Methodology

Although VAR models are a common tool in empirical macroeconomics—used both in forecasting and for analyzing the dynamic impact of shocks to the economy—they suffer from some drawbacks. One problem is their heavy parameterization, which, in combination with small or moderate samples, can result in poor forecasting performance, particularly at longer horizons, because the levels at which forecasts converge are a function of the estimated parameters of the model. As a potential solution to this problem, Villani (2005) suggests a BVAR approach with an “informative prior” on the steady state of the process.

To see the benefits of this approach, consider first the standard BVAR model:

where G(L) = I—G₁L–…–G_pL^p is a lag polynomial of order p, x_t is an n × 1 vector of stationary macroeconomic variables and η_t is an n × 1 vector of i.i.d. error terms fulfilling E(η_t) = 0 and $E\left({\eta }_t{\eta }_t^{\prime }\right)=\mathrm{\Sigma }\mathrm{.}$ It is typically difficult to specify a prior distribution for μ in Equation (1) and the solution has therefore often been to employ a noninformative prior for these parameters. However, the difficulty of specifying a prior for μ is related to the chosen specification. Consider the alternative parameterization of the model suggested by Villani:

where G(L), x_t and η_t all are defined as above. This model—although nonlinear in its parameters—has the feature that ψ immediately gives us the steady state of the series in the system. Hence, it is often the case that the forecaster has an opinion regarding the parameters of ψ and an informative prior distribution can accordingly be specified.

In this paper, we follow Villani (2005) in estimating model (2) with the prior on Σ given by $p\left(\mathrm{\Sigma }\right)\propto {\left|\mathrm{\Sigma }\right|}^{-(n+1)/2},$ the prior on vec(G)—where $G=(G_1\mathrm{...}{G}_p)´-givenbyvec\left(G\right)\sim N_{p{n}^2}({\mathrm{\theta }}_G,{\mathrm{\Omega }}_G)$ and the prior on ψ given by $\mathrm{\Psi }\sim N_n({\mathrm{\theta }}_{\mathrm{\Psi }},{\mathrm{\Omega }}_{\mathrm{\Psi }})\mathrm{.}$ That is, the prior on Σ is noninformative, but the priors on the vectors of dynamic coefficients vec(G) and steady-state parameters ψ—which are characterized by normal distributions centered on particular values—will generally be informative. We will return to and discuss the parameters of these priors below. The priors are then combined with the data through the likelihood function. The conditional posterior distributions of the model are derived in Villani (2005) and the numerical evaluation is conducted using the Gibbs sampler with the number of draws set to 10,000.¹

Empirical Implementation

External conditions that might be relevant for Latin America comprise (at a minimum) three sets of factors: external demand, commodity prices, and global financial conditions. In our model, external demand is proxied by GDP growth of Latin America’s trading partners, weighted using export shares. We refer to this index as world GDP growth; note, however, that the weights are different from the usual PPP-GDP-based weights (in particular, U.S. growth is weighted with about 0.55 rather than its weight of about 0.2 in the world economy). Commodity prices are captured using a net export share-weighted index; and external financial conditions using U.S. Treasury bill rates, and the high-yield corporate bond spread in the United States.² As a measure of Latin American growth, a weighted index for Argentina, Brazil, Chile, Colombia, Mexico, and Peru—referred to as the “LA6” in the remainder of this section—was used.³ In addition, the model included the Latin America subcomponent of JPMorgan’s emerging market bond index, which is influenced both by external financing conditions and domestic fundamentals in Latin America.⁴ Hence:

where $y_t^{world}$ is the logarithm of export-share weighted world GDP; $i_t^{US},$ the three-month treasury bill rate; HY_t, the high-yield corporate bond spread in the United States; y_t, the logarithm of aggregate real GDP for the LA6 countries; c_t, a (net) export commodity price index for these countries; and EMBI_t, the JPMorgan emerging market bond index spread for Latin America.⁵

World growth and U.S. financial variables are treated as block exogenous with respect to the Latin American variables.⁶ The model was estimated on quarterly data, from 1994:Q2 to 2007:Q2, after defining prior distributions for both the vec(G) and ψ parameter vectors. Figure 1 shows our data (see the Appendix for sources).

