Ceasing Export Activities: A Dynamic Analysis of Pre-Exit Financial and Internationalization Predictors

Abstract

This article aims to find out if pre-exit financial (FP) and internationalization (IP) performance indicators can be used for predicting full de-internationalization (ceasing all export activities; CE). To achieve that, a theoretical concept focusing on the behavior of these predictors is built, and three research questions are postulated. Full de-internationalization is an under-researched topic in international business studies, while quantitative studies focusing on its predictors are especially rare. This study fills both gaps by providing population-level evidence for the theoretical concept. The dataset is composed of Estonian exporters that ceased or continued exporting in 2010–2022. IP variables focus on export scale, intensity and scope, while FP variables focus on liquidity, solvency, profitability and revenue-creation capability. The variables cover the timespan of three (pre-exit) years. To outline the significance of predictors and accuracies in the whole population and for different types of exporters, initially, logistic regression is applied, after which the prediction models are also composed with neural networks. Before CE, IP is in a gradual decline, while the bulk of this decline is concentrated shortly before the exit. Before CE, exporters are constantly liquidity- and solvency-constrained, while the problems with revenue creation and profitability are much shorter-lived. That population-level behavior is subject to substantial variation for different types of exporters, especially regarding FP. Prediction models incorporating the full set of variables achieve high accuracy; however, predictive performance declines as the time to exit increases and varies across exporter types. IP variables are more beneficial for predicting CE. The latter also serve as the main practical implications of the paper.

1. Introduction

Firms’ internationalization has been actively studied since the 1970s. Their initial and subsequent foreign market entries have attracted considerable research attention [1,2], but internationalization processes do not only encompass increasing international involvement, although most studies have focused on growth [3,4,5]. De-internationalization—defined as “any voluntary or forced actions that reduce a company’s engagement in or exposure to current cross-border activities” [6]—has been studied less [7,8,9,10]. Still, without paying enough attention to de-internationalization, it is not possible to completely understand internationalization [11]. Given the current business environment, making strategic changes in firms’ local or foreign activities is almost inevitable [12,13,14]. Understanding de-internationalization has been considered crucial for succeeding in international business [4,15,16].

While partial de-internationalization—reducing or ceasing some foreign activities (temporarily) but not all—is very common among firms [17,18,19]; many firms also experience complete or full de-internationalization: they cease all international activities [6,20,21] without intending to resume them again [4,5]. While several authors have stated that complete de-internationalization can occur due to various internal or external factors [11,22,23,24,25,26,27], predicting it remains rare: it is unclear yet if different export variables or financial indicators can be used to do so with a high prediction accuracy [28].

This article aims to fill this research gap by finding out if pre-exit financial and internationalization performance indicators can be used for predicting full de-internationalization (ceasing all export activities) (Full de-internationalization can also take other forms—for instance, firms can divest or de-franchise [15,29], but these de-internationalization forms are not studied in this paper). Due to the lack of a comprehensive theory of de-internationalization predictors, the following literature review section synthesizes a theoretical framework of their predictors, focusing on micro-, small-, and medium-sized enterprises (i.e., MSMEs) based on different streams of literature. The Estonian population of firms subject to (non-exit) from export markets, variables portraying the theoretical concept and methods are explained in the subsequent study design section. Then, the following results and discussion section, divided into specific subsections, answers the three postulated research questions and outlines their theoretical importance. The conclusion section classically includes practical implications, limitations and suggestions for follow-up studies.

2. Literature Review and Research Questions

2.1. Firm Failure and Its Process

As in the international business or economics literature, no core theory is focused on the financial dynamics before ceasing all exporting activities (later “CE”), business failure literature can be applied to set the respective theoretical expectations. In the literature, since the earliest failure prediction studies (e.g., [30,31]), a steady decline in the financial performance of failing firms has been documented, indicated by serious and deepening issues with liquidity, solvency and profitability. Revenue creation capability (in Altman [31], it was defined as the sales-generating ability of the firm’s assets) was not found to be significant in Altman [31], due to potential industrial differences. In addition, a later study by Moulton et al. [32] indicated that a sales drop was characteristic of approximately 57% of bankrupting firms, still with a relatively modest average decrease. According to Moulton et al. [32], the remaining 43% of failing firms with increasing sales are divided between those fighting aggressively for market share and those losing control of their growth, and are more generally subject to (over)expansion issues [33].

Later-emerging literature discussing failure processes [34,35] has largely documented the same phenomena, although firms can differ in respect to the nature and speed of financial decline. Generally, the ideas proposed in the piloting studies about firm failure have remained valid to date regarding what the useful predictors are [36] or which processes exist [37], with one substantial disclaimer. Namely, the detailed results can be (largely) affected by firm characteristics such as size, age, and export behavior (e.g., [28,38]).

Another aspect to be considered is the definition of failure (see [39]), as ceasing exporting itself does not necessarily mean (financial) failure, and it definitely does not automatically equal the most terminal forms of becoming bankrupt and/or being liquidated [40,41,42]. For instance, according to Tangpong et al. [43], a timely de-internationalization can even stop financial performance deterioration, although it can also happen in some cases [16,44]. Logically, the lower the share of foreign sales in total sales, the smaller the effect of CE on financial performance. This is consistent with the general theoretical model of firms’ decline by Weitzel and Jonsson [45], indicating that larger performance declines are associated with more severe definitions of failure. In addition, CE could take different forms, based on how quickly a firm loses its export scope (foreign markets), scale (export sales volume) and intensity (share of export sales from total sales)—a matter largely overlooked in the extant literature [16]. Therefore, the financial dynamics of CE firms with varying export performance could be different in terms of whether and how quickly they change risk classes as portrayed in Korol [46].

