The gift of foresight is one of the many talents that mankind wish to acquire. Since the dawn of time, man has continually sought visions of things to come to better themselves and prevent future catastrophes. Sadly mankind was not designed to have such powers and will have to make do with the gift of hindsight instead. Plato (1955) famously contemplated that necessity is the mother of invention. With hindsight as the vehicle for necessity, the concept of forecasting was conceived. Forecasting is described by Armstrong (2001) as the method of predicting a set of actual outcomes which has not yet happened by the analysis of trends of past and present data. Armstrong (2001) continues by stating that by the analysis of available information, forecasting is a principal way of managing risk. The concept is applied in diverse fields such as meteorology, seismology, costs and sales and financial trading to name a few. Although useful, forecasting is limited given that future outcomes are only probable and not definite (Armstrong 2001).
Forecasting is particularly useful in the trading of financial series (foreign exchange or forex). According to Brooks (1997) the foreign exchange market is the closest thing to the hypothetical efficient market, making it difficult to predict short and long term rates efficiently. Factors that make this financial trade efficient are the diversity and the number of traders and transactions in foreign exchange as well as the near instantaneous reflection of rates affected by externalities (Beechey, Gruen and Vickers 2000). Despite the difficulties, innovative forecasting methods were utilized to maximize returns as opposed to using conventional methods such as benchmarking and "buy and hold" stratagems. According to Dunis & Maio (2005), forecasting methodologies may be categorized into two distinct classes: fundamental (qualitative) and technical (quantitative) trading analysis.
The volume of foreign exchange rate trading has dramatically increased over the decades due to the risk being lower than that of trading securities and stocks (Beechey, Gruen and Vickers 2000). Due to the increased popularity, there is a more vociferous clamour for new methods and innovations have developed to be more efficient and realistic in the prediction of potential returns, risks and costs. Forecasting is deemed to be more enhanced by the application of two or more models, not just by having identified one particular model and utilizing it. (Dunis, Lindemann and Lisboa 2005).
Foreign exchange rate trading, like everything else in life, is not free. Transaction costs are levied for each transaction. So the higher the volume of trades, the higher the associated transaction cost increase accordingly (Hartmann 1999).
The focus of this dissertation is on the financial time series trade forecasting. A financial time series comprises a progression of measures over a predetermined time period. This dissertation uses the daily closing prices of US dollar against the Japanese Yen from January 2005 until January 2011.
The premise of forecasting is that future events and outcomes are predictable by studying past and present results and occurrences (Armstrong 2001). Despite difficulties in predicting possible outcomes for financial series, there are occurrences when patterns emerge and predictions are made easier (Keim 1983). The complexity of forecasting financial series arises out of the plethora of factors (both quantitative and qualitative) that come to play. To thrive within this trading scenario one must be opportunistic. Although Beechey, Gruen and Vickers' (2000) declaration that financial series trading is a near efficient market maybe somewhat disheartening, Beechey,et al. allude to the fact that opportunities may arise not through the market's inefficiencies but through other factors (e.g. natural disaster, geopolitical conflicts, etc.).
The primary intention of the dissertation is to identify the most efficient forecasting model in terms of accuracy when predicting future closing price outcomes for the foreign exchange time series of the US dollar against the Japanese Yen. The following points will be taken into account when assessing each forecasting methodology:
The trend of the USD/JPY financial time series
Characteristics of the financial time series (noisy or otherwise, stationary or non-stationary)
The factors that would affect USD/JPY financial time series
The appropriateness of the method(s) applied to the specific financial time series
The similarities and differences as well as strengths and limitations when weighed against other methods
To build the model, daily closing prices for USD/JPY from January 2006 to January 2011 have been obtained using Thomson Reuters 3000 Xtra. EViews is then used to acquire the necessary data for the autoregressive and moving average methods as well as the naive and ARMA models. On the other hand, the MS Excel is utilized to acquire the moving averages from the initial daily closing prices. All data and information are split into an in sample (period from 02/01/2006 to 05/05/2009) and an out sample (06/05/2009/ to 03/01/2011).
