Expert Systems with Applications 39 (2012) 6168–6173 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa A dynamic meta-learning rate-based model for gold market forecasting Shifei Zhou a,⇑, Kin Keung Lai a, Jerome Yen b a b Department of Management Sciences, College of Business, City University of Hong Kong, Kowloon, Hong Kong Department of Finance, Hong Kong University of Science and Technology, Kowloon, Hong Kong a r t i c l e Keywords: EMD BPNN Meta-learning Forecasting i n f o a b s t r a c t In this paper, an improved EMD meta-learning rate-based model for gold price forecasting is proposed. First, we adopt the EMD method to divide the time series data into different subsets. Second, a back-propagation neural network model (BPNN) is used to function as the prediction model in our system. We update the online learning rate of BPNN instantly as well as the weight matrix. Finally, a rating method is used to identify the most suitable BPNN model for further prediction. The experiment results show that our system has a good forecasting performance. Ó 2011 Elsevier Ltd. All rights reserved. 1. Introduction get trapped (Orr & Leen, 1992), affecting the prediction precision adversely. 1.1. Motivation Forecasting gold price is becoming increasingly important. For long in history, gold has been traded actively on international markets. Many derivatives of gold trading in international gold markets are also traded, such as gold futures, gold options, gold forward contracts, and so on (Grudnitski & Osburn, 1993; Shafieea & Topalb, 2010). Remarkably, since the price of gold varies within a limited range, gold is able to reduce the effect of inflation, control the rise of price and help carry out constrictive monetary policy (Atsalakisa & Valavanisb, 2009). Hence, gold becomes an important tool for risk hedging as well as an investment avenue. Therefore, predicting the price of gold has become very significant and important to investors. However, the current forecasting algorithm offers poor precision in solutions for non-linear problems. We mainly focus on the artificial neural network (ANN) forecasting algorithm. First, the traditional ANN prediction algorithms employ a global fixed learning rate to change the weight matrix. This causes two problems. On one hand, if the learning rate is too small, the weight matrix changes very slowly. Then, the network training process takes a long time to converge (Orr & Leen, 1996). On the other hand, if the learning rate is too large, we may miss the minimal optimal and cause the algorithm to diverge (Fig. 1). Second, unlike the polynomial function, most non-linear problems have complex error surfaces. Hence, there are many local minima in which the training process may ⇑ Corresponding author. Tel.: +852 5344 8097; fax: +852 3442 0189. E-mail addresses: [email protected], [email protected] (S. Zhou). 0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.11.115 1.2. Contributions In this paper, we propose an improved empirical mode decomposition model (IEMD) to forecast the trend and price of gold market. The IEMD model can make efficient use of historical data to predict the future price. During the back-propagation neural network training process, we dynamically change the learning rate by using a global meta-learning rate. Our main contributions are listed as follows: (1) High convergence speed. When the trend of historical data is downward, we change the learning rate to increase the weight matrix in the same direction to address the decreasing speed of the error function. Similarly, when the trend is upslope, the gradient will be increased, to ensure the process will ‘‘climb the hill’’ of error function effectively. Both methods accelerate convergence speed of the neural network training process. (2) High prediction precision. By calculating the gradient of the error function, we minimize the mean square error by changing the weight matrix of the prediction model. The novel system we propose is able to update the learning rate of back-propagation neural network instantly, and capture the trend of data series. As a result, our system is able to yield a high prediction precision for gold market. (3) Significant application. Gold plays an important role in international markets. It is receiving increasing attention from investors. Gold offers a way of risk hedging and it can also be stored without depreciation. A precise prediction of price of gold can help investors save money. This, therefore, shows the potential significance of our system. S. Zhou et al. / Expert Systems with Applications 39 (2012) 6168–6173 2. Related works 2.1. Empirical mode decomposition The empirical mode decomposition method was first proposed by Huang et al. (1998). This time series data decomposition technique applies the Hilbert transform to nonlinear and nonstationary time series data. The main idea of EMD is to decompose a time series into a sum of oscillatory functions. The EMD can make effectively use historical data. Since the EMD is able to divide the original time series into several subsets of data by using the intrinsic mode function (IMF), characteristics of different subsets of series can be identified by training different BPNNs. We can use these trained BPNNs to predict the future trend of time series and select one of the network models for future prediction. 