International Journal of Multidisciplinary Research and Development

International Journal of Multidisciplinary Research and Development

Online ISSN: 2349-4182, Print ISSN: 2349-5979
Impact Factor: RJIF 5.72
www.allsubjectjournal.com
Volume 5; Issue 5; May 2018; Page No. 44-49

Research review on time series forecasting of gold price movement

Wesley Goh Yi Yong, Suliadi Sufahani*
Department of Mathematics and Statistics, Faculty of Applied Sciences and Technology, Universiti Tun Hussein Onn Malaysia,
Johor, Malaysia

Abstract
This paper propose different techniques used in forecasting on time series analysis. Various techniques had been used in the field
of time series forecasting from the traditional Box-Jenkins approach to the most popular neural network technique nowadays.
However, there is no specifically the best method to deal with time series forecasting as the application of different time series
forecasting methods has their own requirements and restrictions. In determining the movement of gold price, there are a lot of
different methods being implemented by various authors to propose their models. Various time series forecasting method have
been discussed in this paper which consists of several journal articles that related to gold price and some of the data mining
techniques in time series forecasting retrieved from Google Scholar in this review.

Keywords: time series, box-jenkins, forecasting, gold price

1. Introduction analysis. Big data could often be heard from various industries
A precious metal is a rare naturally occurring, metallic due to the readily available data that could be access by
chemical element of high economic value; precious metals anyone to perform their business analysis which could reduce
were important as currency in the past but are now regarded the cost and resources on data collection. Hence, data mining
mainly as investment and industrial commodities. Historically, technique is getting more popular as it used to extract the
gold remains resilient and performed better than the S&P 500 needed information from a big pile of data for specific
index during the 1973-1974 market crash, the oil crisis of the analysis.
mid-1980s,the crisis of the new millennium combined with the Traditional way of doing time series forecasting is popular by
tax bubble in the 1990s and the global financial crisis in 2008 using the Box-Jenkins method. However it can only deal with
(The Edge Malaysia, 2017). Hence, gold is known as a safe- univariate time series analysis which includes only the
haven asset as it has an inverse correlation with the US dollar. dependent variable itself. Data mining allow the addition of
Investors tend to preserve the value of their asset when the various independent variables to help to forecast the
economic is uncertain by investing in gold since it does not movement of dependent variable and allow us to understand
have heavy liability or unpredictability (Shafiee and Topal the indicators behind it. Hence, it is value added to any
2010) [2, 12]. Gold act as insurance for investors as it has low businesses.
correlation with most assets thus can be used to reduce Most research only concern about the movement of gold price
portfolio volatility and minimize losses during extreme market with macroeconomic variables such as GDP growth rate,
conditions. inflation rate, and consumer price index. There is no specific
The movement of gold price is concerned by investors, method in data mining used to forecast the gold price with
financial specialist, and government and to who gain their other precious metals. In the study of volatilities of four
profit through gold trading. The impact to the movement of precious metals which are gold, silver, platinum, and
gold prices is influenced by a lot of factors which include palladium prices, the authors claimed that precious metals are
macroeconomic variables and business cycles. However, too distinct to be considered a single asset class, or
related studies also suggested that there is high correlation represented by a single index (Batten, Ciner et al. 2010) [5].
between gold price and crude oil price (Chen and Fang 2013)
[3, 18]. Besides, other author’s result shows that there is a
2. Literature Review
significant relationship between the precious metals and oil, The advancement of technology had boosted the collection
which include gold, platinum and silver. Additional research and storing of data to a whole new level. Nowadays, data can
also indicated that oil impacts gold more than gold impacts oil be easily collected by various gadgets which connect to a
(Gabralla, Jammazi et al. 2013) [4]. computer system. Since data can be easily collected as the
The prices of precious metals varies from day to day, a lot of data storage is getting cheaper, data can also be retrieved and
research had been done by researchers using different obtained with fingertips through internet by using
techniques. The traditional way of doing statistical analysis computational software. In this situation, it brings up a whole
could be more complex and difficult as of today, the data new challenge in the field of data analysis as traditional
storage is getting cheaper and lots of data can be collected for statistical analysis could not be able to perform efficiently as