Each model has been assessed according to various methods of forecasting financial time series. Specifically, the following methods have been used for the assessment:
Autoregressive model or AR
Moving averages (MA)
Naive model of forecasting (NMF)
Autoregressive moving average model or ARMA
The dissertation then proceeds into great detail about the foreign exchange rate. This is then followed by the discussion of past and recent studies in the innovations involving the development of forecasting models for foreign exchange rates. Following this is a discussion of the methodologies for assessing the models. After the analyses, the paper then presents empirical evidence of the assessments. Finally, we conclude the implications of the empirical results as well as make recommendations to improve forecasting efficiency.
The foreign exchange rate
The foreign exchange rate is regarded as the value of a particular country's currency against that of another currency (Hartmann 1999). A currency's foreign exchange rate is considered to be one of many major indicators of a country's economic performance. For large multinational companies, it is imperative that efficient means of forecasting exchange rate movements is established for these organizations to be profitable (Armstrong 2001). But due to the complexities of the foreign exchange markets, there is no single perfect model that is applicable to all currency pairings and financial markets. In fact, Dunis, Laws and Sermpinis (2009) mentioned growing doubts about whether the employment of these models is worthwhile due to conflicting empirical evidence in forecasting.
Trading volumes within the foreign exchange markets have grown significantly over decades. There is prevalent belief that foreign exchange rate trading is less risky than other forms of financial trading (Beechey, Gruen and Vickers 2000). Despite the attractiveness of less risk, trading financial series is still unpredictable as there are an abundance of factors that affect exchange rates. Thus the need for the development of innovations in financial series forecasting exists to this very day.
Transactions are traded between two sides via over the counter market. Banks, hedge funds and financial management companies are just a few of many entities conducting foreign exchange trading. More recently, seminars and classes which cover the essentials for trading foreign exchange rates are more easily available to the public compared to past decades (Schneider 2011). Banks, financial management firms as well as large multinational companies are involved as they have an international portfolio because the transfer of liquid assets (cash) in addition to international imports and exports fall under the influence of foreign exchange rates. Market speculators on the other hand bet on, either for or against, particular changes in currency exchange rates. A prime example would be that market speculators have bets on the eventual and long overdue appreciation of the Yuan (Zuckerberg and Shah 2011), despite efforts of the Chinese government to keep an iron grip on keeping the currency at a low rate through market manipulation (Sanger and Wines 2010).
International economies, specifically those heavily dependent on trade as well as developing and industrial nations (Sanger and Wines 2001), are primarily concerned with foreign exchange rates. Due to this, the foreign exchange rate is one of the most efficient and most regulated markets in the world. Despite the regulations already in place, President Sarkozy has advocated the development of a system of coordinating global currencies as economic superpowers in the 1970s did (Sanger and Wines 2001) before the prevailing view of free market economy took hold.
A number of factors exist that effectively have significant bearing on exchange rates. These are:
Macroeconomic news and information
According to Laakkonen and Lanne (2009), macroeconomic news increases volatility in good times more than in bad times. The effects of positive news are not dependent on times. On the other hand, negative news seems to have more impact in good times than in bad times.
Interest rates
Changes in interest rates, apart from having a significant bearing on inflation, also have an influence on changes in exchange rates. As interest rates are changed, foreign exchange rates are directly influenced and have a near instantaneous change. Investors flock to countries with higher interest rates, thus increasing the country's currency value as the demand for the currency is increased in the short term at least (Dornbusch 1976).
Inflation
The currency value of a country with high inflation suffers regardless of the cause of inflation (too much supply of money in the market and trade deficit) (Dornbusch 1976). Foreign investors would be discouraged in investing in countries with high inflation due to the weakened buying powers of their dollar against their country's currency. A prime example would be Zimbabwe's hyperinflation (Wines 2007). Zimbabwe's years of economic decline as well as failed regulation enforcement have caused the catastrophic 10,000% hyperinflation, making life unbearable for an already impoverished population that they cannot afford to buy staple foods such as bread and eggs.
Gross Domestic Product
Growth in a country's gross domestic product can be seen as a double edged sword (Dornbusch 1976). Dornbusch (1976) argued that growth in GDP would prompt an appreciation in the value of local currency initially. As a result of this, the currency's purchasing power is improved - which then promotes an attractive climate for imports. If left unchecked, the economy would be in the middle of a trade deficit, which causes inflation. This in turn devalues the currency- a traditional cyclical relationship.