2.2. Theoretical Framework of Performance Dynamics Before Ceasing Exporting

With the above literature in mind, the following theoretical narrative examines in more detail the potential expectations for pre-CE performance dynamics. When developing the theoretical framework, a MSME-centric approach is taken, as (very) large firms, especially multinationals, rarely retreat from foreign markets fully, and they are usually impacted by financial constraints differently.

While the usefulness of variables portraying liquidity, solvency, profitability and revenue-creation capability has remained uncontested for failure prediction [36,47], an open question in the literature is whether they are useful for CE prediction and how long before the event. The closest answer to this question was provided by Lukason and Vissak [28], who, based on an empirically validated theoretical concept, concluded that an average bankrupting European exporter was longitudinally subject to (severe) solvency and liquidity issues, while profitability and revenue creation problems evolved more gradually and were less significant. That finding can be linked to the financial constraints literature: as firms starting to export are usually associated with higher levels of liquidity and solvency (e.g., [44,48]), the cessation of exporting could, in turn, be linked to the loss of the respective advantage [49,50].

Although using data on certain market exits rather than full cessation of exporting, Askenazy et al. [51] showed that exiting firms are not only less liquid and solvent but also have more outstanding trade credit and more defaults on it. As solvency has been found to be a long-horizon predictor of failure [52], but also an important contributor to failure risk in different stages of the most common failure process [37], it might be expected to matter for ceasing export activities as well. However, in several studies, liquidity has not been found to be a useful predictor over a long period of time, differently from solvency (e.g., [37,52]), yet it still ranks as the second most useful predictor after solvency [28]. Therefore, it is assumed that firms subject to CE are less solvent and liquid when compared with their counterparts that continue to export.

Concerning low revenue creation capability and profitability, being widely acknowledged as potential causes of CE [53], studies about failure processes have indicated the former to be positioned earlier in the timeline than the latter [33,35]. A similar result has been documented about exporters, as Lukason and Vissak [28] indicated that the gaps between the revenue creation capability and profitability of (non-)surviving firms gradually increased, with the latter being more pivotal shortly before failure. This is consistent with Tangpong et al.’s [43] retrenchment-turnaround model, in which early geographical retrenchment, being logically associated with a drop in revenue, does not necessarily mean a drop in profitability. Therefore, directly contingent on how the average CE process looks like, closer to fully ceasing exporting (e.g., one to two years before CE), the difference in revenue creation could be visible between CE and non-CE firms. In turn, the emergent gap in profitability between CE and non-CE firms could be visible only very shortly before quitting exporting.

With respect to export scale, scope and intensity, it is possible to assume that CE firms witness lower levels than non-CE firms, as less committed exporters have been shown to witness more fluctuations and decline in exporting [54]. As less committed exporters have also been characterized by poorer financial performance [28], they are more likely to abandon (some) foreign markets through retrenchment to restore their earlier performance levels [43]. Rather than witnessing an abrupt CE, the majority of firms are likely to abandon exporting gradually, among others being subject to sunk costs made to start exporting and expectations of performance improvement at foreign market(s) [53]. These arguments lead to the conclusion that, with respect to export behavior, de-internationalization is preceded by a certain “inverse” of the Uppsala model [55], in which the export performance gap between CE and non-CE firms widens gradually.

The theoretical background and assumptions concerning financial and internationalization behavior before quitting exporting have been summarized in

Table 1. Theoretical framework of the study.

Performance Measure	Conceptual Background	Expectation of Behavior
Financial performance measures
Liquidity	Financial constraints in exporting, (exporters’) financial failure processes	Before ceasing or continuing exporting (T), from T-3 to T-1 significant differences exist in liquidity and solvency levels between CE and non-CE firms.
Solvency
Profitability	Retrenchment-turnaround model (exporters’) financial failure processes	Before ceasing or continuing exporting (T), significant differences exist in revenue creation levels starting from T-2 or T-1 and in profitability levels, if present, only in T-1 between CE and non-CE firms.
Revenue creation
Internationalization performance measures
Export scale	(Full and/or partial de-) internationalization processes	Before ceasing or continuing exporting (T), from T-3 to T-1 significant differences exist in export scale, intensity and scope levels between CE and non-CE firms.
Export intensity
Export scope

Note: CE denotes firms that cease exporting; non-CE denotes firms that continue exporting. In Section 2.2, the conceptual background noted in this table has been discussed in the same sequence, i.e., first focusing on liquidity and solvency, then on profitability and revenue creation, and finally, on internationalization performance. The expectations concerning behavior are directly emergent from the narrative provided in Section 2.2.

and are consolidated under the first research question of the study (RQ1). The framework in Table 1 builds on various streams of literature, while the latter have usually focused on other contexts (e.g., bankruptcy as a type of failure) than this study. Based on earlier studies (e.g., [28,43]), we opt for a three-period model for the empirical analysis. The MSME failure prediction studies (e.g., [38,56]), where more severe definitions (such as default or bankruptcy) have been applied on the population level, have indicated that financial performance variables do not possess high accuracy even shortly before the event. Therefore, it is not reasonable to opt for a lengthy horizon in the case of MSMEs, especially when a less severe failure definition is applied. The same internationalization and financial performance variables have been applied in Lukason and Vissak [57] to predict the success (or failure) of a firm after obtaining an export grant, but besides a different context, that study did not apply a dynamic approach in the case of predictors.

