Foreign exchange (FX) intervention involves the purchase and/or sale of foreign currency by the central bank. It is a long-standing policy instrument designed to impact exchange rates and foster stable currency markets. Many emerging economies use it frequently, indicating its (perceived) usefulness (e.g. Gelos et al. 2020). While most central banks in advanced economies use this instrument more sporadically, Switzerland and Israel have intervened extensively since the global financial crisis (Cukierman 2018), and the Bank of Japan (BoJ) has been the most active major monetary authority in the currency markets in recent decades. Indeed, around 22 September 2022, the BoJ sold a hefty US$20 billion to buy and support a sharply weakening Japanese yen. At a time reminiscent of the strong US dollar of the 1980s that led to some of the biggest FX intervention actions coordinated by central banks around the world, the major central banks seem likely to trade in currency markets to support their currencies, independently or in a coordinated fashion. The IMF (2022) has recently fine-tuned its institutional view on the management of capital flows, seeing stabilisation potential not only in times of crisis but also for prevention (Korinek et al. 2022). Thus, FX interventions are not an urchin of capital account management but have been welcomed back into the family of tools policymakers can use.
In contrast to its widespread use by many central banks, hard evidence on the impact of FX interventions remains limited. The main reason is simply the paucity of reliable data because, unlike monetary policy, there is no consistent data source. In recent research (Fratzscher et al. 2022), we contribute to filling this gap by providing a new dataset on FX interventions for 49 countries over up to 22 years. These data are publicly available and come with some advantages over alternatives that we discuss below. Let us start by discussing briefly how these FX intervention data are generated.
Information from news reports
The basic idea is to retrieve information about FX interventions from publicly available news reports to overcome the lack of official data. This procedure was pioneered by Klein (1996) and a few papers have used it, with the limitation that all of them are case studies. Nowadays vast news archives can be searched and analysed with the help of textual analysis. The main constraint is pre-selecting, reading, and classifying the large number of intervention-related news. For this, we develop a text classification algorithm that provides information on the incidence of FX interventions for a large number of countries and over long timeframes. The first step was to train the news classification algorithm so it is capable of correctly classifying news items that cover FX interventions. The news items were retrieved from Factiva. Then, several thousand news items were hand-coded with a double-entry technique to build a training dataset.
This trained algorithm is quite reliable in correctly identifying which news cover intervention. At the monthly frequency, hand-coded news and algorithmically-coded news agree in 99% of cases. Then the trained algorithm is used to classify more currencies and extend the data period from 2011 to 2016 for which news items were not hand-coded. The result is a dataset that indicates every month whether there was an FX intervention or not.
Benchmarking with actual intervention data
Reading the news and comparing them to the published intervention data makes clear that some interventions are not covered in the news. However, doing this only for a few countries that decided to publish their intervention data for certain years risks giving an unrepresentative picture. Thus, we rely on a set of actual, confidential FX intervention data in spot markets and news about FX interventions for 33 countries over up to 16 years that also covers countries that do not make their data publicly available (Fratzscher et al. 2019).
From incidence to volumes
Unfortunately, news reports are typically not enough to identify the direction of intervention, i.e. whether central banks were supporting the foreign or the domestic currency. Even worse, there is usually no information about the volume of intervention – at best, the texts sometimes mention whether interventions were small or large. Thus, for many purposes, it is necessary to complement information about the incidence of FX intervention with information on the volume of intervention. This information is taken from the change in (adjusted) reserve assets in the same month (Neely 2000, Dominguez et al. 2012, Adler et al. 2021). Combining the incidence from news reports and the volume from reserve changes, we generate a ‘news proxy’ that approximates actual FX intervention.
This leads to the obvious question: Why not take the change in reserves as a general measure of FX intervention? Indeed, there is a stream of literature that does exactly this.
Reserve changes as a measure for FX interventions
The problem with using reserves as a proxy for intervention is that reserves change basically every month, while FX interventions occur less frequently (Neely, 2000). In the dataset of Fratzscher et al. (2019), actual interventions occur in about 20% of the observations but not in the remaining 80%. This means that any analysis built on reserve changes massively overestimates the use of FX interventions.
Taking this into account, one way to improve information from reserve changes is to focus only on large changes. It seems plausible (and is indeed true when tested with actual intervention data) that actual interventions have larger volumes than what could be inferred from average monthly reserve changes. Figure 1 shows the result of the procedure discussed above on a dataset in which FX interventions occur in roughly 30% of monthly observations. This is also the starting point of the solid line to the left of the graph. Using no cutoff, i.e. considering all months as intervention months, implies that all of the circa 30% of actual intervention months will be detected. If, for example, only the 50% largest interventions are considered, ‘precision’ will increase to about 40%. Increasing the cutoff further will also further increase precision. The maximum precision will be about 60%, but this would mean focusing only on the largest few percent of interventions. The dashed line in Figure 1 that starts off with about 1,200 intervention months and goes to zero with increasing cutoffs captures this falling number of cases that are covered. Another takeaway that is relevant for any proxy is that designing them involves facing a trade-off between different types of errors. Instead of choosing from the options that reserve changes provide and that are covered in Figure 1, our news-based intervention proxy provides a better alternative.
