observations and we decided to keep the sample to provide a full picture of SME trade credit in European countries.

3 Martinez-Sola, Garcia-Teruel, and Martinez-Solano (2014) have called for a study on trade credit investment effects in a period of economic downturn. We hope our focus on trade credit investment can answer their question and can be a supplement to their study.

4 Casey and O’Toole (2014) argue that trade credit is not necessarily more costly than bank credit for European SMEs during credit supply constrained periods (e.g. in the aftermath of Financial Crisis). Marotta (2005) supports this view in an Italian sample.

5 We do not have access to trade credit agreement data and/or relationships between trade credit suppliers and receivers. We acknowledge that trade credit supplied to sample receivers can come largely from out-of-sample firms in this instance.

6 However, our data cannot indicate if the demand variation is predictable or not by managers.

7 Please see Wang, Han, and Huang (2020) for a discussion on the intrinsic bank financing constraints of SMEs.

8 In contrast, in sound economic conditions, the likelihood of firms’ default rates falls, banks’ costly screening and monitoring process become less marginally beneficial and thus, competition on price becomes fiercer.

9 EU countries in our sample include Austria, Cyprus, Germany, Denmark, Estonia, Spain, France, UK, Greece, Croatia, Hungary, Ireland, Lithuania, Latvia, Malta, Netherlands, Poland, Portugal, and Slovenia. Ideally, we would have data from all the 28 EU countries, instead of 19. However, for the remaining EU countries that are not included in our sample, they either do not have enough observations on the key variables in our sample, are lack of observations for early sample period (e.g. before 2011), or the sample is strongly biased by several industries. Hence, the 9 remaining EU countries are not included in this study. In addition, for the study of hypotheses set 3 (H3s), Cyprus, Denmark, UK, and Ireland are dropped from the initial sample, due to insufficient accounting information to gauge the Z-score that is used for testing H3s.

10 Due to the low quality of SME accounting information in the full Amadeus subscription (e.g. unreasonable and missing values), we use the sub-subscription of Amadeus which contains firms with more valid data coverage. However, our database has a lower representativeness on micro-sized and/or start-up firms compared with existing studies such as Fungacova, Shamshur, and Weill (2017).

11 Estimation is used when partial SME-defining-information is not available (see Amadeus online help).

12 We screen out sample firms based on their activity locations (e.g. overseas territories), industries and legal forms. Details are available from authors on request.

13 We manually manage firm data outliers instead of using winsorisation as the latter could distort and misrepresent original data although the number of observations is not sacrificed. In total, less than 1% of the initially collected data are removed, and full methodologies on outlier cleaning process are available from the authors on request.

14 According to Amadeus user manual and Fungacova, Shamshur, and Weill (2017), short-term and long-term bank credit are respectively recorded and named as ‘loans’ and ‘long-term debt’ in the Amadeus database. Total bank credit is the sum of these two elements.

15 All variables with a prime sign are for Equation (2). We include firm size, cash-richness, tangibility, liquidity and growth opportunity in both equations. At country-level, banking market competition and bank credit supply are added in Equation (1) to capture SME bank financing conditions, and interest rate is added in Equation (2) to reflect credit price. GDP growth rate as an indicator of economic condition is added in both equations.

16 The standard errors for fixed-effect models are clustered at firm-level. Results do not change if they are clustered at industry or country level (same for all F.E. models throughout this paper).

17 ‘t’ and ‘t-1’ in Note 1 respectively denotes a contemporaneous model and an explanatory variable one-year lagged model (same throughout the paper). Macroeconomic variables are all one-year lagged to adapt a response lag of small microeconomic agents toward economic fluctuations (Leon, 2015), despite the unlikeliness of contemporaneous causality between our firm- and macro-level variables.

18 We thank two anonymous referees for this advice. Same check on the different size groups of sample firms is also performed for other hypotheses sets. Firms with less than 50 employees, and turnover less than 10 million euros, are regarded as ‘Micro and Small’ sized. Other firms with less than 250 employees, and turnover less than 50 million euros, are regarded as ‘Medium’ sized.