RQ1. 

How do internationalization (IP) and financial performance (FP) predict cessation of export activities?

2.3. Performance Dynamics of Different Types of Exporters

While the above arguments concerned the FP and IP of the CE and non-CE firm population generally, large variation could exist for firms with different export scope, scale and intensity. For instance, if the firm only exports to one customer, then losing it would mean complete de-internationalization, not depending on the firm’s export share or total exports, while if it has several foreign customers, then losing one would mean that some foreign activities would still continue, therefore, certain scale and intensity would be sustained. Moreover, for some firms, exporting is just a minor activity [16]; thus, for them, a quicker full exit could be more probable than for committed exporters. Therefore, it is vital to validate the results obtained for the population of CE and non-CE firms for different types of exporters, which would also serve as a robustness test of the stability of the general result in sub-populations. Therefore, the second research question of the paper is set as follows. (It is fully exploratory in nature, meaning that the applicability of the theoretical framework created in the previous section is checked in the case of different types of exporters.)

RQ2. 

Does the significance of IP and FP vary in predicting the cessation of export activities for different types of exporters?

2.4. Prediction of Ceasing Export Activities

While several international business and economics studies have focused on the determinants of exporting cessation, most of them have not predicted the respective event. A few studies have predicted the bankruptcy of exporting firms. For example, Lukason and Laitinen [58] focused on French firms, while Lukason and Vissak [28] focused on five European countries (including Estonia, as does this study), obtaining a high classification accuracy (best area under the curve 0.927, equaling 85.7% accuracy with neural networks). Concerning the prediction of ceasing export activities, it might be assumed that if international exposure is very high, the prediction accuracy could be similar to that reported in Lukason and Vissak [28] because the firm is likely to be in severe difficulties or even go out of business. With the reduction in international exposure, especially for export intensity, the accuracy is likely to drop. Therefore, the last research question focuses on the prediction of CE. (It is fully exploratory in nature, as to the knowledge of the authors, there are no comparative predictive studies about CE available.)

RQ3. 

How accurately are IP and FP able to predict the cessation of export activities?

3. Study Design

The dataset is composed of all exiting and non-exiting Estonian exporters (similarly to Hiller et al. [42], all firms with positive export sales are considered exporters) during 2010–2022, with the following restrictions. All firms are subject to a minimum sales revenue level of EUR 40,000 per year, leading to the exclusion of very small microfirms. Before exiting, each firm had to be operational in foreign markets for at least three consecutive years, and the cessation of exporting had to last for at least two years, for the de-internationalization to be substantial enough. The latter condition is similar to Askenazy et al. [51], Chen et al. [50], Beņkovskis et al. [59], Atabek Demirhan [60], Harris and Li [61], Ilmakunnas and Nurmi [44], and Inui et al. [21], but the former three studies focused on ceasing exports to a specific market. (However, for some authors, not exporting for one year was enough to classify a firm as an “exiter” [62,63]. On the other hand, Hiller et al. [42] required three years without export activities.) At least three years of exporting is necessary to exclude intermittent exporters [64,65]. Two years might not be sufficient, as the export revenue per single order can be distributed between two financial years. Therefore, the last three years with export revenues before having only domestic sales originate from 2012–2020, leading to 1363 exit episodes, while the frequency of observations from each year remains in the interval of 100–200. Therefore, the episodes are quite evenly distributed over the lengthy time period. In the case of export exits, we focus on the last three years with export revenues.

As the comparative group on non-exiting firms, we consider all Estonian firms that during the period 2016–2022 were exporting during all years and match the sales revenue criteria, but in their case, we use the following three-year sequences: 2016–2018, 2017–2019, and 2018–2020. Therefore, in the case of each firm, besides the three-year period, two additional years are used to check that it is not subject to CE. Multiple three-year periods are used to avoid the single-period bias. This totals 11,290 non-exit episodes.

Based on the last sales revenue figure, the population of firms divides as 0.8% large, 4.6% medium, 17.0% small and 77.6% micro-sized entities according to the European Union’s classification rules. Based on the last year’s total assets criteria, the proportion of large firms is even smaller (0.6%), while a firm might be large-sized based on one criterion and medium-sized based on another criterion, and thus, they are not excluded from the analysis. The latter can also vary by the year analyzed. Therefore, the empirical analysis has a clear MSME-centric focus, and the results might therefore not necessarily be applicable in the large firms’ segment.

The dependent variable of this study is export exit (EXIT, see

Table 2. Coding and content of variables.

