“Short-term foreign exchange forecasting: decision making based on expert polls”

The paper aims to analyze the decision making based on expert polls for short-term foreign exchange (FX) forecasting from the viewpoint of the economic behavior theory. The paper offers the assessment of the problem of decision making for forecasting and investment into foreign currency. This study analyzes the relative accuracy of expert polls and forecasts, based on historical data, in the prediction of the most liquid currency pairs (EUR/USD, USD/JPY, GBP/USD) as well as USD/RUB currency pair on time horizons 1, 2, 6, and 12 months. Observation period lasted from January 2018 to January 2019. For EUR/USD (56-62 experts), the polls were more accurate than historical simulations. For GBP/USD (28-70 experts), historical simulations were more accurate than polls. For USD/JPY and USD/RUB, historical simulations are better earlier, while polls are slightly better later. The main conclusion is that EUR/USD historical modeling is usually less accurate on the horizon more than half a year as compared with expert polls for making the decisions about the future exchange rate.


INTRODUCTION
Many market participants are interested in being able to predict further exchange rate direction. Whether it is a large company or an individual, a currency forecast is significant for minimizing the risks and increasing the profits. The paper describes the expert polls for shortterm currency forecasting.
Purchasing Power Parity (PPP) principle is based on the theoretical law of one price, according to which identical products in different countries should have the same price. For example, according to this rule, a pencil in Canada should cost as much as the same pencil in the United States, taking into account the exchange rate and excluding the costs of exchange and transportation. Graefe (2018) proposed a structured approach to combining the forecasts based on various types of methods. His approach is correct in predicting the majority of possible political events.
This article intends to compare the accuracy of expert research and mathematical modeling for forecasting the currency exchange rates.
There are the papers about smoothing fluctuations in the accuracy of expert research and mathematical modeling, which is consistent with the professional opinion.

LITERATURE REVIEW
The methodological basis of the analysis is the con- What method is better to use for risk choice: expert polls or historical modeling? The last research findings regarding the reliability of expert surveys can give more accurate results for the forecast, which includes more information (Cooper & Priestly, 2009).
For example, in forecasting the political events, the expert judgment was used much earlier than mathematical models (Kernell, 2000; Silver, 2017). However, FX rate forecasters do not know much about the relative errors of expert surveys for different periods (Na. Morozko, Ni. Morozko, & Didenko, 2018a, 2018b).
Studies on the method of expert polls in various fields of application show that expert knowledge is really limited in forecasting under un-certainty. Expert estimates of exchange rates are sometimes even less accurate than simple statistical models, for example, random walk models (Armstrong, 1980). Macroeconomic experts can use extensive mathematical, statistical approaches as well as empirical data. For example, many studies have shown that surveys tend to reduce the forecast accuracy when lengthening the forecast horizon (Graefe, 2018), and as the forecast date approaches, in the absence of strong shock fluctuations in the foreign exchange market, many surveys become more accurate. In the last century, the researchers proved it for many types of political events (Riker, 1982).
Forecasts are heavily influenced by structural factors as well as the state of the economy, cyclical correlation of market indicators, changes in the availability and attractiveness of currency pairs for speculators, the growth of algorithmic trading, the degree of intervention by regulators and the frequency of significant events (Graefe, Armstrong, Jones, & Cuzán, 2014). These factors can be included in the mathematical model. However, aggregated results are subject to different types of errors (Biemer, 2010;Groves & Lyberg, 2010).
As an example, polls that were devoted to presidential elections in the United States for a week or even a month were the least accurate (Shirani-Mehr, Rothschild, Goel, & Gelman, 2018).
The study presents the empirical data for a reasonable answer to the question about the method of expert estimates, and the relative accuracy of the analysis of expert estimates based on the forecasts of EUR/USD, USD/JPY, GBP/USD and USD/RUB exchange rates for various short-term periods. Many researchers use the same method for oil price forecasting ( At the same time, they usually forget about expert polls as a reliable method of forecasting. However, this article will focus on comparing the method of expert polls with the mathematical method of forecasting the exchange rates (Lopatin, 2019a; Meynkhard, 2019).
In addition, several models are based on the effect of memory, when the current price is the basis for forecasting future prices (Mikhaylov, 2018a; Graefe, Kјchenhoff, Stierle, & Riedl, 2015).

