Using machine learning to assess the impact of deep trade agreements
VoxEU Column

Using machine learning to assess the impact of deep trade agreements

Modern preferential trade agreements contain a host of provisions that go beyond tariff liberalisation. By adapting techniques from the machine learning literature, this column develops data-driven methods for selecting the most trade-increasing provisions and quantifying their impact on trade. The results suggest that provisions related to technical barriers to trade, anti-dumping, trade facilitation, subsidies, and competition policy are associated with enhancing the trade-increasing effect of preferential trade agreements. Based on these findings, the effects of individual agreements can be estimated.

Preferential trade agreements (PTAs) have become more frequent and increasingly complex in recent decades, making it important to assess how they impact trade and economic activity. Modern PTAs contain a host of provisions besides tariff reductions in areas as diverse as services trade, competition policy, or public procurement. To illustrate this proliferation of non-tariff provisions, Figure 1 shows the share of PTAs in force and notified to the WTO up to 2017 that cover selected policy areas. More than 40% of the agreements include provisions such as investment, movement of capital and technical barriers to trade. And more than two-thirds of agreements cover areas such as competition policy or trade facilitation. 

Figure 1 Share of PTAs that cover selected policy areas

Figure 1

Note: Figure shows the share of PTAs that cover a policy area. Source: Hofmann, Osnago and Ruta (2019).

Recent research has tried to move beyond estimating the overall impact of PTAs on trade and tried to establish the relative importance of individual PTA provisions (e.g. Kohl et al. 2016, Mulabdic et al. 2017, Dhingra et al. 2018, Regmi and Baier 2020). However, such attempts face the difficulty that the number of provisions included in PTAs is very large compared to the number of PTAs available to study (see Figure 2), making it difficult to separate their individual impacts on trade flows. 

Figure 2 The number of provisions in PTAs over time

Figure 2

Source: Mattoo et al. (2020).

Researchers have tried to address the growing complexity of PTAs in different ways. For example, Mattoo et al. (2017) use the count of provisions in an agreement as a measure of its ‘depth’ and check whether the increase in trade flows after a given PTA is related to this measure. Dhingra et al. (2018) group provisions into categories (such as services, investment, and competition provisions) and examine the effect of these ‘provision bundles’ on trade flows. Obviously, these approaches come at the cost of not allowing the identification of the effect of individual provisions within each group.

New methodologies 

In recent research (Breinlich et al. 2022), we instead adopt a technique from the machine learning literature – the ‘least absolute shrinkage and selection operator’ (lasso) – to the context of selecting the most important provisions and quantifying their impact. More precisely, we adapt the ‘rigorous lasso’ method of Belloni et al. (2016) to the estimation of state-of-the-art gravity models for trade (e.g. Yotov et al. 2016, Weidner and Zylkin 2021).1

Unlike traditional estimation methods such as least squares and the maximum likelihood that are based on optimising the in-sample fit of the estimated model, lasso balances in-sample fit with parsimony to optimise the out-of-sample fit and to simultaneously select the more important regressors and estimate their effect on trade flows. In our context, the lasso works by shrinking the effects of individual provisions towards zero and progressively removing those that do not have a significant impact on the fit of the model (for an intuitive description, see Breinlich et al. 2021; for more details, see Breinlich et al. 2022). The rigorous lasso of Belloni et al. (2016), a relatively recent variant of the lasso, refines this approach by taking into account the idiosyncratic variance of the data and by only keeping variables that are found to have a statistically large impact on the fit of the model.

Because the rigorous lasso tends to favour very parsimonious models, it may miss some important provisions. To address this issue, we introduce two methods to identify potentially important provisions that may have been missed by the rigorous lasso. One of the methods, which we call ‘iceberg lasso’, involves regressing each of the provisions selected by the rigorous lasso on all other provisions, with the purpose of identifying relevant variables that were initially missed due to their collinearity with the provisions selected in the initial step. The other method, termed ‘bootstrap lasso’, augments the set of variables selected by the plug-in lasso with the variables selected when the rigorous lasso is bootstrapped.

Results and caveats 

We use the World Bank's database on deep trade agreements, where we observe 283 PTAs and 305 ‘essential’ provisions grouped into the 17 categories detailed in Figure 1.2 The rigorous lasso selects eight provisions more strongly associated with increasing trade flows following the implementation of the respective PTAs. As detailed in Table 1, these provisions are in the areas of anti-dumping, competition policy, technical barriers to trade, and trade facilitation. 

Table 1 Provisions selected by the rigorous lasso

Table 1

Building on these results, the iceberg lasso procedure identifies a set of 42 provisions, and the bootstrap lasso identifies between 30 and 74 provisions that may impact trade, depending on how it is implemented. Therefore, the iceberg lasso and bootstrap lasso methods select sets of provisions that are small enough to be interpretable and large enough to give us some confidence that they include the more relevant provisions. In contrast, the more traditional implementation of the lasso based on cross-validation selects 133 provisions. 

Reassuringly, both the iceberg lasso and bootstrap lasso select similar sets of provisions, mainly related to anti-dumping, competition policy, subsidies, technical barriers to trade, and trade facilitation. Therefore, although our results do not have a causal interpretation and, consequently, we cannot be certain of exactly which provisions are more important, we can be reasonably confident that provisions in these areas stand out as having a positive effect on trade.

