2.1 The McKinnon-Shaw hypothesis

The downfall of the Bretton Woods international monetary and financial system in the early 1970s triggered a deep process of deregulated global financial markets. Ronald I. McKinnon and Edward Shaw provided the theoretical justification and the rationale for worldwide financial markets integration.

The McKinnon-Shaw² hypothesis is composed of two elements, the financial repression (FR) and the financial liberalization (FL) theories: ‘If governments tax or otherwise distort their domestic capital markets, the economy is said to be “repressed”’ (^{McKinnon, 1993, p. 11}; cf. also ^{Fry, 1988, caps. 1, 2 and 3}). The main instruments for repressing the economy are:

These instruments of FR sooner or later impair the quality of investment, bring about low productivity, suboptimal economic growth, high unemployment, exchange rate and balance of payments instability and increasing foreign debt to bridge the gap between savings and investment.

The solution to FR is straightforward and leads to FL, according to McKinnon and Shaw. FL involves keeping positive and high interest rates, eliminating interest rate ceilings and predesignated credit lines, removing reserve requirements on bank deposits, and stabilizing domestic inflation to uncover “the true scarcity price of capital” (^{McKinnon, 1973}, ^{1993, p. 12}).

McKinnon (¹⁹⁹³) argues that in the transition from FR to FL countries must follow “the order of economic liberalization”, otherwise the process may fail to deliver a Pareto-optimum result. Not only that: governments must also avoid incurring the overborrowing syndrome during the process of FL. According to Mckinnon, Argentina, Chile and Uruguay failed to follow such a sequence and ended up experiencing financial crash back in the 1970s and early 1980s (^{McKinnon, p. 113, passim}).

The assumptions of the McKinnon-Shaw hypothesis make us believe that the road from FR to FL paves the way to overcome not only technological dependence but, above all, financial dependence. In this context, Mexico -among many other countries- liberalized financially starting in the late 1980s-early 1990s. Unfortunately, after more than three decades of financial liberalization not much seems to have changed in terms of improvements of the economic dynamic-balance of payments trade-off: the empirical evidence in Figure 1 reveals that most of the statistical observations fall in the fourth quadrant, where growth is characterized by increasing balance-of-payments constraints. In Figure 1, a balance-of-payments constraint must be thought of as a significant current account deficit as a share of GDP. The journey from FR to FL left Mexico with the same problem of financial dependence we encountered in the demise of the import substitution industrialization model (^{Vernengo, 2006}). Furthermore, it has been argued that financial liberalization has produced not only the truncation of industrialization, but above all instability, premature financialization and deindustrialization (^{Pérez Caldentey and Vernengo, 2021}).

Note: The fourth quadrant shows that higher GDP growth gives rise to higher current account deficits as a share of GDP.

Source: Authors’ estimations based on data from the National Institute of Statistics and Geography.

Figure 1 Output growth and Balance-of-Payments Trade-off 

2.2 The Thirlwall-Hussain model

Unless one shares the Panglossian outlook that current account disequilibria are immaterial (on the grounds that deficits just signal foreign investors’ preference for domestic assets), admittedly current account deficits reflect financial dependence, at least in the case of today’s Mexican economy. Current account disequilibria imply foreign debt and capital movements.

As discussed previously, Thirlwall (¹⁹⁷⁹) presented a model without considering financial flows or, in other words, if the long-term growth rate of output is approximated by equation (1) yb=x/π , then it is implicitly assumed that exports earnings (() are enough to cover the imports bill totally (( = 1). Actual growth experience may in fact deviate from equation (1) due to either real terms of trade fluctuations or capital flows letting a current account disequilibrium (^{Thirlwall and Hussain, 1982, p. 500}).