METHODS
Expert judgment forecasts regarding the dynamics of currency pairs EUR/USD (Euro/US dollar), USD/JPY (US dollar/Yen), GBP/USD (British Pound/US dollar) and USD/RUB (US dollar/ Russian ruble) from January 31, 2018 to January 31, 2019 were collected over time horizons of 1, 3, 6, 12 months from Thomson Reuters.
Thomson Reuters periodically interviews the representatives of investment banks and research centers relative to the target level of exchange rates. Ratings of all participants in these 16 separate polls are presented in Appendix.
The expert group consists of financial analysts and researchers. The composition of the participants in each poll varies. The number of private traders have ranged from 28 to 70 people. Some experts participated only in the polls regarding the dynamics of the EUR/USD pair; others participated in the polls on 4 four currency pairs.
The standard deviation of estimates is different and varies on average in the range from 2 to 7 percent. The average number of experts for another round of polls is 54 people.
Usually, when constructing an econometric model, values from economic theory are used (Dorofeyev, 2018). However, any variable that has a strong influence on the exchange rate can be added to the calculations. This econometric model is as follows: term financial market rates for corporate clients.
EFA uses the mathematical and statistical methods of prediction based on historical data, which take into account the following factors with varying level of import prices: cyclic recurrence, correlation of market indicators, changes in the availability and attractiveness of the instrument for speculators, electronic and algorithmic trading growth, regulatory intervention risk, and frequency of significant events over time like it was wrote before (Mikhaylov, 2018b; Mikhaylov, 2019).

RESULTS
The results of the forecast based on historical data were compared with empirical data. Then, the accuracy of the above model was compared with empirical indicators in the same way. As a result, a comparison was made of the average errors of these methods.
For many currency pairs and horizons, there were no expert polls of the study, so the number of respondents for each poll varies. Expert polls inaccuracy has been observed due to comparing the errors of experts on different time horizons. Figure 1 shows a comparison of the average absolute error (MAE) of forecasts obtained by the method of expert estimates and the method of mathematical modeling for EUR/USD over four time horizons.
The results were mixed for all four MAE currency pairs of expert forecast (in percent) for 1 monthfrom 1.1 to 2.2, for 3 months -from 2.0 to 3.8, for 6 months -from 2.6 to 9.2, for 12 months -from 3.4 to 6.9.
MAE mathematical model EFA for 1 month is from 0.3 to 1.15, for 3 months -from 1.1 to 2.8, for 6 months -from 3.9 to 8.7, for 12 months -from 4.2 to 8.5.
The method of expert estimates gives a higher accuracy when forecasting the exchange rates for a period of 1 year and more. This, of course, does not mean that mathematical models should be ignored when forecasting the exchange rates for a period of 1 year and more.
Attempting to find a better prediction method is usually not warranted. The reason for the inaccuracy of the application of the method of expert assessments is that it is necessary to form a circle of experts more purposefully and apply a ranking of expert evaluations depending on historical accuracy.

DISCUSSION
The analysis presented in this article is based on small selection of expert forecasts (N = 28 to 70), collected over 1-year period. Further research on various types of financial instruments will help to learn more about the relative accuracy of the method of expert estimates and the shortcomings of expert judgment in forecasting of the exchange rate ( Combining the expert polls may reduce the expert method error. This is a topic for future research. Experts in any field should refrain from attention the specifics of the situation (Lopatin, 2019b; Meynkhard, 2020). In addition, they should be conservative about large changes and take into account all the accumulated knowledge of the situation (Armstrong et al., 2015). A structured approach to combining the forecasts from different methods using different strategies can be a solution to this problem.

CONCLUSION
The authors found out that EUR/USD historical modeling is usually less accurate on the horizon more than half a year as compared with expert polls for making the decisions about the future exchange rate. If one uses a simple mean, then the combined forecast will be more accurate than the average error of the individual forecast as it was found by Armstrong (2001 Secondly, it is extremely difficult in most practical situations to find out in advance which forecast will be more accurate because historical accuracy is not a guarantee of future accuracy. It proves the findings in the papers (Na. Morozko, Ni. Morozko, & Didenko, 2018c, 2018d).
The paper proved the studies, which found a negative relationship between the historical accuracy of expert polls (Graefe et al., 2018) and mathematical models (Graefe et al., 2015). The results showed that the average of the two forecasts is more accurate than a separate forecast if the error is less accurate than the forecast and does not exceed error more than in three times. As noted above, the results regarding the reliability of expert surveys can give more accurate results for the forecast, which includes more information. Therefore, in order to improve the accuracy of expert forecasts, we must look for information that these methods could miss.