19 The first-difference models are estimated by OLS estimator with robust standard errors. The dynamic panel data models are estimated by ‘two-step system GMM estimator’ (Arellano and Bover 1995; Blundell and Bond 1998). In the dynamic model, we treat both lagged dependent variable and main independent variables as endogenous, and the lag length for instrument variables is decided based on the best results of the overidentifying restriction tests for validity and error terms serial correlations tests. We prefer System-GMM over ‘Difference GMM’ (Arellano and Bond 1991) because it is criticised that lagged levels are often rather poor instruments, especially if the variables are close to a random walk (Baum 2006). However, the System-GMM includes both lagged levels and lagged difference in an equation system. Moreover, Roodman (2009) shows that 2-stage estimator is asymptotically more efficient than 1-stage estimator. Same approaches (1st difference and dynamic models) apply to hypotheses set 2. We thank an anonymous referee for suggesting these approaches.

20 We also use short-term bank credit usage (variable name: ST Bank credit, Models 3 and 4) because trade credit is naturally in short-term.

21 See Han, Zhang, and Greene (2017) and Wang, Han, and Huang (2020) for a review on ‘Market Power Hypothesis’ and ‘Information-based Hypothesis’.

22 The relation between trade credit investment (debtors) and bank credit could also be in reverse direction as Ferrando and Mulier (2013) suggest that banks may relieve lending standards if the accounts receivable can be used as collateral. We therefore reverse the bank credit ratio and trade credit ratio of Models 9–11 in Table 5 (not reported) and the negative association remains.

23 These results are not reported for space reason but are available on request from the authors. See Whited and Wu (2006), Hadlock and Pierce (2010) and Mulier, Schoors, and Merlevede (2016) for a review of the WW, HP and ASCL index.

24 However, Baum (2006) states that time fixed-effects should be removed if macroeconomic factors are controlled for because those factors do not vary across firms.

25 These additional robustness check outputs are not reported for space reason but available on request from the authors.

26 Estimations with random-effect maximum-likelihood estimator are also performed but not presented for space reason. All results support the baseline findings and are available from authors on request.

27 Empirical outputs are not reported for space reason, but available from authors on request.

28 Return on equity is calculated as profit (loss) before interest and tax divided by shareholders fund (Amadeus database), in real numbers (not percentage).

29 Added value is the sum of profit (loss) for the period and minority interest, taxation, cost of employees, depreciation and interest paid. All variables are adjusted by GDP deflator (see Ferrando and Mulier 2013).

30 Return on sales is calculated as profit (loss) before interest and tax divided by total sales, in real numbers (not percentage)

31 ROA growth rate is defined as ROA’s difference in two consecutive years divided by the average ROA of these two years.

32 Results are not reported for space-saving purpose but available from the authors on request.

33 Sales volatility is measured by dividing the standard deviation of sales by its mean value over a three-year period.

34 Accounts receivable times 360, divided by operating revenue.

35 Despite the small coefficient, the result is economically significant because the ROA is measured in real number.

36 We thank an anonymous referee for raising this issue.

37 These additional tests’ results are not perfectly consistent with the ones presented, but in general provide similar findings. As there are 6 additional models for these tests, we do not present them for space reason.

38 Such firms could be those which are more informationally opaque, less capable in providing collaterals or borrowing from banks with higher market power.

39 As discussed in Chui, Kwok, and Zhou (2016), under the absence of loan-level data, the cost of bank credit at firm-level is implicitly proxied by the ratio of financial expenses to average total bank credit. The cost of debt ratios interacted with trade creditors are not statistically significant in the regression analysis and results available from the authors on request.

40 These additional test outputs are not reported for space reason, but available from authors on request.

41 The most common goodness-of-fit measure for probit/logit model – Pseudo R², is not available under panel data setting, thus the calculation follows Stata guidance that it compares the log-likelihood value of a model with the same model but all variables besides constant are excluded. This is based on the principal that in probit/logit panel data settings, and log-likelihood should be 0 if the model is perfectly fit.

42 Fixed-effect panel probit estimator is not possible unless time series is close to infinity. See more from Lancaster (2000).

43 The interpretation of ‘40%’ could be exaggerated because the variable Z-Binary treats observations that are originally defined as in the grey zone either distressed or not distressed.