Variable Code (and Domain)	Variable Content
Dependent variable
EXIT (export exit)	1 if the firm had no export revenues for two years after at least three year consecutive exporting, and 0 otherwise (i.e., remaining exporting)
Export performance: independent variables
EXPORTLN (export scale)	Natural logarithm of export revenue
PEXPORT (export intensity)	Share of export sales from total sales
MEXPORT (export scope)	Number of export markets
Financial performance: independent variables
NWCTA (liquidity)	(Current assets—current liabilities) to total assets
NITA (profitability)	Net income to total assets
TETA (solvency)	Total equity to total assets
ORTA (revenue creation)	Sales revenue to total assets

). For both firm groups, independent variables are coded for the three-year sequences (from T-3 to T-1), while in the group of exiting firms, year T reflects the first year and T+1 the second year with no more export revenues, while in the other group, firms remain exporting in years T and T+1. Based on the theoretical concept (see Table 1), variables portray firms’ internationalization (IP) and financial performance (FP) during the three years analyzed (see Table 2), while similar variables have been applied in the extant research. For instance, several authors have used export scale (volume) [50,62], export intensity (share) [28,63,66] and/or export scope (markets) [66,67] to measure export performance [68]. As financial performance variables, classical financial ratios used in failure prediction models are applied, while the chosen four domains (liquidity, profitability, solvency, and revenue creation capability) are all represented among the ten best variables in Bellovary et al. [47], often with the same formulas. In addition, similar ratios have been applied in earlier export exit research (e.g., [51,60,61]).

All the variables are coded for three years (from T-3 to T-1), with a respective marking in tables. Besides the variables reflecting the situation for each of those three years, dynamic variables are also applied in this study. These are calculated as changes of independent variables documented in Table 2 between years T-1 and T-3, with a formula value_T-1–value_T-3, based on Laitinen [69]. The only technical exception is the dynamic export revenue variable, in the case of which original values are subtracted, then the natural logarithm is applied on the absolute value and then multiplied by “−1” in the case the original change was negative. Such a dynamic context is marked with DELTA.

The analysis is divided into three consecutive stages to answer three research questions. In Stage 1, four logistic regressions (for T-3, T-2, T-1, DELTA) are composed with all seven variables from the respective period. These enable the obtaining of the general behavior of the respective variables in the population. A similar dynamic approach was applied in a predictive study by Iwanicz-Drozdowska et al. [52]. In the case of T-3, T-2 and T-1 models, a significant negative coefficient would indicate the existence of a performance gap between CE and non-CE firms, while the firms continuing to export perform better. The significant negative coefficient in the DELTA model would indicate the enlargement of the performance gap during the three years. Multicollinearity is not an issue in any of the models, as the average variance inflation factor (VIF) ranges from 1.50 to 1.67, while the maximum individual VIF in the models is 2.37. To avoid the bias from using a single method, the calculations in Stage 1 have been repeated with the probit regression.

In Stage 2, the population of CE and non-CE firms is divided into three sub-groups (types of exporters) in three different ways based on the firms’ export scale, intensity and scope. This is achieved for each of those three dimensions by dividing the population into three by quantiles based on the value of either export revenue, export share from sales or number of markets. The respective two breakeven points for these variables are EUR 61,295 and EUR 385,349 export revenue, 12.3% and 76.4% export share, 1 and 3 markets (leaving all the respective figures to the lower quantile). These groups are respectively called small/medium/high export scale/intensity/scope, and firm frequencies for them can be found in Appendix A.

Then, analysis in each of those nine sub-groups is repeated for the four periods (T-3, T-2, T-1, DELTA), totaling 36 logistic regressions. Because in Stage 2, the most important aspect is the potential varying behavior of the respective variables, the results are presented not as full regressions, but instead, in the case of significant (also by distinguishing different levels of significance) variables, it is marked whether the coefficient is positive or negative. The latter provides evidence whether the higher values for the respective variables increase (positive coefficients) or decrease (negative coefficients) the likelihood of export exit in different firm sub-groups. To enhance the analysis concerning the behavior of seven performance measures and their match with the theoretical concept in Table 1, an additional table summarizing the latter has been presented in Section 4.

In the final stage, the potential of variables from three periods (T-3, T-2, and T-1) in predicting export exit is assessed. For that purpose, for the firm population and for the nine sub-groups separately, prediction models using the logistic regression are composed with the following timing logic reflecting the availability of the respective information. First, only the seven variables from T-3 are applied, then from T-3 and T-2, and finally, from all three periods. Therefore, these prediction models include 7, 14 and 21 variables, respectively. This enables us to outline how the prediction accuracy increases nearer to the year when a firm has fully ceased exporting. Each prediction model, unlike the earlier regressions outlining the significance of variables, is composed by weighting the populations of exiting and non-exiting firms equally in logistic regression. Such provision of equal weights to both groups is important, as otherwise, the prediction models are likely to prefer the majority group (non-exiting firms), leaving the characteristics of exiting firms largely irrelevant. As an example, without this weighting option, the prediction accuracy in the exiting firms’ group can be very low, and therefore, it is not technically correct to consider it as the prediction of ceasing exporting activities. The latter is a usual strategy implemented in earlier failure prediction research (e.g., exactly the same approach was administered in [38]). For potential practical applicability purposes, logistic regression-based prediction models are also presented in the Appendix F.