Data

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Sources: See Appendix.Note: Growth rates are given as percentage changes with respect to the same quarter in the preceding year.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Data

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Sources: See Appendix.Note: Growth rates are given as percentage changes with respect to the same quarter in the preceding year.

• Download Figure
• Download figure as PowerPoint slide

Data

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Sources: See Appendix.Note: Growth rates are given as percentage changes with respect to the same quarter in the preceding year.

• Download Figure
• Download figure as PowerPoint slide

Slightly modified “Minnesota priors” (Litterman, 1986) were used for the dynamic coefficients, vec(G). Based on the assumption that a univariate random walk with drift is a good starting point for modeling GDP and commodity prices in levels (see Table A1), prior means on the first own lag for variables modeled in first differences were set equal to zero. Accordingly, the prior means for all higher order lags and for all cross-coefficients—that is, coefficients relating a variable to another variable in the system—were also set to zero.⁷ However, prior means on the first own lag of variables modeled in levels were set to 0.9. The reason for this is that a traditional Minnesota prior—that is, a prior mean on the first own lag equal to 1—is theoretically inconsistent with the mean-adjusted model (2), as a random walk does not have a well-specified unconditional mean.

Steady-state priors are shown in Table 1 (first column). They can be justified as follows:

Table 1.

Steady-State Prior and Posterior Distributions

Source: Authors’ estimations.Note: Units are percentage points for Latin American EMBI and U.S. high-yield bond spread, and in percent for all other variables.

¹Refers to a normal distribution.

Table 1.

Steady-State Prior and Posterior Distributions

	95 Percent Probability Interval
	Prior1	Posterior
World GDP growth	(3.0, 4.3)	(3.1, 3.7)
U.S. treasury bill rate	(3.0, 5.0)	(3.6, 4.9)
U.S. high-yield bond spread	(3.0, 6.0)	(3.7, 5.4)
LA6 GDP growth	(3.5, 5.0)	(3.5, 4.6)
Commodity price growth	(−2.0, 4.0)	(−0.9, 3.9)
Latin EMBI	(2.0, 5.0)	(2.3, 4.6)

Source: Authors’ estimations.Note: Units are percentage points for Latin American EMBI and U.S. high-yield bond spread, and in percent for all other variables.

¹Refers to a normal distribution.

View Table

Table 1.

Steady-State Prior and Posterior Distributions

	95 Percent Probability Interval
	Prior1	Posterior
World GDP growth	(3.0, 4.3)	(3.1, 3.7)
U.S. treasury bill rate	(3.0, 5.0)	(3.6, 4.9)
U.S. high-yield bond spread	(3.0, 6.0)	(3.7, 5.4)
LA6 GDP growth	(3.5, 5.0)	(3.5, 4.6)
Commodity price growth	(−2.0, 4.0)	(−0.9, 3.9)
Latin EMBI	(2.0, 5.0)	(2.3, 4.6)

Source: Authors’ estimations.Note: Units are percentage points for Latin American EMBI and U.S. high-yield bond spread, and in percent for all other variables.

¹Refers to a normal distribution.

• Priors for world growth were based on medium-term projections from the IMF’s WEO.

• Following standard convention, the prior for the U.S. three-month Treasury bill rate was based on a U.S. inflation target and an equilibrium real interest rate of approximately 2 percent each. These values are in line with Taylor (1993) and Clarida, Gali, and Gertler (1998).

• The steady-state prior for Latin American growth, centered on 4.25 percent, was based on econometric studies of the impact of economic reforms on long-run growth in Latin America (see Loayza, Fajnzylber, and Calderon, 2004; and see Zettelmeyer, 2006, for a survey).