Figure 1 Various cutoffs of reserve changes, the number of covered FX interventions (right axis, dashed line) and share of correct predictions (left axis, solid line)
Characteristics of the news proxy
The core contribution of the news proxy is its information about the incidence of interventions. Again using the benchmark data, this news-based information is much more likely to be true. On average, the trained algorithm is correct in 75% of cases. By comparison, in the exact same dataset, the accuracy of the reserve changes is only 31%. Yet, when relying on news, one will miss some interventions. One needs a unified measure to account for this trade-off.
The noise-to-signal ratio
A standard measure that combines how well a proxy correctly signals an event while penalizing false alarms is the noise-to-signal ratio. For this, it divides the ‘probability of false alarm’ by the ‘probability of detection’. To understand these definitions, it is useful to take a look at the matrix of four possible outcomes shown in Table 1. If there is a country-month in which an intervention took place and an intervention was also predicted, this is called ‘true positive’ (A). The remaining fields are constructed analogously and labeled (B) to (D).
Table 1 Types of prediction errors
The ‘probability of detection’ is the share of correctly predicted interventions (true positives) over all interventions, i.e. A/(A+C). The ‘probability of false alarm’ is defined as the share of false positives denominated by all actual non-interventions, i.e. B/(B+D). The ratio of the latter over the former is then the noise-to-signal ratio which forms a kind of relation between success and failure. We apply these standard metrics to judge the quality of the three measures approximating actual FX intervention: (i) reserve changes, (ii) reserve changes with cutoff, and (iii) news proxy. Table 2 provides results.
Table 2 Outcomes of predictive quality
Starting in the top left corner, the accuracy of reserve changes is 0.308. When comparing the three types of intervention proxies, we see the trade-off mentioned above: the reserve proxies’ probability of detection is 1; this means that if a proxy classifies almost every month as having an intervention, it will not miss any. Yet, obviously, the value of such a proxy is quite limited. By contrast, the news proxy in the last column is worse regarding the probability of detection but better regarding the probability of false alarm. With a mere 0.043, the latter is so low that the noise-to-signal ratio of the news proxy clearly dominates the two alternative intervention proxies.
Thus researchers have basically three alternatives when analysing FX intervention data.
- They may rely on officially published data, implying that there are only a few data with limited country and time coverage. Also, publishing intervention data is a policy decision, resulting in a sample that is not representative of all countries. Still, using such published data is the preferred way for individual case studies where actual data are made available by the central bank.
- Researchers may rely on reserve changes because this measure is available for most countries in the world over long periods. The price to be paid is lots of noise and, depending on the research question, the very low degree of precision of this measure may be troubling. As a countermeasure, one may prefer to focus only on larger interventions by using reserve changes with cutoff, as this increases precision at the cost of missing more actual interventions.
- Finally, one may rely on the news proxy. This measure is the best available to reduce the probability of false alarm. The price is that a number of actual interventions are missed. If one aims to analyse the impact of FX interventions that are relatively well-identified, for example by using the data as a shock series in a model, an analysis with missing interventions seems more attractive than an analysis dominated by many false positives.
The debate about the effectiveness of FX interventions and their role in capital account management requires data to run useful empirical analyses. We provide such a publicly available database. Aside from the coverage that goes far beyond officially available intervention data, its main advantage is that it has much less noise compared to other proxies that are solely based on reserve changes.
Adler, G, K S Chang, R Mano, and Y Shao (2021), “Foreign exchange interventions: A dataset of public data and proxies”, IMF Working Paper 2021/047.
Cukierman, A (2018), “Forex intervention and reserve management in Switzerland and Israel since the financial crisis: Comparison and policy lessons”, VoxEU, November 2.
Dominguez, K M E, Y Hashimoto, and T Ito (2012), “International reserves and the global financial crisis”, Journal of International Economics 88: 388-406.
Fratzscher, M, O Gloede, L Menkhoff, L Sarno, and T Stöhr (2019), “When is foreign exchange intervention effective? Evidence from 33 countries”, American Economic Journal: Macroeconomics 11(1): 132-156.
Fratzscher, M, T Heidland, L Menkhoff, L Sarno, and M Schmeling (2022), “Foreign exchange intervention: A new database”, CEPR Discussion Paper 17558, IMF Economic Review, forthcoming.
Gelos, G, L Gornicka, R Koepke, R Sahay, and S Sgherri (2020), “Capital flows at risk: Taming the ebbs and flows”, VoxEU, February 4.
IMF (2022), “Review of the institutional view on the liberalization and management of capital flows”.
Klein, M W (1996), “The accuracy of reports of foreign exchange intervention”, Journal of International Money and Finance 12: 644-653.
Korinek, A, P Loungani, and J D Ostry (2022), “A welcome evolution: The IMF’s thinking on capital controls and next steps”, VoxEU, April 8.
Neely, C J (2000), “Are changes in foreign exchange reserves well correlated with official intervention?” Federal Reserve Bank of St. Louis Review 82(5): 17-32