44
International Journal of Multidisciplinary Research and Development

the data sets get larger because it will take a whole lot more of research and a SARIMA model of order (0, 1, 1) (0, 1, 1) 12 is
time and energy for a human being to do so. Hence, it is the model chosen by the authors. The authors suggested that
required to have both computational skills and analytical skills the Bayesian information criterion (BIC) is preferred to
in order to improve the performance of data analysis. Akaike’s information criterion (AIC) for comparing different
Big data is being more and more often to be heard nowadays. models as BIC will penalize the addition of extra parameters
Big data usually includes data sets that are so large and more severely than AIC does.
complex which the sizes are beyond the ability of standard A comparison has been made by (Shafiee and Topal 2010) [2,
12] to compare the accuracy of forecasting model of gold price
statistical software tools to capture, curate, manage, and
process data within a tolerable elapsed time (Snijders, Matzat by using reverting jump and dip diffusion model, and ARIMA
et al. 2012) [6]. Big data requires a set of techniques and model. The study focus on the gold price trends over the past
technologies with new forms of integration to reveal insights 40 years from January 1968 to December 1988 and analysis
from datasets that are diverse, complex, and of a massive scale was conducted by using monthly data to forecast the gold
(Hashem, Yaqoob et al. 2015) [7]. price for the next 10 years. The unit root test Augmented
Data mining is a process of extracting information from a Dickey Fuller (ADF) test was conducted by the authors and
huge data set and transform it into understandable structure for the gold price was found non-stationary and the first
future use. Data mining is the interdepartmental field of study differential transformation was applied. The reverting jump
that are able to discover patterns and models from a large and dip diffusion model was claimed to perform better than
amount of information stored in data warehouse (Gaber, ARIMA model through comparison of RMSE and MAE of
Zaslavsky et al. 2005) [8]. Data mining is also refers to the both model. The new model proposed by the author is as
extracting or mining knowledge from large data stores or sets follows which consider the drift, diffusion components, and
(Al-Radaideh, Assaf et al. 2013) [9]. However, data mining is jump and dip period of the gold price.
not something new as various author had been using the
technique in their research for past decade. X t   1   2 t   2 (1      3 D 1   4 D 2  u t (2)

2.1 Box-Jenkins Approach Where

The Box-Jenkins Auto-Regression Integrated Moving
 2 t = First component or drift
Average (ARIMA) is one of the most sophisticated techniques
in time series forecasting. A study of gold price forecasting in  2 (1    =Second component or the range of random
Indian was conducted by (Guha and Bandyopadhyay 2016) movement
[10]. The study is based on the historical data of monthly gold
 3 D 1   4 D 2 = Third component or jump or dip
price from November 2003 to January 2014 and ARIMA
model was applied to forecast fluctuation of the gold price. (Khan 2013) [13] proposed ARIMA model for accuracy
The price of gold in this study is valued in Rs per 10 grams. checking on the prediction of gold price. The samples consist
Three model parameters which are Autoregressive (AR) of gold price data in USD per ounce from 2 January 2003 to 1
indicating weighted moving average over past observations, March 2012. Most of the macro economy data are non-
integrated (I) indicating linear trends or polynomial trend, and stationary are claimed by the authors. Augmented Dickey
Moving Average (MA) indicating weighted moving average Fuller (ADF) test is constructed to determine the stationarity
over past errors are combined to form ARIMA (p, d, q) where: of the series, gold price is said to be non-stationary and only
be found stationary after first difference is applied. Akaike
p = order of autocorrelation Information Criteria (AIC) and Schwarz Information Criteria
d = order of integration (differencing) (SIC) are used to compare the ARIMA models. Root Mean
q = order of moving averages Square Error (RMSE), Mean Absolute Error (MAE), Mean
Absolute Percentage Error (MAPE) is used to evaluate the
The ARIMA model can also be expressed as follow: accuracy of the models. It was found that ARIMA (0, 1, 1) is
the best model selected to predict the future gold price.
(1) (Yaziz, Azizan et al. 2013) [14] proposed a hybrid method of
ARIMA-GARCH modelling to forecast the gold price by
According to the researcher, ARIMA (1, 1, 1) is the best using a total of 40 daily gold price. ARIMA is said to be
model chosen for prediction as all the significant values are powerful and flexible in forecasting but it is unable to deal
less than 0.05. Furthermore, Root Mean Square Error with volatility and nonlinearity in a data set. GARCH is used
(RMSE), Mean Absolute percentage Error (MAPE), Mean to model the nonlinear patterns of the residuals. The authors
Absolute Error (MAE), Bayesian Information Criterion (BIC), claim that the gold price is non-stationary by conducting
and Lungs Box Q statistics are used as the statistical measure Augmented Dickey-Fuller (ADF) test which the first
of the model. The authors also claimed that the gold price differencing is needed to make the series stationary. The
shows high first order auto-correlation through Durbin- results show that ARIMA-GARCH model outperform several
Watson (DW) test. models proposed by other authors.
Another Box-Jenkins analysis was conducted by (Faraway and
Chatfield 1998) [11] which is the analysis of the well-known 2.2 Arimax model
airline data. A total of 132 monthly data were used for the The paper by (Wongdhamma 2016) [15] proposed a comparison