2.2. Online learning algorithm In the online learning algorithm, weight vectors of input data are updated immediately after presentation of each data point (Saad, 1999). Therefore, the online learning algorithm is able to adjust weights of input data and capture the trend of time series instantly. Online learning algorithm has a number of advantages. As the weight matrix is updated recursively, this algorithm can be used when there is no fixed training set and new data keeps coming in. Besides, local minimum is a problem for gradient descent in nonlinear models whereas online learning can easily escape from local minima when a noisy data is input. In this paper, we use on- 6169 line learning algorithm to update the weight matrix. At the same time, we also update the learning rate to make sure the network model can be trained at a fast convergence speed. To achieve this goal, we apply a meta-learning rate to the model. Meta-learning means learns from historical data, which studies how to choose the right bias dynamically rate rather than a fixed base-learning (Abraham, 2004; Ricardo & Youssef, 2002). 3. Improved empirical mode decomposition model (IEMD) 3.1. IEMD model structure Yu et al. proposed an EMD-based neural network ensemble learning model (Yu, Lai, Wang, & He, 2007; Yu, Wang, & Lai, 2008). In this model, they use forward neural network as the prediction model. The time series data is divided into several subsets by using IMF (Huang et al., 1998). Each subset is used as an input data for the FNN model for training. The FNN model is finally used to predict future outcome. All predicted results are assigned a weight and are combined together. This model is quite effective for dealing with historical data. However, prediction results are greatly affected by the weight of each FNN model. Besides, if one of the FNN models has made a precise prediction, the final prediction may be imprecise because all predicted results are summed together. To overcome these problems, we propose an improved EMD online learning-based back-propagation neural network model. The structure of the IEMD–BPNN–PMR model is shown in Fig. 2. To make efficient use of the historical data, we adopt the EMD method to partition the data into several subsets. Each subset of data is used to train the back-propagation neural network. The time series are decomposed according to the sifting procedure proposed by Yu et al. (2008). This procedure is repeated until all data are divided and each subset of data has only one local minimal or maximal. At the end of the sifting procedure, data series x(t) can be expressed as follows: xðtÞ ¼ n X ci ðtÞ þ r n ðtÞ ð1Þ i¼1 where n is the number of IMCs, namely number of data subsets, ci(t) is the ith IMF, and rn(t) is the final residue of the procedure. Thus, any time series data can be decomposed by using this EMD method. Frequency components contained in each frequency band are different. They also change the variance of data series x(t), while rn(t) represent the trend of data series x(t). 3.2. Back-propagation neural network Back-propagation neural network (BPNN) is a type of artificial neural network (White, 1990) which is a class of typical intelligent learning paradigm and is widely used in the field of data forecast- Fig. 1. Effect of learning rate on algorithm convergence. Fig. 2. An improved EMD online learning-based BPNN model. 6170 S. Zhou et al. / Expert Systems with Applications 39 (2012) 6168–6173 ing. In this paper, we use a three-layer neural network with error-back-propagation algorithm. This network model is used as the predicting model of our system. There are many nodes in the hidden layer. Hence, there are multiple combinations of weights and data points. BPNN is able to provide flexible mapping between inputs and outputs. Hornik, Stinchocombe, and White (1989) had proved that a three-layer of feed-forward neural network with an identity transfer function in the output unit and logistic functions in the hidden layer can well approximate any arbitrary continuous function. Therefore, in this paper, we use the three-layer BPNN as the forecasting model. Assume that yi represent the ith hidden node, then yi ¼ n X wi xi þ w0 ð2Þ tion will be accordingly increased. Therefore, this relationship helps modify the weight matrix to control the errors. Then, following Saad (1999), we can calculate the increment of weight as follows, using Eq. (6). Dwij ¼ lij @Ej ¼ lij ðt j yj Þf 0 ðyi Þxi @wij