RDP 2025-09: Forecasts of Period-average Exchange Rates: Insights from Real-time Daily Data 1. Introduction

Exchange rate forecasting is crucial for guiding economic decisions in policymaking and investment strategies. Of particular interest in macroeconomic analysis are period-average exchange rates – such as effective or bilateral rates expressed in real terms – which are constructed as the arithmetic average of daily closing exchange rates over a specified time interval, typically a month or a quarter. Relative to exchange rates sampled at specific points in time (‘point-sampled’ rates), period-average exchange rates are more relevant to variables measured as flows over time, such as net exports, inflation, revenue and costs, and broader economic conditions. For this reason, many policymakers, including international organisations and central banks, make assumptions about period-average exchange rates as part of their routine forecasting processes (see, for instance, Wieland and Wolters (2013), Glas and Heinisch (2023), International Monetary Fund (2023)).

In this paper, we ask whether period-average exchange rates are forecastable in real time. The core contribution is a new real-time dataset of daily nominal and real effective exchange rates (EERs) for all countries, including the first daily real effective exchange rates (REERs). This dataset enables the first formal testing of real-time forecasts of period-average exchange rates against the traditional random walk benchmark. It also allows us to benefit from the efficiency gains from using high-frequency data to construct forecasts. With these data in hand, we revisit longstanding concerns in the literature that have not been possible to resolve until now due to data limitations.

Forecasting period-average exchange rates presents distinct challenges. One issue is that using period-average data when constructing forecasts can introduce information loss that diminishes forecasting accuracy (e.g. Wei 1978; Kohn 1982; Lütkepohl 1986). Since period-average exchange rates are constructed from daily point-sampled data, efficient forecasts require access to the underlying high-frequency observations. A second issue is the potential for spurious predictability due to time averaging (Working 1960; Marcellino 1999; Bork, Rovira Kaltwasser and Sercu 2022; Ellwanger and Snudden 2023). This explains why Meese and Rogoff (1983a) and much of the literature forecast end-of-period bilateral exchange rates instead. In contrast, forecasts of REERs have relied on period averages due to the historical absence of point-sampled or daily effective exchange rate data (Meese and Rogoff 1983a, p 9).

To navigate the body of existing work, we contribute a detailed survey of the temporal assumptions used in the exchange rate forecasting literature in Section 2. This survey identifies three gaps in the literature, which our paper aims to fill. First, the survey shows that the literature has yet to test the predictability of period-average exchange rates by comparing them with the traditional random walk benchmark. Second, it shows the literature forecasting period-average exchange rates uses models estimated on period-average inputs, suggesting the forecasts are potentially inefficient. Third, the survey shows that no paper has done a real-time evaluation of forecasts for period-average or point-sampled EERs, or for point-sampled bilateral real exchange rates (RERs).

To address these limitations, we construct a new exchange rate dataset. For all available countries, we construct real-time vintages at the daily frequency for the four types of daily exchange rate: nominal effective exchange rates (NEERs), REERs, bilateral nominal exchange rates (NERs) and bilateral RERs.

Using the new dataset, we evaluate how temporal aggregation affects the accuracy of real-time forecasts for period-average exchange rates. We assess both model-based and no-change forecasts, and test, for the first time, real-time out-of-sample forecasts of monthly averages against the traditional random walk benchmark.^[1] The results reveal three key empirical findings that underscore the importance of temporal aggregation bias in exchange rate forecasting.

The first empirical finding is that, for all measures of exchange rates and for almost all countries, the month-average no-change benchmark is less accurate than the end-of-month no-change benchmark. The difference in performance is large – for example, directional accuracy is improved by up to 40 per cent. This evidence for all exchange rates and almost all countries confirms the evidence for the USD/JPY bilateral rate by Ellwanger and Snudden (2023) and occurs because daily exchange rates are highly persistent. Our first finding suggests that the monthly average no-change benchmarks used for EERs since Meese and Rogoff (1983a) may provide too lenient a benchmark against which to assess forecast performance. However, importantly, we find that unlike bilateral exchange rates, the forecast precision gains from temporal disaggregation of daily EERs are far smaller. This is insightful, as it provides evidence that EERs exhibit properties distinct from random walk processes.

The second empirical finding is that both direct and recursive model-based forecasts estimated with daily or end-of-month inputs perform substantially and significantly better than forecasts estimated with month-average data. This is found to be robust across exchange rate measures and countries. Once again, this substantiates theoretical advantages regarding the gains in forecast accuracy when exchange rates are temporally disaggregated. These findings are encouraging; they show that one can substantially improve the accuracy of model forecasts for period-average exchange rates using real-time information from daily exchange rates. Interestingly, we also find evidence that end-of-period point-sampled exchange rate forecasts are useful for forecasts of period-average exchange rates. This implies that methods in existing studies examining forecasts of end-of-period exchange rates are potentially very useful for forecasts of period-average exchange rates.^[2]

The third empirical finding is that when applying efficient estimation and testing methods, made possible for the first time by the daily data, we find new evidence of real-time predictability for period-average EERs in up to 50 per cent of countries. In contrast, for period-average bilateral exchange rates, there is substantial spurious predictability, in both directional accuracy and mean-squared precision, when temporally disaggregated model-based real-time forecasts are compared against period-average no-change forecasts.^[3] This raises an important distinction between exchange rate measures as, for both real and effective measures of exchange rates, the use of high-frequency information substantially increases the likelihood of rejecting the random walk hypothesis. This suggests a refinement to the current consensus on the inability of exchange rates to outperform naive benchmarks. While bilateral NERs, derived from financial markets, behave like random walks, forecasts of bilateral RERs and EERs are more accurate when efficiently constructed using the realtime daily data.

Our findings contribute to the broader understanding of temporal aggregation bias. We provide quantitative evidence to support theoretical claims that temporal averaging reduces forecast accuracy (Tiao 1972; Amemiya and Wu 1972; Kohn 1982; Lütkepohl 1986). Importantly, we show that the loss in accuracy from daily to monthly aggregation is considerably larger than what has been documented for aggregation from monthly to quarterly or quarterly to annual frequencies (e.g. Zellner and Montmarquette 1971; Lütkepohl 1986; Athanasopoulos et al 2011). This suggests that the gains from temporal aggregation should be understood to be of first order importance when the forecast target is a period average of daily data.

These results have broader implications for macroeconomic measurement. Although our focus is on month-average and end-of-month exchange rates, the informational loss from temporal aggregation is even more severe at quarterly frequencies. More generally, we recommend that end-of-period values for EERs be routinely reported, just as they are for bilateral exchange rates.

This paper serves as a guide to understanding temporal aggregation in exchange rate forecasting. The findings offer a framework for interpreting existing empirical results and for designing future forecast evaluations. By incorporating high-frequency data, we demonstrate that real-time forecasts of period-average exchange rates can be substantially improved, providing insights that are directly relevant for economic decision-making.

Footnotes

For forecasts of period averages, the traditional random walk no-change is given by the last observed end-of-period value (Ellwanger and Snudden 2023; McCarthy and Snudden forthcoming). [1]

Zhang, Dufour and Galbraith (2016), Kohlscheen, Avalos and Schrimpf (2017) and Ca' Zorzi et al (2022) forecast point-sampled EERs. Froot and Ramadorai (2005) and Chen et al (2014), among other articles, forecast point-sampled bilateral RERs. [2]

For example, forecasts of period-average bilateral NERs that rely on disaggregated data are found to outperform the month-average no-change forecast almost universally when end-of-month or daily inputs are used to construct forecasts. In contrast, we find little evidence that such forecasts can improve upon the traditional random walk no-change forecast. [3]