45
International Journal of Multidisciplinary Research and Development

of ARIMA and ARIMAX model on time series forecasting. affecting the gold price include Bureau future (CRB) index,
The authors implement wavelet analysis (WA) to decompose USD/Euro Foreign Exchange Rate EUROUSD, inflation rate
the trend, seasonality, process variations, and noises of time and money supply (M1).
series. The decomposed data is then used as exogenous
variables on constructing an ARIMA model. The results show 2.4 Support vector regression
that WA are good regressors in ARIMAX which are able to Support Vector Regression (SVR) is used as a learning
capture nonlinear patterns well. ARIMAX-WD performs algorithm to predict the gold price by understanding the
better than ARIMA in long term and nonlinear time-series pattern of the historical gold price (Navin 2013) [20]. An
forecasting application. assumption that made by the authors is that the historical data
incorporate all the external effects behaviour. The data is
2.3 Regression transformed into generic dataset followed by cross validation
A linear regression approach is proposed by (Gharehchopogh, process to feed them as inputs into support vector regression
Bonab et al. 2013) [16] in the case study of predicting the S&P model. The type of kernel and special parameters are selected
500 index. The authors selected the volume as the dependent and applied to the model, the process was repeated until the
variables and average daily price as independent variables. accuracy of the model is good enough. The model is then used
The results shows that the accuracy is satisfied as there is to test with the out of sample dataset for gold price prediction.
61.35% of similarities observed when compared to the out
samples. However, the model has a low R-squared value, 2.5 Neural network
which indicates that the model does not contribute much for Neural network and genetic algorithm were implemented by
the prediction of volume. The model could be improved by (Mirmirani and Li 2004) [21] to analyse the movement of gold
considering other factors that could affect the daily volume price. Application of Genetic Algorithm along with neural
trade via variable selection through data mining. networks is said to improve the learning power and robustness
Multiple regression method is proposed by (Baker and Van of the system. The analysis was done by using daily cash
Tassel 1985) [17] based on fundamentalist approach to prices of gold from 12/31/1974 to 12/31/1998 which consist of
determine the variables that affecting the gold price and model 6008 data points. The results show that the historical prices of
building for long term gold price prediction. The changes in gold strongly affect the gold price in future. The authors also
commodity price index, changes in the value of US dollars, claimed that there is short-term time dependence of gold price
and future inflation rate is found to be affecting the monthly movements with the time lag of 36 days.
change of gold price. The results suggested that the price of Neural network has also be applied by (Grudnitski and Osburn
gold can be expected to rise if there is a general increase in 1993) [22] for forecasting S&P and gold futures price. Other
commodity prices. than focusing on the price trend itself, the authors also
Another multiple regression approach is conducted by (Chen consider general economic conditions and traders’
and Fang 2013) [3, 18]. The return on investment in gold has expectations about the market in future. The authors conclude
higher correlation with other commodities (aluminium, that neural network can be applied to forecast price changes of
petroleum, zinc, and etc.) than the business cycle and the main the markets with four factors considered which are:
macroeconomic variables. The authors aimed to explore the 1. Derivation of the parameters of the networks relies
long term determinants of gold price by considering the US heavily on published commitments.
money supply M2, CRB index, US dollar index, the Dow 2. Selection of futures whose price are relatively insulated
Jones Industrial Average, and the SPDR holdings as the from natural phenomena.
exogenous variables in building the regression model. The 3. Appropriate length for the training period.
study concludes that M2 of US money supply is the most 4. Trading selectively rather than in every period.
affecting factor with gold price which they are highly positive
correlated. CRB index of inflation levels also shows high In constructing the neural network model, the researchers may
positive correlation with the gold price. However, the model face a hard time on estimating the weights to be used in the
showed that gold price has low negative correlation with the model. It is not an easy task as the number of weights may be
Dow Jones Industrial Average, low positive correlation with large and the objective function may have local minima
the US dollar index and SPDR gold trust positions. (Faraway and Chatfield 1998) [11]. The application of neural
According to experts, there are a lot of economic factors network model is still a doubt for researchers as some study
which could affect the movement of gold price. Multiple had proposed that the neural network model is no better than
linear regression method has been used to predict the gold other methods. Neural networks has several drawbacks which
price based on several economic factors such as inflation, include excessive training times, difficulty in obtaining and
currency price movements and others (Ismail, Yahya et al. replicating a stable solution, overfitting of model, and the
2009) [19]. Stepwise regression is used to remove the black box nature of the solutions. Although, neural network
correlation between variables. At the same time, the authors modelling is non parametric analysis which the process can be
proposed that the effects of significant lag in the cause-and- completely automated on a computer, Black boxes can
effect process in order to improve the performance of the sometimes give silly results.
model. Durbin-Watson statistics is used to check whether
autocorrelation exist in the error terms. Error mean square 2.6 Other Methods in Time Series Forecasting
MSE and mean squared prediction error MSPR are used to Ensemble machine learning was proposed by (Gabralla,
evaluate the model accuracy. From the study, the factors Jammazi et al. 2013) [4] to study the forecasting performance