In order to find out the comparative performance of machine learning in prediction, for the population and nine sub-groups, models inclusive of 21 variables have also been composed with neural networks (NN). A two-layered NN with sigmoid function in both layers and automatic computation of units in both layers is applied, while each time, frequencies of (non-)exit firms have been equalized with the synthetic minority oversampling technique. The latter is a substitute for weighting in machine learning. A batch type of training is used, and the optimization algorithm is scaled conjugate gradient. Variables have been standardized before application in NN. A proportion of 70% of the firms are used for training the models, while the accuracies are reported from the test set composed of the remaining 30% of firms. For the population and each sub-group, three NN models are composed, as the results of NN can vary for each run. While the highest accuracy has been reported over these three runs, the additional analysis of variables’ importance in NN has been averaged over the three runs, as even for the same accuracy, different variables can have varying importance.

Despite being a single-country study, earlier research has indicated that the financial decline of (non-)failing exporters is relatively similar for firms from small open economies such as Estonia (see e.g., [28]). For instance, the study by Lukason and Vissak [28] concerning bankruptcies of exporters from different European countries indicates that the largest gap in revenue creation exists shortly before failure. Similarly, the study by Lukason and Laitinen [70] indicates that the financial failure processes of firms can be very similar in certain countries, even when comparing countries of different sizes from different European regions. Therefore, it is assumed that the results obtained in this study could be applicable in certain other small export-oriented European economies.

4. Results and Discussion

4.1. Internationalization and Financial Performance Variables as Predictors of Ceasing Export Activities (RQ1)

The four population-based logistic regressions (LR) documented in

Table 3. Logistic regressions for the population with the dependent variable EXIT and seven independent variables for different periods (T-3, T-2, T-1, DELTA).

Variable	T-3			T-2			T-1			DELTA
	B	p-Value	dy/dx	B	p-Value	dy/dx	B	p-Value	dy/dx	B	p-Value	dy/dx
EXPORTLN	−0.260	0.000	−0.0213	−0.295	0.000	−0.0233	−0.373	0.000	−0.0260	−0.028	0.000	−0.0025
PEXPORT	−0.727	0.000	−0.0595	−0.779	0.000	−0.0615	−1.110	0.000	−0.0774	−2.503	0.000	−0.2284
MEXPORT	−0.545	0.000	−0.0445	−0.576	0.000	−0.0455	−0.831	0.000	−0.0579	−0.027	0.082	−0.0025
NWCTA	−0.420	0.001	−0.0344	−0.490	0.000	−0.0387	−0.665	0.000	−0.0464	−0.342	0.061	−0.0312
NITA	0.454	0.002	0.0372	0.033	0.831	0.0026	−0.294	0.068	−0.0205	−0.236	0.052	−0.0216
TETA	−0.371	0.015	−0.0303	−0.469	0.002	−0.0370	−0.400	0.013	−0.0279	−0.054	0.806	−0.0049
ORTA	−0.015	0.378	−0.0012	−0.052	0.005	−0.0041	−0.074	0.000	−0.0051	−0.025	0.281	−0.0023
Constant	2.440	0.000		3.072	0.000		4.264	0.000		−2.272	0.000
Pseudo-R2	0.253			0.289			0.403			0.081

Note: B—coefficient of the variable in the model, dy/dx—average marginal effect of the variable.

lead to several generalizations. The variables of internationalization (IP) and financial performance (FP) for the three periods match the theoretical concept postulated in Table 1.

All IP variables are highly significant for all three periods (from T-3 to T-1), indicating the existence of a performance gap between CE and non-CE firms. While non-CE firms indicate higher IP, the DELTA of these variables indicates the enlargement of the performance gap between CE and non-CE firms. As the marginal effects are the largest for the period T-1, the performance gap has widened, especially before the CE firms quit exporting. This is also confirmed by the descriptive statistics documented in Appendix B. Namely, based on the median and mean values of IP variables, the largest performance gap is visible in T-1. While non-CE firms largely sustain the same level of export performance during the period analyzed, a substantial drop is visible for CE firms.

The FP variables indicate a constant performance gap with respect to liquidity and solvency, the non-CE firms being characterized by larger values for these measures. While for liquidity, the performance gap becomes wider during the viewed period, this does not apply for solvency. Liquidity can be considered more important for sustaining everyday operations, and therefore, its deterioration is a crucial negative signal. As theorized in Table 1, a profitability gap exists only for T-1, the CE firms obtaining a lower profitability. It is also noteworthy that in T-3, CE firms are even more profitable than their counterparts. Revenue creation gap, according to expectations, exists for T-2 and T-1, while it does not significantly widen over the three-year period. The additional probit regression results presented in Appendix C confirm the findings obtained with logistic regression, and therefore, the results are not subject to a single-method bias.

These results corroborate the theoretical expectation that CE firms are constantly underperforming their non-CE counterparts in respect to export performance [16], while a gradual reduction in that performance is visible, with a bulk of it being concentrated in the period T-1. This is consistent with the gradual decline logic outlined in Weitzel and Jonsson [45]. CE firms are constantly constrained in terms of liquidity and solvency, serving as a potential main financial driver of their export exit. Therefore, export exit is similar to non-entry subject to financial constraints [48,49], while, likewise with earlier exporters’ failure studies, liquidity and solvency serve as useful predictors [28]. Logically, as the export revenues start to decline, this starts to contribute to the emergence of a revenue creation gap, but with a distinct time lag. Finally, having lost much of their export revenues, CE firms are also subject to reduced profitability shortly before ceasing export activities. The appearance of profitability issues after the problem with revenue creation is consistent with earlier studies about failure processes [33,35]. Therefore, population-level evidence was provided that an average full de-internationalization is a gradual process linked with constant liquidity and solvency constraints, with issues concerning revenue creation and profitability emerging sequentially and later towards an imminent exit. This finding contributes to the extant literature, which so far lacks evidence in respect to how the average de-internationalization process occurs in terms of the pre-exit dynamics of export scale, intensity and scope. Moreover, the extant knowledge about financial constraints hindering export entry is transferable to the exit context, as firms ceasing exporting are also constrained by liquidity and solvency [51].