• For the U.S. high-yield bond and EMBI spreads, we did not have guidance from either theory or the previous literature. Consequently, we did not impose strong priors, and instead defined wide distributions in line with the observed behavior of these variables since the late 1980s and early 1990s, respectively—that is, based on a somewhat longer sample period than the one used for estimation.

• Commodity prices are assumed to be reasonably well described by a random walk with a small drift component. The steady-state growth rate in commodity prices is accordingly centered on 1 percent and is not particularly wide despite the historically high variability of commodity prices.

The table shows that the estimated posterior distributions are within, and usually narrower than, the assumed prior intervals. We also confirmed that the short-run dynamics of the model were not affected by the steady-state priors chosen.⁸ Hence, the assumed steady-state priors do not prejudge the model’s short-run forecasts.

Impulse Response Functions and Variance Decompositions

A standard Cholesky decomposition of the variance-covariance matrix was used to identify independent standard normal shocks ε_t based on the estimated reduced-form shocks; that is, we used the relationships Σ = PP´ and ε_t = P^–1η_t, with the variables ordered as in x_t in Equations (3) and (4). Hence, world GDP growth is assumed to be contemporaneously independent of all shocks except its own; U.S. interest rates are assumed to contemporaneously depend only on world GDP shocks; and so on.⁹

Figure 2 shows the response of LA6 growth to various shocks; the full set of impulse response functions appears in the Appendix.¹⁰ The magnitude of standard deviation shocks is as follows: about 0.37 percentage point for world growth, 28 basis points for U.S. Treasury bill rate, 67 basis points for the U.S. high-yield bond spread, 5.5 percent for commodity prices, and 110 basis points for the Latin American EMBI. These shocks are estimated to have the following effects on Latin American growth:

Response of LA6 GDP Growth with Respect to One Standard Deviation External Shocks

(Percent change from four quarters earlier; shown over 12-quarter horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Response of LA6 GDP Growth with Respect to One Standard Deviation External Shocks

(Percent change from four quarters earlier; shown over 12-quarter horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Response of LA6 GDP Growth with Respect to One Standard Deviation External Shocks

(Percent change from four quarters earlier; shown over 12-quarter horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

• Increases in world growth are passed on to Latin America about one-for-one: a 0.37 percent world growth shock leads to an increase in (four-quarter) Latin American growth by about 0.52 percentage point after four quarters. This is similar to the impulse response of world growth with respect to its own shock, which also reaches a maximum of about 0.44 (though it gets there faster; see Figure 1).

• The reaction of Latin American growth to U.S. interest rates is more muted; a hike leads to a reduction in growth, but the effect is reasonably small.

• In contrast, a standard deviation (67 basis points) rise in the U.S. high-yield bond spread, interpreted as reflecting a retreat of investors from risk, has a very strong effect, leading to a decline of four-quarter growth in Latin America by about 0.7 percentage point after three quarters. Note that the U.S. high-yield bond spread also appears to have strong effects on the Latin American EMBI as well as an effect on world growth (or, more strongly on U.S. growth); both of these channels could play a role in transmitting the shock.

• A standard deviation commodity shock—which in this sample is a change of almost 5.5 percent in a quarter, illustrating how volatile Latin American commodity prices have been—leads to a change in four-quarter Latin American growth of about 0.4 percentage point after three quarters.

• Finally, a 110-basis point rise in the Latin American EMBI is associated with a drop in four-quarter growth by 0.4 percent after four quarters.

Figure 3 shows the variance decompositions for LA6 growth, see the Appendix for the full system). More than half of the medium-term (10-20 quarter horizons) forecast error variance of Latin American GDP growth is explained by external factors: approximately 18 percent by world growth shocks, 6 percent by commodity prices, and a remarkable 34 percent by U.S. financial conditions (the combined influence of U.S. short-term interest rates and the U.S. high-yield bond spread).