46
International Journal of Multidisciplinary Research and Development

of daily WTI crude oil price with consideration of a number of the market index will go up, go down or stay in the next 1, 6,
influential features as inputs. Best-first search and genetic and 21 days. In the mean times, the authors claimed that
algorithm were used by the authors in selection of variables although NN models have high accuracy, many of them does
that influences the crude oil price. According to the authors, it not put into practice due to inability of neural networks to
is possible to use commercial material instead of gold price explain its reasoning. For the results, D-Tree with 3 classes is
for crude oil price prediction. Their results show that the considered as the best model in the study. The authors applied
ensemble method performed better than SMOReg and IBL chart patterns discovery to drill down into the model to find
using only three attributes. valuable patterns which could be easily used in decision
Another approach of using data mining classification in stock making.
market prediction is proposed by (Paliyawan 2006) [23]. The
author performs the analysis by using four different 3. Discussion
approaches which are neural network, decision tree, naïve This section highlights the key point that expressed by the
Bayes, and k-nearest neighbours. The authors predict whether authors of previous study.

Table 1: Summary of time series forecasting

Study Technique Research Problems
Qasem A. Al-Radaideh, To analyse the historical data of stocks using decision tree technique to help investors to decide
Decision tree
Adel Abu Assaf & Eman when to buy or sell their stocks. The variables selected for the study involve only the historical
classifier
Alnagi (2013) [9] price of the stock which only represents the market behaviour.
To implement forecasting of gold price through time-series ARIMA model. The ARIMA model
Traditional Box-
Banhi Guha & Gautam is selected through comparing AIC and SIC. The forecasting performance model selected is
Jenkins methods
Bandyopadhyay (2016) compared by using MAE, MAPE, RMSE, ARIMA model is suitable for short run only, it can
[10] Khan (2013) [13] (ARIMA and
only predict by using univariate data set. ARIMA is unable to deal with volatility and
SARIMA)
nonlinearity in a data set.
To fit variety of neural network models to the well-known airline data and compares the
resulting forecast with Box-Jenkins and Holt-Winters method. Neural network model were
Julian Faraway & Chris Neural network
found to be possible that the fitting routine may not converge or may converge to a local
Chatfield (1998) [11] and ARIMA
minimum. Besides, NN models that fit too well may give poor out of sample forecast. The
complicated black box structure of NN model is hard to understand and interpret.
Decision tree and Develop a forecasting model for predicting and forecasting gold prices based on past historical
Navin & Dr. G. Vadivu
support vector prices of gold. External factors are not being considered as it assume the historical prices
(2013) [20]
regression (SVR) incorporate all the external behaviour.
To compare the accuracy of forecasting model of gold price by using reverting jump and dip
Reverting jump