Supply of money
With a country's central bank's lowering of interest rates comes a risk of lowering the currency's value (Dua and Ranjan 2012). The intention of lowering interest rates is to increase economic activity, encouraging businesses as well as individuals to secure loans. As the market is swamped more and more with credit generated money, inflation is increased. This inflation, in turn, devalues the buying power of foreign currencies against that of the local's prompting some foreign investors to pull out of the country and look to other places to invest (Dornbusch 1976).
Ironically, the stated factors do not exhibit frequent motions as much as exchange rates do. It should be noted that changes in exchange rates are also affected by changes in the above factors in other countries.
Despite conflicting arguments between forecasting systems, there is consensus between researchers that predicting foreign exchange rates is indeed possible. Due to volatility (or noise) in short term data series, short term prediction is very difficult to undertake (Troutt and Elsaid 1996). This somewhat reinforces Gale's (2004) statements in defence of buy and hold strategy. Gale (2004) alluded to the fact that the premise of having a general idea of the future change in foreign exchange rate implies that future long term change is very possible to predict. But regardless of forecasting methods, traders are hopeful that innovative forecasting techniques would be developed to further improve forecasting in both short and long term financial time series.
Literature Review
In this section, we review past literature regarding forecasting models which are to be applied in the USD/JPY series. The section aims at re-evaluating past research conducted on techniques; specifically: the autoregressive model (AR), moving average and naive forecasting and the autoregressive moving average models (ARMA).
Dunis, Lindemann and Lisboa (2005) mentioned that the utilization of more than one kind of forecasting system to predict future time series outcomes would be better than just opting for one type of system. This statement is echoed by Armesto, Engemann and Owyang (2010). In their work with mixed frequency forecasting models, they argue that there is no such thing as a golden rule for both high and low frequency time series forecasting (Armesto, Engemann and Owyang 2010).
ArAmstrong (2001) has established that in order to develop an easy and straightforward method of random walk models, the naive strategy would particularly be most appropriately utilized. Despite the fact that foreign exchange rate changes are not completely random, Tyree and Long (1995) believe that random walk models (e.g. naive model) are more useful than neural network regression models (e.g. autoregressive and autoregressive moving average models). Aktas, Karan and Aydogan (2003) agreed with this sentiment. As a result of the research conducted by Aktas, Karan and Aydogan (2003), it was found, that of all models utilized (benchmark and neural network regression), and the naive method was able to beat the other models in terms of forecasting ability. Despite the effectiveness, Tyree and Long (1995) conceded that random walk models, and in turn the naive model, is not the best forecasting technique for predicting foreign exchange rate changes despite being more efficient in doing so than that of neural network regression models.
The common practice of buy and hold strategies have been put to question by Yochanan (et al 2001). Yochanan (et al 2001) concluded that, in the Tel Aviv 25 Index at least, the moving average model has been more efficient and reliable than the buy and hold strategy, even after having taken note of transaction costs. In Yochanan's (et al 2001) work, the returns generated by a number of transactions have mitigated the effects of volatility and costs. As the buy and hold strategy lacks the means of predicting future outcomes despite its premise, profit is not generated by holding on to a security but by trading (Yochanan et al 2001).
In the earlier days of trading securities, the buy and hold strategy was prevalent before the conception of forecasting methods for such have been developed (Armstrong 2001). The strategy itself is based on the basic premise that financial markets would generate a good rate of return in the longer term despite the market exhibiting periods of volatility (Gale 2001). If the foreign exchange market is truly efficient then this would imply that the price paid for buying and selling a currency pair is fair - no anomalies and therefore no difference in buying and selling rates. If this is the case then the traders will be at a loss at precisely the amount of transaction costs at a precise point in time. As Beechey, Gruen and Vickers (2000) implied that the foreign exchange rate market is the closest thing to being an efficient market, Beechey, Gruen and Vickers (2000) has alluded that there are inefficiencies that may be exploited for profit. Gale (2001) believes that trades can only be won due to the differences in rates due to market influences at a different time period. Gale (2001) encouraged that the number of transactions be done at a minimum (due to transaction costs and commissions) and only if the trader is satisfied with the profit (sell price less trading costs at a different point in time) he will achieve. High amount of transactions only benefit the broker (Gale 2001).