Variance Decomposition from Mean-Adjusted BVAR Model

(Shown over 20-quarter forecasting horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Variance Decomposition from Mean-Adjusted BVAR Model

(Shown over 20-quarter forecasting horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Variance Decomposition from Mean-Adjusted BVAR Model

(Shown over 20-quarter forecasting horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Out-of-Sample Forecasts

In addition to the impulse responses and variance decompositions just shown, we will be analyzing the effect of external conditions on Latin America using conditional forecasts (see below). It hence makes sense to ask first how the model performs as a forecasting tool relative to other models that use the same variables. We do this by comparing the out-of-sample forecasting performance from our mean-adjusted BVAR with two benchmarks: a conventional BVAR with diffuse priors on the steady-state values, and a classical (non-Bayesian) VAR.¹¹ In addition, we compared the model’s out-of-sample forecasts with forecasts published by the IMF’s WEO at roughly the same time.

Forecasts from the two BVAR models are generated in a straightforward manner. For every draw from the posterior distribution of parameters, a sequence of shocks is drawn and used to generate future data. This leads to as many paths for each variable as we have iterations in the Gibbs sampling algorithm (namely, 10,000). For each of the two models, a central forecast is then generated as the median forecast based on the forecast density at each horizon. These central forecasts are compared with each other and to the point forecast from the classical VAR.

We initially estimate all models—the two BVAR models and the classical VAR—using data from 1994:Q2 to 2001:Q2. Using these estimates, we generate forecasts to 2004:Q2, that is, for all quarterly horizons h between 1 and 12. We then extend that sample one period, re-estimate the models and generate new forecasts 12 periods ahead, and so on. Once the estimation sample reaches 2004:Q2, we forecast over consecutively shorter periods, because the actual data that we need to compare the forecasts with ends in 2007:Q2. The last evaluation is conducted on models estimated from 1994:Q2 to 2007:Q1 and forecast only one period ahead. This yields N₁₂ = 13 forecasts at the 12-quarter horizon, N₁₁ = 14 forecasts at the 11-quarter horizon, and so on, and N₁ = 24 forecasts at the one-quarter horizon.

The forecasting performance of the three VAR models is then compared using the horizon h root mean square error (RMSE), given by

where x_{t + h} is the actual value of variable x at time t + h and ${\hat{x}}_{t+h,t}$ is an h step ahead forecast of x generated at time t. The relative RMSE at horizon h is defined as

where RMSE_{ma, h} is the RMSE of the mean-adjusted model at horizon h and RMSE_{alternative, h} is the corresponding RMSE for the traditional BVAR or classical VAR.

Figure 4 shows the relative RMSE at horizons 1-12 for all variables in the system and both alternative models.¹² A number smaller than one indicates that our model outperforms the alternative at a particular forecasting horizon. Furthermore, a higher value for “traditional BVAR” compared with “classical VAR” means that the traditional BVAR does better than the classical VAR.

Forecasting Performance of Mean-Adjusted BVAR Model Compared with Traditional BVAR and Classical VAR

(Relative root mean square errors; shown over 12-quarter forecasting horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Forecasting Performance of Mean-Adjusted BVAR Model Compared with Traditional BVAR and Classical VAR

(Relative root mean square errors; shown over 12-quarter forecasting horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Forecasting Performance of Mean-Adjusted BVAR Model Compared with Traditional BVAR and Classical VAR

(Relative root mean square errors; shown over 12-quarter forecasting horizon)

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Figure 4 shows (unsurprisingly) that the mean-adjusted BVAR model decisively outperforms the classical VAR; the relative RMSE against the classical VAR is almost always smaller than unity. It also shows that the mean-adjusted BVAR generally performs better than the traditional BVAR. For four out of six variables, it produces smaller root mean squared errors at all horizons. In the other two cases—forecasts for world growth and commodities—the comparison depends on the horizon, with little difference in performance at the short end (1-3 quarter), somewhat better performance of the traditional BVAR at medium horizons (4-8 quarter); and better performance of the mean-adjusted BVAR at longer horizons. The latter likely reflects the role of steady-state priors, which help the forecasts converge to a sensible level.