ð7Þ where lij represents the online learning rate. We can see that the increment of weight will depend on the learning rate. If l ij is large, each change of weight will be similarly large. The online weight update that uses a local and time-varying learning rate for each weight can be defined as follows. wij ðt þ 1Þ ¼ wij ðtÞ þ lij ðtÞDwij ðtÞ ð8Þ where Dwij(t) is the increment of weight, defined in Eq. (7). i¼1 where xi is the data input and wi is the corresponding weight. Meanwhile, yi is also called a net input, used in latter calculations. After the hidden layer’s nodes are generated, we use the activation function f() to transform the net input into output, namely, yj ¼ f ðyi Þ ð3Þ where j = 1, . . . , m and m represents the number of nodes in the hidden layer. By using this network model, we finally make the prediction for the time series. 3.3. Prediction model rating 1 ðt j yj Þ2 2 ð4Þ where tj is the target value of the time series; yj is the predicted outcome from the jth BPNN model; and j ranges from 1 to m. Then, we compare all these Ej and find the minimal one, say Ej, and increase the rate of jth model by 1. This procedure is repeated throughout the training process. Finally, we get a BPNN prediction model with the largest rate for prediction. 4. Improved online learning algorithm 4.1. Online weight update In this section, an improved online learning algorithm is proposed. We start with the mean squared error and formulate some mathematical methods to deduct the relationship between weight and MSE. First, we use the chain rule to decompose the gradient into two factors: @Ej @Ej @yj ¼ @wij @yj @wij ð5Þ where wij represents the weight value from the ith data input to the jth hidden node. By combining Eqs. 2, 3 and 5, we have the following result: G¼ @ @Ej ¼ @wij 1 ðt 2 j 2 yj Þ @yj ¼ ðyj t j Þf 0 ðyi Þxi @yj @wij As discussed above, the online learning algorithm has the ability to escape from the local minimum. The noise in the stochastic error surface bounces the network out of local minimums as long as they are not too severe. When the weight matrix is modified, we simultaneously change the local learning rate by gradient descent method. In order to reduce error function, we change the learning rate before changing the weights. Then, the gradient can be defined as follows. nij ðt þ 1Þ ¼ We have to ensemble the prediction results from different BPNN models. Thus, the prediction model rating (PMR) method is proposed to handle this problem. In this section, we calculate the mean squared error E(MSE) of every output of the BPNN model. Ej ¼ 4.2. Online learning rate update ð6Þ where G represents the gradient of MSE. G is not necessarily a positive value since f0 (yi) may be a negative value. Here, we assume G is positive. Then, if there is an increment in weight wij, the error func- @Eðt þ 1Þ @Eðt þ 1Þ @wij ðt þ 1Þ ¼ @ lij ðt þ 1Þ @wij ðt þ 1Þ @ lij ðtÞ ¼ ðyj tj Þf 0 ðyi Þxi Dwij ðtÞ ð9Þ By applying the result of Eq. (9) and introducing a meta-learning rate, we can define the online learning rate as follows. lij ðtÞ ¼ lij ðt 1Þ kðyj tj Þf 0 ðyi Þxi Dwij ðt 1Þ ð10Þ where k is a global meta-learning rate. There are two disadvantages of using fixed global meta-learning rate. Since the meta-learning rate is fixed, the increment of learning rate is also fixed from Eq. (10). This fixed increment may sometimes speed up the movement of error cost towards the minima when the movement is down-slope. A fast accelerating speed causes the training process to pass by the local minimal and diverge from data trend, the process of which is shown on Fig. 1(b). Besides, a fixed learning rate increment cannot adaptively adjust the converge speed of the training process. It can only increase or decrease the speed at a fixed acceleration. This, therefore, can not improve the learning speed effectively. Hence, we mainly focus on finding an adaptive meta-learning rate in the remaining part. In general, the error cost is consisted of k output nodes. And each individual output node is corresponding to different estimated output. According to the different output, we can calculate the error cost. Hence, we define the error cost function as follows. Eðt þ 1Þ ¼ 1 eðt þ 1ÞT eðt þ 1Þ 2 ð11Þ where e(t + 1) is the difference between the target output and prediction results, namely ej ðtÞ ¼ yj ðtÞ t j ðtÞ ð12Þ Since there exist k output nodes, there exists an output error vector, which is defined as e(t) = [e1(t), e2(t), . . . , ek(t)]T 2 Rk1, and yj(t) is the jth entry of estimated result at time t. As mentioned above, even though the learning rate is updated instantly, the fixed global learning rate may cause problems and have poor performance for training convergence. Meanwhile, in S. Zhou et al. / Expert Systems with Applications 39 (2012) 6168–6173 order to make the neural network to re-learn from new input data, the neural network learning process should be recursive and iterative. From Yu et al. (2007), the error cost vector can be rewritten as follows. eðt þ 1Þ ¼ eðtÞ þ lðtÞnðtÞ ð13Þ By combining Eqs. (11) and (13), we get the following equation. 