Pruitt and White's (1988) introduced the idea to use technical analyses such as the CRISMA (cumulative volume, relative strength, moving average) in predicting time series outcomes. Despite Pruitt and White's (1998) belief that CRISMA fared better than random walk models, the system has limitations. Goodacre and Kohn-Speyer (2001) found CRISMA, despite the multi nature component of the system, is essentially a moving average trading rule. The limitations of the CRISMA system were noticed by Goodacre and Kohn-Speyer (2001) when factors such as market movement, risk and volatility, and transaction costs are taken into account. Dunis and Miao (2005) agree that the moving average model does well in the short term if there is low to negligible volatility in price levels. But despite the moving average trading rule's limitations, Egelkraut, Garcia and Sherrick (2007) implied that in terms of risk at least that implied forward volatilities of corn options proved to be unbiased and nearly as effective as the moving average system and significantly superior to naive forecasting (on early year options at least). Unfortunately, Egelkraut, Garcia and Sherrick's (2007) model have exhibited relatively less forecasting power on later year options due to the reduction in trading volume. Despite the moving average trading rule's limitations, Egelkraut, Garcia and Sherrick (2007) support Yochanan's (et al 2001) that the forecasting method significantly outperforms buy and hold strategies in terms of returns in certain circumstances at least. Yochanan et al (2001) maintains that the returns at the very least compensate for the volatility and transaction costs of the financial time series.
On a different note, the autoregressive model is effective enough in itself to perform better than other more sophisticated forecasting methods (Pesaran and Timmermann 2005) even accounting for external volatilities and macroeconomic variations (Marcellino, Stock and Watson 2006). Having integrated auto regression with generalised random coefficients along with Owen's (2001) empirical likelihood, Zhao and Wang (2011) found that empirical likelihood is more effective than the conventional approximation based methods of forecasting. Owen (2001) believes that since empirical likelihood, as it deals more with possibilities, has inherent advantages over models that utilize sampled data. Although there are recent innovations, developments in forecasting methods superior to that of the autoregressive models are few and far between as new developments do not necessarily mean they are better and more efficient (Pesaran and Timmermann 2005).
Another forecasting method covered by this dissertation is the autoregressive moving average. Box, Jenkins and Reinsel (1994) believe apart from the effectiveness of autoregressive moving average method, the cost of utilizing the method is comparatively less than that of other systems. There is strong support for the autoregressive moving average forecasting method. Since the forecasting method has been around for ages, the autoregressive moving average model along with the naive method are used as the standard for which newer forecasting techniques are compared against (Dunis, Laws and Naim 2005). Box, Jenkins and Reinsel (1994) further add that the ARMA model is quite efficient in short term forecasting. Pourahmadi (2007) adds that the Box-Jenkins forecasting model provides accurate results in out sample data set, which outperforms random walk forecasting methods. Pourahmadi (2007) alludes that the ARMA has the capability to outperform non-linear models. Falk (2011) explained that the ARMA model operates on the premise that the future outcomes are reflected on past outcomes of the time series. Although the ARMA model has still some distance to cover in terms of perfection, Falk (2011) admits that the ARMA (particularly the Box-Jenkins method) have been a useful forecasting model which have existed for decades; though Falk (2011) concedes that ARMA lacks the means for an efficient conditional forecast.
Armstrong (2001) believes that the wide use of ARMA and ARMA based models is down to the easy computation and interpretation of the results. In addition, Box, Jenkins and Reinsel (1994) state that ARMA is capable of generating profits even after transaction costs are considered. In economic terms, Insel, Sualp and Karakas (2010) found ARMA outperformed neural network regression forecasting models when predicting future real gross domestic product.
According to Marcellino, Stock and Watson (2006), the autoregressive moving average model (even the widely used Box-Jenkins model (Box, Jenkins and Reinsel 2004 cited by Marcellino, Stock and Watson 2006)) is limited by its inability to take into account the random movements (also known as noise) in financial time series. Non-linear variants of ARMA model have been proposed to counteract that particular limitation (Pourahmadi 2007). Dunis, Laws and Sermpinis (2012) have doubts about the ARMA model's ability altogethers, citing that neural network regression forecasting have performed better than traditional forecasting models (e.g. ARMA, moving average). This sentiment is echoed by Chortarea, Jiang and Nankervis (2011) citing that artificial neural networks are adaptive, non-parametric, and considers noise in their forecasts - basically the things that ARMA is not capable of handling. (Marcellino, Stock and Watson 2006). Insel, Sualp and Karakas (2010) established that neural network models outperformed ARMA in predicting inflation, interest and exchange rates.