We next compare the forecasting performance of the mean-adjusted BVAR model with the IMF’s WEO forecasts published by the IMF (Table 2). Following the WEO, we focus the comparison on forecasts of annual average growth rates (derived from the model’s quarterly forecasts) over a two-year horizon. We compare forecasts based on roughly similar information periods. Hence, Spring WEO forecasts (published in April of each year) are compared with model forecasts based on data through the fourth quarter of the previous year, but Fall WEO forecasts (typically published in September) are compared with model forecasts based on data through the second quarter of the ongoing year. This gives the WEO forecasts an informational advantage, because these are typically influenced by news in the first and third quarters, respectively, but the model estimates are not. (Of course, the WEO forecasts also have an informational advantage because they are based on many different data series, in addition to the six that we use in our models.) We focus on forecasts for the 2002-07 period, to allow the model a reasonably long estimation sample before comparing it with the WEO.

Table 2.

LA6 GDP Growth: Comparison of Model-Based and WEO Forecasts

(In percent)

Source: Authors’ estimations; and IMF, World Economic Outlook (WEO).

¹Preliminary.

Table 2.

LA6 GDP Growth: Comparison of Model-Based and WEO Forecasts

(In percent)

	2002	2003	2004	2005	2006	2007
Actuals, based on:
Data in Spring WEO of year after forecast	0.3	2.0	5.4	4.1	5.2	4.91
Most recent data (used in model estimation)	0.3	2.7	5.6	4.2	4.9	4.91
Forecasts of:
Fall 2002 WEO	−0.5	3.0
Model; data through 2002:Q2	0.5	4.7
Spring 2003 WEO		2.7	3.8
Model; data through 2002:Q4		2.2	3.5
Fall 2003 WEO		2.2	3.4
Model; data through 2003:Q2		2.8	5.1
Spring 2004 WEO			3.8	3.7
Model; data through 2003:Q4			6.0	6.5
Fall 2004 WEO			4.5	3.6
Model; data through 2004:Q2			5.8	6.3
Spring 2005 WEO				4.2	3.6
Model; data through 2004:Q4				5.2	4.4
Fall 2005 WEO				4.0	3.8
Model; data through 2005:Q2				4.6	4.5
Spring 2006 WEO					4.3	3.6
Model; data through 2005:Q4					4.1	5.1
Fall 2006 WEO					4.6	4.2
Model; data through 2006:Q2					4.9	5.0
Spring 2007 WEO						4.8
Model; data through 2006:Q4						4.3

Source: Authors’ estimations; and IMF, World Economic Outlook (WEO).

¹Preliminary.

View Table

Table 2.

LA6 GDP Growth: Comparison of Model-Based and WEO Forecasts

(In percent)

	2002	2003	2004	2005	2006	2007
Actuals, based on:
Data in Spring WEO of year after forecast	0.3	2.0	5.4	4.1	5.2	4.91
Most recent data (used in model estimation)	0.3	2.7	5.6	4.2	4.9	4.91
Forecasts of:
Fall 2002 WEO	−0.5	3.0
Model; data through 2002:Q2	0.5	4.7
Spring 2003 WEO		2.7	3.8
Model; data through 2002:Q4		2.2	3.5
Fall 2003 WEO		2.2	3.4
Model; data through 2003:Q2		2.8	5.1
Spring 2004 WEO			3.8	3.7
Model; data through 2003:Q4			6.0	6.5
Fall 2004 WEO			4.5	3.6
Model; data through 2004:Q2			5.8	6.3
Spring 2005 WEO				4.2	3.6
Model; data through 2004:Q4				5.2	4.4
Fall 2005 WEO				4.0	3.8
Model; data through 2005:Q2				4.6	4.5
Spring 2006 WEO					4.3	3.6
Model; data through 2005:Q4					4.1	5.1
Fall 2006 WEO					4.6	4.2
Model; data through 2006:Q2					4.9	5.0
Spring 2007 WEO						4.8
Model; data through 2006:Q4						4.3

Source: Authors’ estimations; and IMF, World Economic Outlook (WEO).