4.2. Variation in the Behavior of Internationalization and Financial Performance Variables by Exporter Types (RQ2)

The analysis is followed by studying the behavior of IP and FP by different exporter types (see Appendix D for the results of the respective logistic regressions). These types have been created by dividing the population into three groups, either based on the export scale (volume of exports in EUR), intensity (share of export from total sales) or scope (number of foreign markets) terciles. The results reveal numerous differences when compared with the general theoretical concept (Table 1) and its empirical validation (Table 3).

To facilitate the discussion concerning the behavior of variables and their correspondence to the theoretical framework, the results have been summarized in

Table 4. The behavior of internationalization (IP) and financial performance (FP) measures in sub-groups of exporters and their correspondence to theory.

Sub-Group of an Exporter	Performance Measure							IP Matches with Theory (Max = 9)	FP Matches with Theory (Max = 12)	Total Matches with Theory (Max = 21)
	Export Scale	Export Intensity	Export Scope	Liquidity	Profitability	Solvency	Revenue Creation
Small export sales	+W		+W	+W		+	+	6	11	17
Medium export sales	+W	+	+W	+				9	9	18
Large export sales	+W	+W	+W					8	5	13
Small export share	+W		+			+		9	9	18
Medium export share	+W		+	+W				7	8	16
Large export share	+W	+W	+W					9	7	16
Small number of markets	+W	+W	+	+		+	+W	8	10	18
Medium number of markets	+W		+					6	5	11
Large number of markets	+W		+W		+			5	6	11
Average match with theory								82.7%	64.8%	73.0%

Note: For performance measures from export scale to revenue creation, “+” indicates the existence of a performance gap in favor of non-exiting firms for at least two years before the full exit year, while W indicates the widening of the respective performance gap. The matches with theory for IP and FP indicate how many years for the three internationalization (a total of 9 observations) and four financial (12) performance measures match the theoretical concept in Table 1, while in the final column, these matches have been summed. The last row indicates the average match with theory over nine types of exporters.

. Before CE, for different types of exporters, the performance gap is much more frequently visible concerning IP than FP variables (see columns about performance measures in Table 4). Moreover, the performance gap concerning IP usually widens during the three-year period (see Appendix D), indicating the retrenchment of international activities before CE. Therefore, it can be concluded that generally different types of exporters subject to CE have a poorer IP than non-CE firms [16], and this difference becomes more substantial before export exit. As summarized earlier on the population level, the results show that even for different types of exporters, de-internationalization does not happen “overnight”—it is a more gradual process. Although not fully universally valid, such a gradual decline is more visible in the case of larger international exposure. Still, CE firms are more subject to loss of their export volume and markets, rather than intensity, the latter being significant only in four sub-groups out of nine. The other two measures (export volume and markets) are significant for all types of exporters. In turn, performance gap with respect to FP is rather infrequent, and it usually does not widen during the three-year period. Generally, the results indicate that the FP gap is more subject to lower international exposure. As expected based on the theoretical concept in Table 1, for different types of exporters, the performance gap between CE and non-CE firms more often emerges from liquidity and solvency [49,50].

Table 4 (the last three columns) analyzes the correspondence of the IP and FP behavior in relation to the theoretical concept in Table 1. To achieve that, for each exporter type, it has been outlined on how many occasions the IP and FP variables from T-3 to T-1 behave exactly as postulated in Table 1, while in the last column of Table 4, the summed behavior of IP and FP has been outlined. As already emergent from the earlier narrative, IP measures indicate a better match with theory. Over nine types of exporters, the average correspondence to the theoretical concept is 82.7%, with no remarkable differences for pooled types portraying small, medium and large international exposure. In turn, on average, around two-thirds (64.8%) of exporter types match the theorized FP behavior. With the increase in international exposure, the correspondence of FP to the theoretical concept decreases. On average, for firm types with large export scale, intensity and scope, only half of the FP measures from T-3 to T-1 behave according to the theoretical concept. Finally, the average match with the theoretical concept over exporter types is 73.0%, with types subject to lower international exposure indicating a better match.

As the main theoretical implication, it can be concluded that different types of exporters generally tend to follow a similar pathway of de-internationalization with respect to the reduction of exporting. This is supported by the fact that, although not fully universal, export intensity loss is more pronounced in the case of firms with larger international exposure (specifically, with larger export volume and share). In turn, the presence of a financial performance gap and its enlargement can vary more between different exporter types. There are some exporter types (especially those with larger international exposure), in the case of which CE and non-CE firms (almost) do not differ financially. This was not surprising, as according to Lukason and Vissak [71], in Estonia, different exporter types did not significantly differ in terms of financial performance.Still, the opposite behavior is also present, especially for those with smaller international exposure.