Historical growth experience of most -if not all- developing economies shows that these countries tend to accumulate in crescendo current account deficits through time (Figure 1 summarizes Mexico’s experience). These deficits must be financed somehow by capital inflows, so international financial markets allow developing economies to temporarily expand at rates faster than the speed predicted by equation (1). This is where financial dependence kicks in: “growth becomes constrained ultimately by the rate of growth of capital inflows, and, by itself, the simple growth rule enunciated would not be a good predictor of long run growth performance.” (op. cit., p. 501). Hence the Thirlwall-Hussain (¹⁹⁸²) extension of the original BPCG model accounting for the influence of financial flows on long-run economic growth³:

where c _t stands for the growth of real capital flows, ( and (1-() are the shares of the import bill financed by export earnings and real capital flows, respectively, x _t and ( are as before, the growth rate of exports and the income elasticity of imports, respectively, and yb* is the balance of payments long-run constrained growth rate of output with initial current account disequilibrium.

It can readily be seen from equation (2) that, ceteris paribus, capital inflows can cause yb* to deviate from yb , unless ct = 0 . Yet, initial disequilibrium (yb*>yb) requires ct > 0 to offset the proportion of the import bill not covered by export earnings. Thirlwall-Hussain (¹⁹⁸²) amends the original BPCG model by making explicit the role of financial transfers in the dynamics of developing open economies. An important point in this regard is that current account deficits are generally financed by debt-creating financial flows which must be repaid sooner or later.

All in all, while amending the canonical BPCG model the Thirlwall-Hussain extension appears to leave with us two extreme solutions: first, unless international financial markets will never become reluctant to finance an ever-increasing trade imbalance, which is not the case, deficit economies must generate a current account surplus through deflationary and recessionary regimes to make up for the accumulated financial hole. This implies that equation (1) must hold in the long-term, in which case ( = 1 and c _t = 0. Secondly, there is a sustainability issue⁴: a) since current account imbalances are financed by debt-generating financial flows, interest payments on accumulated liabilities should play a part in the model, and b) an upper bound to current account imbalances and, therefore, a limit to debt-accumulation should also be in place, in case the economy becomes fundamentally unstable and deteriorates into a Ponzi scheme.

3.1 Long-term analysis

Johansen’s cointegration tests are based on VAR models and are thus multivariate, which means that they can be useful to identify more than one cointegrating relationship. To perform such tests, we must first ensure that the benchmark and the alternative VAR models are congruent, which means that their residuals must not be plagued by serial correlation and heteroscedasticity. To eliminate or lessen such problems, we tested several lag structures for both VAR models. In the case of the benchmark model, nine lags for each variable in each equation was found to be the best possible lag length specification. In the case of the alternative VAR model, the more convenient lag length specification was ten lags for each variable in each equation. For the sake of brevity, we report the test results only for the benchmark VAR model.⁷ In the case of the alternative VAR model we will only present the outcome of the cointegration tests, the long-term equations, and the impulse response functions in appendixes 1, 2 and 3, respectively, for comparative purposes. Table 2 displays the outcome of the serial correlation Lagrange multiplier tests for the benchmark model:

The probability values corresponding the null hypothesis of absence of serial correlation suggest that VAR residuals are essentially free of serial correlation up to lag order 10. Now, table 3 presents the result of the White heteroscedasticity test.

Table 3 shows that VAR residuals are homoscedastic. Therefore, the benchmark VAR model is congruent.⁸ However, VAR residuals do not follow a normal distribution according to the Jarque-Bera normality test.⁹ Achieving multivariate normality is not an easy task in this case because the 2007-2009 global economic crisis and the COVID-19 pandemic brought about macroeconomic volatility and thus atypical observations. Nonetheless, Johansen (^{1995, p. 20}) relaxes the normality requirement to perform cointegration analysis, whereas Cheung and Lai (^{1993, p. 314}) acknowledge that such a requirement can be somewhat restrictive in econometric analysis.

For a set of variables to be cointegrated, the following requirements must be satisfied: 1) The variables must be nonstationary, 2) there must be at least one stationary linear combination involving those variables, and 3) such a linear combination must have a reasonable economic interpretation (^{Johansen, 1995}). When those requirements are satisfied, the Granger representation theorem (^{Engle and Granger, 1987}) states that equation (3), which is an unrestricted VAR model, can be rewritten as a vector error-correction (VEC) model of the following form:

where Π=∑i=1pΒi-I6 and I6 is a 6x6 identity matrix, given that the model includes six variables. Furthermore, Γi=-∑j=i+1pBj . In this manner, the p-order VAR model can be rewritten as a (p-1) VEC model. The VEC model is the restricted version of the VAR model, where the restrictions are given by the cointegrating relationships that the VEC model incorporates. Now, if such a VEC model exists, then, according to the Granger representation theorem, matrix Π must have a reduced rank and the following two conditions must be fulfilled:

Further, the number of cointegrating equations in θ'Zt-1 is equal to the rank of matrix Π . The intuition behind equation (4) is that the variables of the model share one or more long-term or equilibrium relationships, given by θ'Zt-1 . In fact, θ'Zt-1=VECTt-1 , where VECT stands for the vector of error-corrections terms. In this manner, VECTt-1 contains the set of cointegrating equations or long-term relationships. However, when innovations or disturbances (captured by vector ηt ) occur, then the variables in VECTt-1 temporarily deviate from equilibrium (i.e., an error is generated). Such deviations or “errors” are then corrected through an adjustment mechanism given by matrix α and by matrices Γi (i=1, 2,…, p-1), so that equilibrium is eventually restored (i.e., variables go back to their cointegrating relationship). In this manner, equation (4) can be rewritten as follows:

To determine the number of cointegrating equations, if any, Johansen cointegration tests must be performed. Such tests were carried out under the assumption that the VAR model includes a vector of constant terms (i.e., Wt in equations (3), (4) and (5) is a vector of intercepts), which is a standard assumption. Moreover, an intercept term is also included in each cointegrating equation, so that the long-term relationships are not forced to cross the origin (^{Patterson, 2000, p. 625}). The result of the cointegration tests for the benchmark model is depicted in Table 4:

Johansen cointegration tests are carried out sequentially and the conclusion regarding the number of cointegrating equations can be drawn as soon as we fail to reject the null hypothesis. Moreover, Johansen cointegration tests yield two test statistics: the trace statistics and the maximum eigenvalue statistics. The probability values corresponding to the trace statistics suggest that we have three cointegrating equations at the 5% significance level, whereas the probability values of the maximum eigenvalue statistics indicate that we have only one. When a contradiction like this occurs, Johansen’s (¹⁹⁹⁵) recommendation is to decide based on the number of cointegrating equations that is more consistent with economic theory. After estimating several normalizations of the cointegrating equations, we decided in favor the trace statistics, so we have the following three cointegrating equations:

Equations (6), (7) and (8) were normalized for the real output (Yt) , exports of goods and services (Xt) , and imports of goods and services (Mt) , respectively. This means that each of these variables cannot be included as regressors in any other cointegrating equations, which poses a severe limitation for specification purposes. Nonetheless, the estimated coefficient of FDI is statistically significant at the 1% level in the three long-term equations.¹⁰ Equation (6) shows that the long-term elasticity of real output with respect to FDI is 0.5559, which highlights the relevance of this variable to sustain economic growth. The rationale is that to be able to grow in the long run, the Mexican economy requires a wide range of imported intermediate inputs and FDI is largely responsible for financing their acquisition. Therefore, this evidence is consistent with the balance-of-payments restraint on economic growth explained by the Thirlwall-Hussain model. Such a restraint can be alleviated by a prominent and relatively stable source of foreign exchange like FDI.

Furthermore, equation (8) shows that FDI has a positive long-term impact on imports of goods and services. The long-term elasticity of imports with respect to FDI is 2.3697. Lastly, FDI translates into capital stock and technology for the recipient country. This, in turn, allows economies like Mexico to export goods and services. In this perspective, the long-term elasticity of exports with respect to FDI is 2.3849 (equation (7)). This means that the long-term impact of FDI on exports and imports is positive and very similar. Lastly, the estimated coefficients of the other regressors (the peso-dollar real exchange rate and the terms of trade) are not statistically significant in any of the three equations, whereas the t-statistics and probability values of the intercepts terms are not provided by the econometric software.¹¹