¹Preliminary.

The “actuals” that we compare both forecasts with are slightly different. The model’s out-of-sample forecasts are compared with the latest available GDP growth data, which are also used to estimate the model; but the WEO forecasts are compared with data published at the time (in practice, we use the “actuals” published in the Spring WEO in the year following the time of the forecast). This is because we cannot expect WEO forecasts to anticipate revisions in GDP that are reflected in the data series used to estimate the model.¹³

As in the previous comparison, the forecasts shown in Table 2 can be used to estimate root mean squared errors. Because we are limiting the comparison with the one and two-year forecasting horizons, we only compute two RMSEs for each set of forecasts. For the one-year horizon (meaning the current year for the Spring WEO forecasts), the model, perhaps surprisingly, does slightly better than the WEO: the RMSE is 0.8 percentage point for the WEO and only 0.5 for the model. At the two-year horizon, the WEO’s RMSE is 1.1, but the model RMSE is 1.2. These are very small differences. Overall, the forecasts of the model appear to be about as good as those of the WEO, in spite of the WEO’s much richer information basis.

It is instructive to examine where the model and WEO forecasts differ. Although the comparison period is obviously too short to draw firm conclusions, it appears that the model is more sensitive than the WEO in picking up turning points, but sometimes exaggerates the cyclical change. The model did much better than the WEO in predicting the strength of the 2004 recovery based on information through mid-2003. However, it predicted the recovery to strengthen further in 2005, which was not the case. Similarly, the model, in response presumably to lower U.S. growth in the second half of 2006, predicted a moderate slowing in Latin American growth in 2007. In the event, Latin American growth did not slow in 2007 relative to 2006, driven by buoyant domestic demand, and the higher persistence of the WEO forecast turned out to be an advantage in this case.

Conditional Forecasts and Scenario Analysis

In addition to producing unconditional (or “endogenous”) forecasts, as discussed so far, the mean-adjusted BVAR model turns out to be a convenient machinery for conditional forecasts, that is, forecasts based on assumptions about the future paths of some of the endogenous variables.¹⁴ Conditional forecasts can serve two purposes. First, they are a way of incorporating extra-model information—” judgment,” in Svensson’s (2005) terminology—into the forecasting process. For example, assumptions about world growth or about the future path of commodity prices could be fed into the model. To the extent that these are based on information outside the model (such as commodity price forecasts based on futures prices), this might improve overall forecasting performance. Second, conditional forecasts can be used for scenario analysis, that is, to examine how growth would respond to specific external events. It is in this sense that conditional forecasts will be used extensively in this section.

We generate conditional forecasts as follows (see Österholm (2006) for details). As described in the previous section, we are interested in generating a distribution of future paths of the endogenous variables. To generate each path, we require the historical data, a draw from the posterior distribution of parameters, and a sequence of orthogonal shocks (ε_T+1,…, ε_T+H). These shocks are then used together with the definition ε_t = P^-1η_t to generate the reduced-form shocks and hence—given history and the realization of the parameters—the future data. The only difference between the unconditional and conditional forecasting exercises is that in the unconditional case, the entire vector ε_T+h is generated randomly at each horizon, through independent draws from a normal distribution. In contrast, in the conditional case, only the orthogonal shocks belonging to a subset of the endogenous variables are created randomly; the shocks of the conditioning variables are implied by the assumed conditioning path. For a given set of randomly generated orthogonal shocks of the variables that have not been conditioned upon and a given path of the conditioning variables, the implicit shocks of the conditioning variables and the forecasts for all variables are generated sequentially, one horizon at a time.

Figure 5 shows two conditional forecasts based on the model estimated through the second quarter of 2007. First a “quasi-unconditional” forecast which conditions only on the realizations of financial and commodity price variables in Q3—which by now are observable—and an estimate for Q3 external growth, and after that projects all variables endogenously. Second, a “baseline” conditional forecast based on world growth and commodity price paths projected by the October 2007 WEO, as well as interest rate paths consistent with that outlook. In particular, following a weak fourth quarter, output in the United States is assumed to gradually recover in 2008 in the conditional baseline scenario, leading to a moderate increase in short-term interest rates, and a moderate decline in the high-yield bond spread from levels observed at the beginning of the fourth quarter. In the figure, theses conditioning paths are recognizable by the fact that they do not have a probability “fans” around them.