Although it is counterintuitive that CE firms with exporting as their main source of revenue witness less financial hardship, there is a potential logical explanation for it. CE firms that export only as a “side-business” and are underperforming financially might choose to retreat from foreign markets [59] by means of swift operational decisions in order to restore their financial position by focusing only on domestic activities. In turn, for CE firms in cases where exporting is an important or primary activity, this is a well-grounded strategic decision [16], and they follow a step-by-step international retrenchment to avoid severe financial consequences. This is supported by the fact that for firm types with medium or large international exposure, the revenue creation gap did not exist, being an indication that they are able to exchange their foreign sources of income quite smoothly against domestic ones.

4.3. Prediction of Ceasing Exporting with Internationalization and Financial Performance Variables (RQ3)

The export exit prediction models are composed of all variables known at the period when the prediction model would be implemented. Certain general tendencies are visible from

Table 5. Prediction accuracies of logistic regression models in the population and in different sub-groups using seven independent variables from different periods, and comparative accuracy of neural networks using data from all three periods.

Firm (Sub-)Group		Period Used in Logistic Regression Prediction			Period Used in Neural Networks Prediction
		T-3	T-3 and T-2	T-3 and T-2 and T-1	T-3 and T-2 and T-1
All firms		75.7%	77.6%	82.1%	84.1%
Export volume	Small export sales	66.0%	68.1%	73.8%	75.1%
	Medium export sales	65.7%	72.3%	82.4%	89.3%
	Large export sales	68.4%	76.0%	89.4%	97.3%
Export intensity	Small export share	71.9%	75.6%	78.0%	81.9%
	Medium export share	75.0%	78.3%	82.5%	90.1%
	Large export share	67.6%	75.9%	84.2%	95.1%
Export scope	Small number of markets	68.6%	70.1%	76.9%	80.6%
	Medium number of markets	67.0%	72.1%	79.9%	89.4%
	Large number of markets	72.0%	77.4%	87.0%	96.9%

Note: All independent variables from Table 2 are applied from the periods noted in the column heading.

documenting the prediction accuracies. First, the prediction accuracy increases systematically nearer to the exit, making the latter more easily predictable. Second, generally, the increase in prediction accuracy in percentage points between periods T-3 and T-2 combined versus T-3 and T-2 and T-1 combined is (much) larger than that from the comparison of the earlier two periods. Therefore, the prediction models indicate that more substantial IP and FP drops occur between the years T-2 and T-1, leading to a considerably increased performance gap between (non-)exiting firms. Third, the larger the export volume, intensity or scope, the higher the prediction accuracy, except for some minor exceptions for the period T-3. This tendency is especially visible in the case of export volume sub-groups, as the exit of those with the smallest export volume can be predicted with a 73.8% accuracy, while for those with the largest export volume it is 89.4%. In conjunction with results presented in Table 4, this points to the fact that the larger the export volume, intensity and scope, the more likely that the international performance level and dynamics in the case of exiting firms diverge from those of non-exiting firms.

The best prediction accuracy of 89.4%, corresponding to an area under the curve (AUC) of 0.960, can be considered excellent. Comparatively, in Altman et al.’s [38] all-European study focusing on insolvency prediction of MSMEs, which is usually situated further in the failure timeline, the general model provided an AUC of 0.771, while in Estonia it was 0.890. The obtained AUC also exceeds that obtained in Lukason and Vissak [28], predicting the bankruptcy of European exporting firms with financial ratios, indicating that in the case of exporting firms, the dynamics of their international behavior is a valuable supplement among classical failure predictors. This is additionally supported by the fact that, although not provided in Table 5, for the firm group with large export volume, the usage of only financial performance variables from three periods would provide an accuracy of 75.3%, while the usage of only internationalization performance variables 88.1%.

Based on the results reported in Table 5, using neural networks leads to a higher accuracy than logistic regression, while this tendency is especially visible in sub-groups with larger export engagement. For the sub-groups with large export sales, large export share and large number of markets, the prediction accuracies even exceed the 95% thresholds, the corresponding AUCs being close to the ideal. The additional analysis of variables’ importance in Appendix E indicates that while export scale and scope are the most useful predictors in neural networks (based on the average importance in sub-groups), export intensity’s importance is more similar to the financial dimensions of liquidity, solvency and profitability. Each of the latter is characterized by one predictor among the eleven best ones, while revenue creation is the only one not positioned there. With the exception of profitability, this largely reconfirms what was provided in Table 4 earlier.

This leads to a summarizing conclusion that export exit is accurately predictable, and that the result is mostly subject to using internationalization performance variables, not the classical financial ratios as performance measures. In a different context of being successful or unsuccessful after receiving an export grant, Lukason and Vissak [57] found a similar tradeoff between these two types of variables in a single-period setting.

4.4. Main Findings of the Study and Their Theoretical Contribution

In this section, the main findings of the study are generalized, and their main contributions to the literature are discussed.

4.4.1. Process of Ceasing Export Activities

Ceasing exporting (CE) in the MSME sector has two distinct features. First, CE firms are less exposed to international trade, as expressed through having fewer foreign markets, a lower export share from sales and a smaller volume of exporting. Such a discrepancy especially widens shortly before CE. Financially, CE is associated with such firms being constantly more constrained by liquidity and solvency. In turn, the performance gaps with respect to revenue creation and especially profitability for such firms are more short-lived. The latter relationships in mind, the study contributes to the extant research by outlining the typical process of ceasing exporting. To date, the exact attributes of such processes have been absent from the extant literature, although based on the theoretical framework created by merging different streams of literature, an array of these attributes was assumed.