1 ½eðtÞ þ lðtÞnðtÞT ½eðtÞ þ lðtÞnðtÞ 2 Eðt þ 1Þ ¼ ð14Þ The error cost at time (t + 1) is the function of error cost at time t and learning rate and gradient of learning rate on error cost. This equation shows that the history data is used to compute the error cost function. When using online learning algorithm, we consider the increment process of the error cost as a continuous one. This means that we can take derivative of the error cost function and find the local minimal data point. For a continuous function, we obtain a local optimal when the first derivative of this function is equal to 0. Namely, dEðt þ 1Þ 1 1 ¼ nT ðtÞ½eðtÞ þ lðtÞnðtÞ þ ½eðtÞ þ lðtÞnðtÞT nðtÞ ¼ 0 dl l¼l ðtÞ 2 2 ij ð15Þ where lij ðtÞ is the local optimal learning rate, at which we obtain a minimal square error. To continue the process, we take the second derivative of error cost function. We find that the derivative result is larger than 0, which means that we obtain a lower bound local minimal at lij ðtÞ. 2 d Eðt þ 1Þ dl2 ¼ nT ðtÞnðtÞ > 0 ð16Þ l¼lij ðtÞ By solving Eq. (15), we obtain the expression of learning rate, namely lðtÞ ¼ nðtÞeðtÞ ð17Þ nT ðtÞnðtÞ From the definition of l(t), we find that the learning rate is affected by the estimated output and the change of weight matrix at the last epoch training. Therefore, we can draw a relationship among the meta-learning rate, estimated output and change of weight matrix through combining the Eqs. (10) and (17). Then, we can obtain the meta-learning rate as follows. k¼ nT ðtÞeðtÞ þ lðt 1ÞnT ðtÞnðtÞ nT ðtÞnðtÞ n1 ðt 1Þ ð18Þ 6171 From the above definition of k, we know that this meta-learning rate changes with time data series. At different data point, we obtain different meta-learning rate. When considering the definition of n(t) and l(t), we find that the change of k is proportional to the change of weight matrix Dw(t). Like what is shown in Fig. 3, when the error cost moves in the direction of down slope, Dw(t) will decrease, which accordingly causes the meta-learning rate to decrease. When k is decreasing, it will decrease the learning rate in turn and, therefore, lower down the learning speed and converge to the local minimal. Conversely, when the error cost function makes upslope motions, the weight matrix will increase. With a positive change of weight matrix, the meta-learning rate is also increased. Since the moving process of error cost function is trying to escape from the local minimal, the increment of metalearning rate will accelerate the divergence process. 5. Experiments 5.1. Evaluation criteria In this part, three classes of measurement criteria including level prediction, convergence speed and directional forecasting are employed to evaluate the performance. First, unlike Bo, Wang, and Lai (2007), we use the mean absolute error (MAE) (Willmott & Matsuura, 2005) rather than the criteria of root mean squared error (RMSE) because RMSE varies with variability within the distribution of error magnitudes and MAE is a more natural measure for the average error. Thus, we use MAE to compare the prediction precision of different forecasting models with our novel model. The definition of MAE is given by Eq. (11). MAE ¼ N 1 X ^ i yi j jy N i¼1 ð19Þ ^i is the predicwhere N represents the total number of data points; y tion result of ith data point; and yi is the target value of ith data point. The smaller the value of MAE, the more precise the prediction model will be. Second, we compare our model with other prediction models in terms of convergence speed. Though an algorithm can run at a high speed, this does not mean the algorithm is good because if the prediction result has a large value of MAE, the algorithm will be useless to investors. Therefore, we consider the running time of the models together with the prediction precision. Third, in order to measure the direction of movement of time series accurately, we use a directional statistic (Dstat) (Yu, Wang, & Lai, 2005), which is defined by the following equation. Dstat ¼ N 1X ai 100% N i¼1 ð20Þ ^iþ1 yi Þðyiþ1 yi Þ > 0, otherwise, ai = 0.N reprewhere ai = 1, if ðy sents the total number of testing samples. The larger this factor is, the more proper direction the forecasting model will have. 5.2. Experimental results Fig. 3. Effect of meta-learning rate on convergence. In this section, we use daily gold future price data from the database of global financial data (GFD). The database provides the most extensive time series data on commodities available anywhere. It includes historical data of about 100 different commodities. In this paper, we use data spanning from 1 Jan 2000 to 13 Jul 2010, a total of 2636 observations. And we use the rest of the time series spanning from 29 Dec 2009 to 14 