Shafiee and Topal (2010) diffusion model, and ARIMA model. The new model proposed by the author consider the drift,
[2, 12] and dip diffusion
diffusion components, and jump and dip period of the gold price which performed better than
model
ARIMA model.
Proposed a comparison of ARIMA and ARIMAX model on time series forecasting.
Wavelet analysis (WA) is applied to decompose the trend, seasonality, process variations, and
Wongdhamma (2016) [15] ARIMAX
noises of time series. The decomposed data is then used as exogenous variables on constructing
an ARIMA model.
To predict the S&P 500 index using linear regression by using volume as dependent variable and
Gharehchopogh, Bonab Fuzzy Linear
average daily price as independent variable. The model shows low R-squared value, other
et al. (2013) [16] Programming
factors could be identified to improve the model contribution.
Baker and Van Tassel
To build a model to predict gold price by considering exogenous factors.
(1985) [17] Ismail, Yahya Multiple linear
Several economic factors including commodity price index have been found to affect the gold
et al. (2009) [19] Chen and regression
price movement.
Fang (2013) [3, 18]
To predict the gold price by understanding the pattern of the historical gold price.
Support vector
Navin (2013) [20] Selection of kernel function is complicated. There is no benchmark for SVR approach. There is
regression
no optimal method on free parameter selection.
To apply Genetic algorithm accompanied with neural networks to find an optimal net
Neural network
Mirmirani and Li (2004) architecture in model selection in time series analysis. The results show that the historical prices
[21] and Genetic
of gold strongly affect the gold price in future. The authors also claimed that there is short-term
algorithm
time dependence of gold price movements.
To forecast S&P and gold futures price. The authors conclude that neural network can be applied
to forecast price changes of the markets with four factors considered which are: Derivation of
Grudnitski and Osburn
Neural network the parameters of the networks relies heavily on published commitments. Selection of futures
(1993) [22]
whose price are relatively insulated from natural phenomena. Appropriate length for the training
period. Trading selectively rather than in every period.
Gabralla, Jammazi et al. Ensemble To study the forecasting performance of daily WTI crude oil price with consideration of a
(2013) [4] machine learning number of influential features as inputs.
Neural network, To apply data mining classification in stock market prediction.
Paliyawan (2006) [23]
decision tree, The results show D-Tree with 3 classes is the best model in the study and chart patterns

47
International Journal of Multidisciplinary Research and Development

naïve Bayes, and discovery can be applied to drill down into the model to find valuable patterns.
k-nearest Although NN models have high accuracy, many of them does not put into practice due to
neighbours inability of neural networks to explain its reasoning.
To propose a hybrid method of ARIMA-GARCH modelling to forecast the gold price by using a
Yaziz, Azizan et al.
ARIMA-GARCH total of 40 daily gold price. ARIMA and GARCH are able to complement each other to achieve
(2013) [14]
a better forecasting model than traditional ARIMA model.