In appendix 1 we display the outcome of Johansen cointegration tests for the alternative model. In this case, the conclusion is the same: there are 3 cointegrating equations at the 5% significance level. In appendix 2, we can see the three cointegrating equations corresponding to the alternative VAR model. Broadly speaking, the estimated coefficient of US real GDP is positive and statistically significant at the 1% level in the three equations. In the growth equation, we can see the elasticity of Mexican output with respect to US output is 1.1594, which highlights the extent to which domestic growth is conditioned by foreign growth. The influence of US GDP on Mexican trade is reflected in the export and import equations as well. The elasticity of exports with respect to US output is 3.6468, whereas the elasticity of imports with respect to the same variable is 3.7345. Therefore, US economic activity abroad increases Mexican exports through an enhanced external demand for domestic goods and services. On the other hand, US economic activity raises Mexican imports, presumably by way of stimulating economic growth in Mexico. Lastly, in this case, the estimated coefficient linked to the real bilateral exchange rate is statistically significant in the three equations. In this perspective, real currency depreciation encourages exports, economic growth and even imports. The positive effect of the real exchange rate on imports is counterintuitive but can be explained by the fact that higher exports translate into higher imports due to processing trade (Hicks’s supermultiplier, ^{Perrotini and Vázquez, 2019}). Moreover, real currency depreciation fuels economic growth and thus elevates imports as the Mexican economy is highly dependent on imported capital goods, parts, and components.

3.2 Short-term sensitivity analysis

The VEC model previously estimated works to obtain the long-term equations. Unfortunately, the short-term coefficient matrix (α) indicates that real output (Yt) , exports (Xt) and imports (Mt) are weakly exogenous, given that none of the short-term coefficients (i.e., the coefficients of matrix α ) are statistically significant for the first three equations of the VEC model,¹² represented by equation (5). The weak exogeneity of these variables means that they pertain to the long-term equations but cannot be included in the impulse-response analysis stemming from the VEC model (^{Johansen, 1995}; ^{Patterson, 2000, pp. 674-676}).

In this context, we resort to a stationary VAR model of the form: ΔZtB=ΔYt,ΔXt,ΔMt,ΔQt, ΔFDIt,ΔTTt' . Such a model corresponds to the benchmark specification, except that the variables are in first differences. When this model is estimated with five lags for each variable in each equation, its residuals are free from serial correlation and heteroscedasticity.¹³ Figure 2 shows the dynamic response of the Mexican output to a one-standard deviation increase in each of the other variables of the model. This is done through a set of generalized impulse-response functions (GIRFs) with 95% confidence intervals.¹⁴ Such GIRFs are estimated over a six-quarter time horizon.

Note: The 95% confidence intervals were estimated through Hall’s (¹⁹⁹²) percentile bootstrap procedure.

Source: Authors’ estimations based on data from the National Institute of Statistics and Geography, the Bank of Mexico, the Mexican Secretariat of Economy, and the Federal Reserve Bank of Saint Louis.

Figure 2 Generalized impulse-response functions with 95% confidence intervals 

First, we can observe that an increase in exports raises Mexican output on impact and this effect dissipates at the end of the first quarter. This can be inferred from the confidence interval, which is above zero during the first quarter and includes zero as of the second quarter. Second, an increase in imports renders a similar effect on output. Third, real currency depreciation lowers output initially, but the effect is short-lived. The intuition behind this is that a real depreciation of the peso makes foreign intermediate inputs more expensive, thereby discouraging production. This is the supply-side effect of currency depreciation. In the long term, however, the alternative VAR model shows that a real depreciation of the peso stimulates production, presumably because it fosters exports by lowering its price in terms of dollars. This is the demand-side effect of exchange rate depreciation. It is worth mentioning that FDI does not yield a statistically significant effect on output in the short term. Lastly, an improvement in the terms of trade produces a slight positive effect on output, which can barely be observed at the beginning of the first quarter.

In the case of the alternative VAR model, ΔZtA=ΔYt,ΔXt,ΔMt,ΔQt, ΔYt*,ΔTTt' , which is also a five-lag stationary VAR model whose residuals are free of serial correlation and heteroscedasticity, the evidence indicates that an increase in US GDP yields a strong positive effect on Mexican production. Furthermore, the dynamic effect of the other variables on Mexican output is similar to the one observed through the benchmark VAR model, except that the negative effect of real currency depreciation and the positive effect of a terms-of-trade improvement are more noticeable (see appendix 3).