Jul 2010 for 6172 S. Zhou et al. / Expert Systems with Applications 39 (2012) 6168–6173 evaluation of prediction performance, which contains a total of 136 observations. To begin the experiment, we firstly use the EMD technique to handle the time data series. Before we use this technique, there are two requirements which should be satisfied. (1) The number of extreme points should be equal to zero points of the series or differ by only one; (2) The mean value of envelope of local minimal and local maximal must be zero. With these two requirements, the IMCs contained in EMD may become more harmonics when we apply EMD to decompose the data series (Huang, Shen, & Long, 1999). The decomposition procedure is descried as follows: (1) Identify all the extreme points in the data series, including local maximal and local minimal; (2) Generate upper and lower envelops xup(t) and xlow(t) for the all local extrema; (3) Compute the mean envelope m(t) for all the extrema from xup(t) and xlow(t); (4) Calculate the difference between mean envelope at point t and x(t) dðtÞ ¼ xðtÞ mðtÞ ð21Þ (5) Check d(t): (a) if d(t) meets the above two requirements, an IMC is identified and replace x(t) with r(t) = x(t) d(t); (b) if d(t) does not meet, then replace x(t) with d(t); (6) Repeat (1)–(5) until the following stopping condition: T X ½dj ðtÞ djþ1 ðtÞ2 2 t¼1 dj ðtÞ < SC ð22Þ where SC is the stopping condition, usually it is set between 0.2 and 0.3. Compared with the traditional Fourier analysis, this EMD technique owns several advantages. First, this method is robust to handle the nonlinear and nonstationary data since it can automatically and adaptively select the fluctuations from the original time series. Second, EMD can divide the time series into several independent IMCs and a residual component. The behavior of these decomposed components only depends on the nature of the original data. Final- Table 1 The MAE and running time comparisons. Forecasting model MAE Rank Time (s) Rank BPNN FNN EMD–FNN–ALNN IEMD–BPNN–PMR 9.8292 8.9748 3.3549 2.1677 4 3 2 1 83 72 124 79 3 1 4 2 ly, this method is easy to understand and use and owns wide applications. The decomposition results are shown on Fig. 4. Each of the subfigures represents an IMC. These sub-data series are then used for the BP neural network training. Since the IMC have extracted the characteristics from the original data, the training of neural network will have much high convergence speed by using these data series. In many literatures, researchers use the logistic sigmoid f ðuÞ ¼ 1þe1 u as activation function for units in the hidden layer. This function has a small asymmetric range from 0 to 1. Then, this logistic sigmoid function worsens the condition of the network. Accordingly, in our prediction model, we choose tanh as the activation function f(), which has a symmetric range from 1 to 1 and is more suitable as activation function for hidden units. Since tanh0 (u) = 1 tanh (u)2, then we have f 0 ðyi Þ ¼ 1 y2i . From Table 1, MAE values of different models are different. BPNN has large MAE of 8.9748 and ranks the last. Our model, IEMD–BPNN–PMR, ranks at the first place with MAE value of 2.1677 and is followed by EMD–FNN–ALNN model, BPNN model and FNN model. This sequence represents a descending order of prediction precision. In terms of running time, we can see that the FNN model spends less time to converge. It takes only 72 s, which is only 20.11% of the total time of all models. Our model ranks at the second place and takes 79 s to converge. The FNN model is a simple type of neural network. For a three layer FNN, the model consists of a linear combination of input data and the weight matrix and an activation function to transfer the combined results. Hence, the FNN model will take the least running time to converge. However, the precision of this model is also the worst. Conversely, our model adopts a BPNN model as the prediction model. Although it takes more time to get converged, prediction results are more precise than the FNN model. Our model runs faster than the individual BPNN model. This is because we have employed an EMD method to decompose the input data, which accelerates the convergence speed to some extent. Besides, the usage of improved online learning algorithm makes sure the weight matrix is changed instantly at each epoch. As a result, the training process has a high convergence speed. Table 2 shows results of comparison of different models on the aspect of directional forecasting. Ranking at the first place is the IEMD–BPNN–PMR model, which is followed by EMD–FNN–ALNN model. The reason why our model can predict correctly the trend of data series most of the time is that we make some changes to the original BPNN model and use tanh as the activation function, Table 2 The Dstat (%) comparisons. Fig. 4. The decomposition of gold market price. Forecasting model Dstat (%) Rank BPNN FNN EMD–FNN–ALNN IEMD–BPNN–PMR 50.74 59.56 66.91 86.76 4 3 2 1 S. Zhou et al. / Expert Systems with Applications 39 (2012) 6168–6173 which has a symmetric range from 1 to 1. This helps the activation function to capture the series trend flexibly and accurately. 