Previous studies showed that several approaches have been machine learning. Computing, Electrical and Electronics
implemented in the time series forecasting in economic and Engineering (ICCEEE), International Conference on,
financial time series analysis. However, there is no IEEE, 2013.
specifically the best method to deal with time series 5. Batten JA, et al. The macroeconomic determinants of
forecasting as the application of different time series volatility in precious metals markets. Resources policy.
forecasting methods has their own requirements and 2010; 35(2):65-71.
restrictions. Therefore, further improvement of traditional 6. Snijders C, et al. Big Data: big gaps of knowledge in the
Box-Jenkins analysis in time series analysis could be field of internet science. International Journal of Internet
improved. Table 2 describes the contributions that could be Science. 2012; 7(1):1-5.
recognized in the future research. 7. Hashem IAT, et al. The rise of big data on cloud
computing: Review and open research issues. Information
Table 2: Contributions of the research Systems. 2015; 47:98-115.
To forecast the movement of gold price with the consideration of 8. Gaber MM, et al. Mining data streams: a review. ACM
the price of other precious metals. Sigmod Record. 2005; 34(2):18-26.
To identify the relationship of gold to other precious metals. 9. Al-Radaideh QA, et al. Predicting stock prices using data
To compare the accuracy of forecasting model using data mining mining techniques. The International Arab Conference on
forecasting technique and traditional Box-Jenkins approach. Information Technology (ACIT’2013), 2013.
10. Guha B, Bandyopadhyay G. Gold price forecasting using
4. Conclusion ARIMA model. Journal of Advanced Management
The paper presents a review of literature concerned with the Science, 2016, 4(2).
variety of methods that had been applied in the time series 11. Faraway J, Chatfield C. Time series forecasting with
forecasting in the field of economics and finance. The neural networks: a comparative study using the air line
growing of recorded data is faster than ever before nowadays data. Journal of the Royal Statistical Society: Series C
which leads to huge data sets. Traditional statistical analysis (Applied Statistics). 1998; 47(2):231-250.
methods may need to modify in order to improve the 12. Shafiee S, Topal E. An overview of global gold market
performances and speed of conducting analysis. In short, and gold price forecasting. Resources policy. 2010;
statistical knowledge and computational skills have to be 35(3):178-189.
combined in order boost the performance of statistical 13. Khan MMA. Forecasting of gold prices (Box Jenkins
analysis. Further study could be done by considering the other approach). International Journal of Emerging Technology
industrial commodities as exogenous variables that that could and Advanced Engineering. 2013; 3(3):662-670.
have driven the gold price. Data mining techniques that could 14. Yaziz S, et al. The performance of hybrid ARIMA-
complement with the traditional statistical analysis should be GARCH modeling in forecasting gold price. 20th
considered in time series analysis and forecasting in future International Congress on Modelling and Simulation,
work. Adelaide, 2013.
15. Wongdhamma W. Upgrade from ARIMA to ARIMAX to
5. Acknowledgement Improve Forecasting Accuracy of Nonlinear Time-Series:
We would like to say thank you to Universiti Tun Hussein Create Your Own Exogenous Variables Using Wavelet
Onn Malaysia (UTHM) and Office for Research, Innovation, Analysis, 2016.
Commercialization and Consultancy Management (ORICC), 16. Gharehchopogh FS, et al. A linear regression approach to
UTHM for kindly providing us with the internal funding TIER prediction of stock market trading volume: a case study.
1for proofreading (Vot U903). International Journal of Managing Value and Supply
Chains. 2013; 4(3):25.
6. References 17. Baker SA, Van Tassel RC. Forecasting the price of gold:
1. The Edge Markets. Cover Story: Backed By Solid Gold. A fundamentalist approach. Atlantic Economic Journal.
Retrieved on 2017-2018, from http://www.theedgemar 1985; 13(4):43-51.
kets.com/article/cover-story-backed-solid-gold 18. Chen X, Fang Y. Enterprise systems in financial sector–
2. Shafiee S, Topal E. An overview of global gold market an application in precious metal trading forecasting.
and gold price forecasting. Resources policy. 2010; Enterprise Information Systems. 2013; 7(4):558-568.
35(3):178-189. 19. Ismail Z, et al. Forecasting gold prices using multiple
3. Chen X, Fang Y. Enterprise systems in financial sector-an linear regression method. American Journal of Applied
application in precious metal trading forecasting. Sciences. 2009; 6(8):1509.
Enterprise Information Systems. 2013; 7(4):558-568. 20. Navin DG. Big Data Analytics for Gold Price Forecasting
4. Gabralla LA, et al. Oil price prediction using ensemble Based on Decision Tree Algorithm and Support Vector

48
International Journal of Multidisciplinary Research and Development

Regression (SVR), 2013.

21. Mirmirani S, Li H. Gold price, neural networks and
genetic algorithm. Computational Economics. 2004;
23(2):193-200.
22. Grudnitski G, Osburn L. Forecasting S&P and gold
futures prices: An application of neural networks. Journal
of Futures Markets. 1993; 13(6):631-643.
23. Paliyawan P. Stock Market Direction Prediction Using
Data Mining Classification. 2006; 5:6.

49