4.4.2. Variation in the Process by Exporter Types

Different types of exporters subject to CE are more universally similar in respect to the nature of their decrease in internationalization performance. The few emergent differences in the process originate solely from the dynamics of export intensity (i.e., export share from total sales). This can be logically explained by the fact that the latter indicator, unlike the number of foreign markets and export volume, is the only one also affected by the volume of domestic sales. In turn, different types of exporters subject to CE are characterized by largely varying financial processes. Therefore, while the study confirms a certain theoretically motivated financial process for an average CE firm, it cannot be generalized over different types of exporters. To date, such evidence by types of exporters was absent from the extant literature.

4.4.3. Predictability of Ceasing Exporting

Ceasing export activities is predictable, but as logically derived from the respective cessation process, with a higher accuracy closer to the event. The event is easier to predict in the case of more committed exporters. Generally, the internationalization performance variables are more useful predictors than those reflecting financial performance. This is corroborated by the lower importance of the latter variables in machine learning models, but also their lower significance in statistical models. The findings complement the large body of literature on business failure prediction by indicating that ceasing exporting, which may be considered a failure, is similarly well predictable. However, the latter is subject to a substantial disclaimer, meaning that, unlike with extant business failure literature, financial performance measures are not as good predictors when compared with the domain-specific variables portraying internationalization performance.

5. Conclusions

This article contributed to the literature by validating the usefulness of internationalization and financial performance predictors for forecasting cessation of export activities based on a theoretical concept. The results on the Estonian firm population show that a performance gap exists among exiting and non-exiting firms with respect to export scale, intensity and scope during the three years before the exit, while that gap is widening closer to the exit. Exiting firms are more constrained by liquidity and solvency throughout that period, while their revenue creation capability and profitability become weaker with a certain time lag. Export exit is predictable with high accuracy, especially closer to the event.

The study leads to practical implications for various stakeholders. Export support agencies and creditors financing exporting firms can account for the fact that before ceasing exporting, international engagement is likely to reduce, i.e., withdrawal from foreign markets does not happen overnight. In addition, firms in lack of equity or liquid assets are more likely to withdraw. Therefore, when such signals are apparent, funding of such firms should be considered with higher caution. For instance, providing export support or loans to companies at risk of withdrawal from exporting will question the rationality of allocating such public funds to them. Ideally, the recipients of export-related public support should not only sustain their international engagement but also be able to increase it.

For private creditors, foreseeing export withdrawals has severalfold importance, as often these mean failed investments and the need to find additional markets for vacant production capacity, both of which can increase the risk of a company at least in the short-term. For practical purposes, Appendix F includes the weighted logistic regressions, which can easily be applied to calculate the risk of withdrawal. Still, knowing the useful predictors in a longer horizon and the potential benefits of machine learning methods based on the results of this study, these stakeholders have certain guidelines to build in-house predictive tools as well.

There are also important takeaways for the managers of exporting firms. Namely, as activities at foreign markets are usually concerned with additional investments and costs, the decision to engage in exporting should be accompanied by careful financial planning. This especially concerns the availability of slack financial resources, because the paper demonstrated the foreign retreat to be associated with more limited liquidity and solvency of a firm. Therefore, firms should potentially hold reasonable financial buffers or at least have easy access to additional financial resources when aiming to remain in foreign markets. As the average export cessation process concerns a gradual retreat from a single foreign market, such companies seem to be too concentrated on specific choices, and potentially, instead of monitoring that foreign environment, are stuck in letting things happen rather than managing them. Because of that, for instance, the loss of one or a few clients might come as a surprise without providing short-term alternatives in a certain market. The potential solution would be a gradual diversification of the foreign market portfolio, while obviously, the costs and benefits of that action should be separately assessed.

The paper’s main limitation to be addressed in future studies is that the developed theoretical concept can be extended by means of performance measures and timespan. While the applied framework is likely to be valid for MSMEs in other similar countries, which requires further testing, and based on the concept developed in Tangpong et al. [43], it is assumed that for large and listed firms, a lengthier time frame would be beneficial. In addition, a longer time frame could also be applied for this study, as it would especially enable us to understand how the predictors behave in the case of temporary de-internationalization, permanent de-internationalization and ceasing operations fully. In terms of performance measures, it would be interesting to understand how international withdrawal appears marketwise. While an average analyzed full de-internationalizer had a single foreign market, scholars could study whether retreats differ in respect to how far the market is, or even, if such information is accessible, whether differences exist for varying market-product combinations. With respect to financial performance, as the applied measures serve as usual inputs to probabilistic failure prediction models, a possibility for the paper’s development is to look at the failure probability dynamics also after the cessation of export activities. The latter would be especially topical for firms with high export intensity, as these are the likely candidates of going fully out of business, if timely reorganization is not implemented. While this study is unambiguously focused on predictive context, no conclusions can be drawn regarding what exactly causes the cessation of exporting. Therefore, future studies could be directed to finding out whether these predictors, and if so, in which exact sequence, also serve as causes of retreat from foreign markets. Finally, as Estonia is a small export-oriented open economy, studies could be conducted in countries with huge domestic markets, where returning home from abroad would probably be a much swifter option.