Besides identifying the set of provisions that are more likely to have an impact on trade, our methods also provide an estimate of the increase in trade flows associated with the selected provisions. We use these results to estimate the effects of different PTAs that have already been implemented. Table 2 summarises the estimated effects for selected PTAs obtained using the different methods we introduce. As, for example, in Baier et al. (2017 and 2019), we find a wide variety of effects, ranging from very large impacts in agreements that include many of the selected provisions to no effect at all in agreements that do not include any.3 

Table 2 also shows that different methods can lead to substantially different estimates, and therefore these results need to be interpreted with caution. As noted above, our results do not have a causal interpretation. Therefore the accuracy of the predicted effects of individual PTAs will depend on whether the selected provisions have a causal impact on trade or serve as a signal of the presence of provisions that have a causal effect. When this condition holds, the predictions based on this method are likely to be reasonably accurate, and in Breinlich et al. (2022), we report simulation results suggesting that this is the case. However, it is possible to envision scenarios where predictions based on our methods fail dramatically; for example, it could be the case that a PTA is incorrectly measured to have zero impact despite having many of the true causal provisions. Finally, we note that our results can also be used to predict the effects of new PTAs, but the same caveats apply.

Table 2 Partial effects for selected PTAs estimated by different methods

Table 2


We have presented results from an ongoing research project in which we have developed new methods to estimate the impact of individual PTA provisions on trade flows. By adapting techniques from the machine learning literature, we have developed data-driven methods to select the most important provisions and quantify their impact on trade flows. While our approach cannot fully resolve the fundamental problem of identifying the provisions with a causal impact on trade, we were able to make considerable progress. In particular, our results show that provisions related to anti-dumping, competition policy, subsidies, technical barriers to trade, and trade facilitation procedures are likely to enhance the trade-increasing effect of PTAs. Building on these results, we were able to estimate the effects of individual PTAs.

Authors’ note: This column updates and extends Breinlich et al. (2021). See also Fernandes et al. (2021).


Baier, S L, Y V Yotov and T Zylkin (2017), "One size does not fit all: On the heterogeneous impact of free trade agreements",, 28 April. 

Baier, S L, Y V Yotov and T Zylkin (2019), "On the Widely Differing Effects of Free Trade Agreements: Lessons from Twenty Years of Trade Integration", Journal of International Economics 116: 206-228.

Belloni, A, V Chernozhukov, C Hansen and D Kozbur (2016), "Inference in High Dimensional Panel Models with an Application to Gun Control", Journal of Business & Economic Statistics 34: 590-605.

Breinlich, H, V Corradi, N Rocha, M Ruta, J M C Santos Silva and T Zylkin (2021), "Using Machine Learning to Assess the Impact of Deep Trade Agreements", in A M Fernandes, N Rocha and M Ruta (eds), The Economics of Deep Trade Agreements, CEPR Press.

Breinlich, H, V Corradi, N Rocha, M Ruta, J M C Santos Silva and T Zylkin (2022), "Machine Learning in International Trade Research - Evaluating the Impact of Trade Agreements", CEPR Discussion paper 17325.

Dhingra, S, R Freeman and E Mavroeidi (2018), “Beyond tariff reductions: What extra boost to trade from agreement provisions?”, LSE Centre for Economic Performance Discussion Paper 1532.

Fernandes, A, N Rocha and M Ruta (2021), “The Economics of Deep Trade Agreements: A New eBook”,, 23 June. 

Hofmann, C, A Osnago and M Ruta (2019), "The Content of Preferential Trade Agreements", World Trade Review 18(3): 365-398. 

Kohl, T S. Brakman and H. Garretsen (2016), "Do trade agreements stimulate international trade differently? Evidence from 296 trade agreements", The World Economy 39: 97-131.

Mattoo, A, A Mulabdic and M Ruta (2017), "Trade creation and trade diversion in deep agreements", Policy Research Working Paper Series 8206, World Bank, Washington, DC.

Mattoo, A, N Rocha and M Ruta (2020), Handbook of Deep Trade Agreements, Washington, DC: World Bank.

Mulabdic, A, A Osnago and M Ruta (2017), "Deep integration and UK-EU trade relations," World Bank Policy Research Working Paper Series 7947.

Regmi, N and S Baier (2020), "Using Machine Learning Methods to Capture Heterogeneity in Free Trade Agreements," mimeograph.

Weidner, M, T Zylkin (2021), "Bias and Consistency in Three-Way Gravity Models," Journal of International Economics: 103513. 

Yotov, Y V, R Piermartini, J A Monteiro and M Larch (2016), An advanced guide to trade policy analysis: The structural gravity model, Geneva: World Trade Organization.


1 Our approach complements the one adopted by Regmi and Baier (2020), who use machine learning tools to construct groups of provisions and then use these clusters in a gravity equation. The main difference between the two approaches is that Regmi and Baier (2020) use what is called an unsupervised machine learning method, which uses only information on the provisions to form the clusters. In contrast, we select the provisions using a supervised method that also considers the impact of the provisions on trade.

2 Essential provisions in PTAs include the set of substantive provisions (those that require specific integration/liberalisation commitments and obligations) plus the disciplines among procedures, transparency, enforcement or objectives, which are required to achieve the substantive commitments (Mattoo et al. 2020).

3 It is worth noting that lasso based on the traditional cross-validation approach leads to extremely dispersedestimations of trade effects, with some of them being clearly implausible. This further illustrates the superiority of the methods we propose.

525 Reads