6. Conclusions In this paper, we propose an improved EMD online learningbased model for gold price forecasting. We first use the EMD method to partition the time series data into several subsets of data series, which are used for training the neural network. Then, back-propagation neural network is employed to function as the prediction model in our novel system since BPNN is one of widely studied and used learning algorithms in the academia. By studying the essential process of BPNN mathematically, we formulate a deduction process to show that online learning can be updated instantly. As the learning rate is updated online, the weight is also updated after every data input. This ensures that the BPNN model is able to escape from the local minimum of the error function and, therefore, is able to show a good performance in terms of quick convergence speed. The experiment results show that our system has offers good prediction results as well as running time. Although we apply the novel algorithm to the gold market and obtain some impressive results, there is still some improvement we can make. The global meta-learning rate can be further studied to see whether there exists a relationship between MAE and data samples. Besides, our model can also be applied to stock and warrants market (Huang, Nakamori, & Wang, 2005; Lu, 2010). There will be more interesting and practical results. References Abraham, A. (2004). Meta learning evolutionary artificial neural networks. Neuro Computing, 56, 1–38. Atsalakisa, G. S., & Valavanisb, K. P. (2009). Suveying stock market forecasting techniques- Part II: Soft computing methods. Expert Systems with Applications, 36(3, Part 2), 5932–5941. 6173 Bo, W., Wang, S., Lai, K. K. (2007). A hybrid ARCH-M and BP neural network model for GSCI futures price forecasting. In International conference on computational science (3) (pp. 917–924). Grudnitski, G., & Osburn, L. (1993). Forecasting S& P and gold futures prices: an application of neural networks. Journal of Futures Markets, 13(6), 631–643. Hornik, K., Stinchocombe, M., & White, H. (1989). Multilayer feedforward networks are universal approximators. Neural Networks, 2, 359–366. Huang, W., Nakamori, Y., & Wang, S.-Y. (2005). Forecasting stock market movement direction with support vector machine. Computers & Operations Research., 32(10), 2513–2522. Huang, N. E., Shen, Z., & Long, S. R. (1999). A new view of nonlinear water waves: The Hilbert spectrum. Annual Review of Fluid Mechanics, 31, 417–457. Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q., et al. (1998). The empirical mode decomposition and the Hilbert spectrum for nonlinear and nonstationary time series analysis. Proceedings of the Royal Society A: Mathematical Physical & Engineering Sciences, 454, 903–995. Lu, C.-J. (2010). Integrating independent component analysis-based denoising scheme with neural network for stock price prediction. Expert Systems with Applications., 37(10), 7056–7064. Orr, G. B., & Leen, T. K. (1992). Learning in neural networks with local minima. Physical Review A, 46(8), 5221–5231. Orr, G. B., & Leen, T. K. (1996). Using curvature information for fast stochastic search. Neural Information Processing Systems, 606–612. Ricardo, V., & Youssef, D. (2002). A perspective view and survey of meta-learning. Artificial Intelligence Review, 18(2), 77–95. Saad, D. (Ed.). (1999). On-line learning in neural networks. New York, NY, USA: Cambridge University Press. Shafieea, S., & Topalb, E. (2010). An overview of global gold market and gold price forecasting. Resources Policy, 35(3), 178–189. White, H. (1990). Connectionist nonparametric regression: multilayer feedforward networks can learn arbitrary mappings. Neural Networks, 3, 535–549. Willmott, C. J., & Matsuura, K. (2005). Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance. Climate Research, 30(3), 79–82. Yu, L., Lai, K. K., Wang, S., He, K. (2007). Oil price forecasting with an emd-based multiscale neural network learning paradigm. In International conference on computational science (3) (pp. 925–932). Yu, L., Wang, S., & Lai, K. K. (2005). A novel nonlinear ensemble forecasting model incorporating GLAR and ANN for foreign exchange rates. Computers & OR, 3, 2523–2541. Yu, L., Wang, S. Y., & Lai, K. K. (2007). An on-line learning algorithm with adaptive forgetting factors for feedforward neural networks in financial time series forecasting. Nonlinear Dynamics and Systems Theory, 7(1), 97–112. Yu, L., Wang, S., & Lai, K. K. (2008). An EMD-based neural network ensemble learning model for world crude oil spot price forecasting. Soft Computing Applications in Business, 261–271.
© Copyright 2026 Paperzz