Quasi-Unconditional vs. Conditional Baseline Forecasts

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.Note: Units are percentage points for Latin EMBI and U.S. high-yield bond spread; and percent for all other variables.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Quasi-Unconditional vs. Conditional Baseline Forecasts

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.Note: Units are percentage points for Latin EMBI and U.S. high-yield bond spread; and percent for all other variables.

• Download Figure
• Download figure as PowerPoint slide

Quasi-Unconditional vs. Conditional Baseline Forecasts

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.Note: Units are percentage points for Latin EMBI and U.S. high-yield bond spread; and percent for all other variables.

• Download Figure
• Download figure as PowerPoint slide

As can be seen from the figure, the conditioning paths turn out to be close to the “quasi-unconditional” forecasts for most variables. The largest difference is with respect to external demand growth, where the quasi-unconditional forecast envisages a modest rebound in 2008, to 3.5 percent growth on average, but the WEO forecast implies a slight decline. Hence, the comparison of the quasi-unconditional and conditional baseline forecasts gives a sense of the effects of the moderate slowing of the world economy—without a further deterioration of financial conditions—on Latin America. The main result is that Latin America would slow only slightly, by about 0.3 percent on average in 2008 relative to the quasi-unconditional forecast, and by about 0.6 percent relative to the expected 2007 outturn. The conditional point forecast for average annual growth in 2008 is 4.4, about in line with the October 2007 WEO projection for the LA6 countries (4.2).

We next examine how this conditional forecast changes in a number of scenarios, which represent particular risks to the external environment that are suspected to have an impact on Latin American growth. Appendix Table A2 summarizes these scenarios and compares them to the “quasi-unconditional” forecast and the conditional baseline forecast.

A U.S. recession and credit crunch

The most obvious external risk looming over Latin America today is the possibility that a deepening U.S. housing crisis may trigger a U.S. recession in 2008, as consumer confidence collapses and financial market turbulence begins to affect corporate credit conditions. For illustrative purposes, we assume a scenario in which U.S. growth declines sharply in the fourth quarter of 2007 and remains around zero in the first half of 2008, before beginning a slow recovery in the second half. Average annual growth in 2008 would be reduced from just under 2 percent in the baseline to about 0.8. The Federal Reserve is assumed to cut interest rates aggressively to prevent a deeper and longer downturn, leading to a decline in U.S. Treasury bill rate to about 2 percent by the end of the second quarter. At the same time, corporate credit is likely to decline, and credit spreads for sub-investment-grade borrower would likely rise sharply, to over 700 basis points, in line with previous U.S. recessions.

Slower U.S. growth and tighter credit market conditions are likely to spill over to industrial growth outside the United States. In line with Bayoumi and Swiston (2007), for each one point reduction in U.S. growth we assume a reduction of about 0.4 to 0.5 percent in the growth of major non-U.S. importers of Latin American goods. Given the high weight of the United States in Latin American exports, this implies an overall reduction of export demand by about 0.8 percentage point below baseline, on average, in 2008.

We make no assumption about the price path of commodities; that is, commodity prices are left endogenous. Not surprisingly, the model predicts that they would fall in reaction to the assumed U.S. and world slowdown, by about 22 percent by mid-2008, before recovering slightly (see Table A2).

Finally, regarding the Latin EMBI, we make two alternative assumptions:

• One option is to simply leave the EMBI endogenous. Because the typical response of the EMBI to changes in the high-yield bond spread during this period was at least one-for-one during the 1994-2007 estimation period (see impulse responses in Figure A2), this implies a very sharp rise, from about 230 basis points on average in the third quarter to 670 basis points by the first quarter of 2008.

• Alternatively, one can assume a path for the EMBI in line with the much more muted reactions of the EMBI to changes in the high-yield bond spread observed in the last months. In this case, the EMBI would rise by only about half the amount of the endogenous rise, to 370 basis points by the first quarter.

Figure 6 and Table A2 summarize the conditioning paths and compare them to the baseline paths, but Figure 7 shows the results.

Conditioning Paths for Baseline and U.S. Recession Scenario

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Conditioning Paths for Baseline and U.S. Recession Scenario

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Conditioning Paths for Baseline and U.S. Recession Scenario

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Effects of a U.S. Recession and Credit Crunch

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Effects of a U.S. Recession and Credit Crunch

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Effects of a U.S. Recession and Credit Crunch

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Not surprisingly, the magnitude of the predicted slowdown depends on the behavior of the Latin EMBI—that is, on the extent of financial contagion.

• If this follows the average pattern during the sample period, the predicted effect of a 2008 U.S. recession and credit crunch is large, with about zero growth during the first half of 2008 relative to the second half of 2007, just short of a recession. The 2008 average growth falls by 2.1 percentage points relative to baseline, to 2.3 percent.

• If external financing premia in Latin America were to continue their partial decoupling from U.S. financial markets, the model predicts a milder slowdown, to about 3.2 percent in 2008, or about 1.2 points relative to baseline.

Note that in either case, the predicted fall of Latin American output relative to the baseline (1.2 to 2.1 percent) exceeds the assumed fall in external demand (0.8). The reason for this is that the scenario consists of a combined demand and financial shock, given the assumed sharp rise of the U.S. high-yield bond spread. Even in the more muted scenario, the financial transmission channel has not been shut down completely, and the commodity price channel is also at work in both variants of the scenario.

Declines in commodity prices

As discussed, declining commodity prices constitute one channel through which a slowdown in external growth could hurt Latin America. It is also interesting to see how much a commodity price fall of about the same magnitude would slow Latin American growth if it occurred in isolation. This can be done by modifying the baseline scenario to include a fall in commodity prices by 20 percent over three quarters, beginning in the last quarter of 2007 (see Table A2). All other baseline paths are retained, and the EMBI is allowed to adjust endogenously.

The model predicts that a 20 percent fall in commodities prices in late 2007 and early 2008 would lead to noticeably lower, but still robust, 2008 growth in Latin America (see Figure 8 and Table A2). Annual average growth is reduced to 3.7 percent. This partly reflects the direct effect of commodity prices on growth, but also higher external risk premia, with the EMBI rising endogenously by about 60 to 80 basis points above baseline (see Figure 8).

Effects of a Decline in Commodity Prices

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Effects of a Decline in Commodity Prices

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Effects of a Decline in Commodity Prices

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

An emerging market financing shock

Finally, a scenario illustrating much tighter emerging market financing conditions was considered, in the form of a 200-basis-point shock to the Latin EMBI spread and a 100-basis-point increase in the U.S. high-yield bond spread—perhaps triggered by an emerging market crisis outside Latin America, with a limited spillover effect to U.S. high-yield credit markets. To isolate the effects of a “pure” emerging market financing shock, we retain baseline assumptions for world growth, commodity prices, and U.S. Treasury bill rates. The shock to the EMBI spread and the U.S. high-yield bond spread is assumed to occur in the fourth quarter, after which these variables are allowed to be endogenous.

The model suggests that a financing shock of this kind would significantly reduce growth in Latin America, though by less than the combined U.S. recession/credit crunch scenario considered earlier (Figure 9). The 2008 growth is predicted to fall to about 3.3 percent on an annual average basis, about 1 percent below baseline.

Effects of Emerging Market Financing Shock

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Effects of Emerging Market Financing Shock

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide

Effects of Emerging Market Financing Shock

Citation: IMF Staff Papers 2008, 001; 10.5089/9781589067257.024.A003

Source: Authors’ estimations.

• Download Figure
• Download figure as PowerPoint slide