0 calificaciones0% encontró este documento útil (0 votos)
16 vistas6 páginas
IPASJ International Journal of Computer Science (IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org
Volume 2, Issue 5, May 2014 ISSN 2321-5992
IPASJ International Journal of Computer Science (IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org
Volume 2, Issue 5, May 2014 ISSN 2321-5992
IPASJ International Journal of Computer Science (IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org
Volume 2, Issue 5, May 2014 ISSN 2321-5992
IPASJ International Journal of Computer Science(IIJCS)
Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm
A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org Volume 2, Issue 5, May 2014 ISSN 2321-5992
Volume 2 Issue 5 May 2014 Page 31
Abstract In this paper, we discuss results of forecasting nonstationary financial time series such as IIP usi ng yield spread between long and short-term rates with Wavelet Neural Networks. Using the yields on securities with maturities ranging from three months to ten years, we analyzed four different long minus short yield spreads (policy horizon and non-policy horizon) . We observed the predictive power of each of these spreads for economic activity within aggregate and time scale framework using Wavelet Neural Networks gives slightly better forecasting ability than Neural Networks and significantly better results can be obtained than other conventional techniques. We also recommend the use of lagged IIP for greater efficiency that enables the yield spread to produce better IIP forecasts.
1. INTRODUCTION Financial time series such as stock market index, exchange rate etc are inherently non-stationary, it is regarded as one of the most challenging applications of time series forecasting. Whilst non-linear models are often used for a variety of purposes, one of their prime uses is for forecasting, and it is in terms of their forecasting performance that they are most often judged. However, a casual review of the literature suggests that often the forecasting performance of such models is not particularly good. The relatively poor forecasting performance of some non-linear models calls for substantive further research in this area, given that one might feel uncomfortable asserting that non-linearities are unimportant in describing economic and financial phenomena[1]. Wavelet Neural Networks for financial time series forecasting with regard to certain parameters and expect better results than other techniques such as Fourier methods, statistical techniques, numerical methods and neural networks etc.
2. Wavelet Neural Networks in Financial Time Series Forecasting In recent years, wavelets have become a very active subject in many scientific and engineering research areas. Especially, wavelet neural networks (WNN) inspired by both the feedforward neural networks and wavelet decompositions have received considerable attention and become a popular tool for function approximation[2]. Wavelet neural networks are feedforward neural networks with one hidden layer, comprised normally of radial (e.g. Mexican hat) wavelets as activation functions, and a linear output layer. The output layer of the wavelet neural network represents the weighted sum of the hidden layer units, i.e. wavelet basis functions[3]. The structure of the wavelet network is similar to that of the radial basis function (RBF) network except that in the wavelet network the radial basis functions are replaced by orthonormal scaling functions that are not necessarily radial-symmetric. In contrast to classical sigmoidal-based artificial neural networks (AAN), wavelet networks provide efficient network construction techniques, faster training times, and multiresolution analysis capabilities. At present, there are two different kinds of wavelet neural network structure, one with fixed wavelet bases, where the dilation and translation parameters of wavelet basis are fixed, and only the output layer weights are adjustable. Another type is the variable wavelet bases, where the dilation parameters, translation parameters and the output layer weights are adjustable[4]. Wavenet is another term to describe wavelet networks. Originally, wavenets did refer to neural networks using dyadic wavelets. In wavenets, the position and dilation of the wavelets are fixed and the weights are optimised by the network so they belong to the first category of wavelet neural networks. The theory of wavenets has been generalised to many different settings such as biorthogonal wavenets, fuzzy wavenets, multiresolution analysis wavenets etc. We limit the scope of Using Wavelet Neural Networks to Forecast IIP Growth with Yield Spreads
M. Yasin Pir 1 , Firdous Ahmad Shah 2 M. Asger 3
1 Department of Computer Applications, Govt. College for Women, M.A.Road, Srinagar-190010, J &K,India.
2 Department of Mathematics, South Campus, University of Kashmir, Anantnag-192101, J &K, India.
3 School of Mathematical Sciences & Engineering, BGSB University, Rajouri-185131, J &K, India. IPASJ International Journal of Computer Science(IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org Volume 2, Issue 5, May 2014 ISSN 2321-5992
Volume 2 Issue 5 May 2014 Page 32
this section to the first category of wavelet neural networks[5]. Wavelet Neural Networks can be used for financial time series forecasting with regard to certain parameters and better results are expected than other techniques.
Frame work of the Forecasting Model
1. Decompose original time series using wavelet decomposition and obtain the corresponding approximation and details series at a predetermined level of resolution. 2. Design neural network predictive model for each of the decomposed components of the original time series 3. Input variables of the neural network models for each of the decomposed series, comprising of technical indicators. 4. Use the model for arriving at the results.
3. DATA AND METHODOLOGY The data is collected fromRBI [6]. In this study, monthly data fromOctober 1996 to June 2012, which gives a data window size of 189 observations, have been used to test the predictability of yield spreads for forecasting output growth. Index for Industrial Production (IIP) is considered to derive the output growth For the construction of yield spreads, we use the spread between the long termyields to maturity (YTM) on Government of India (GOI) securities and short term Treasury bill rates constructed at shorter end, longer end and policy relevant areas of yield curve. These spreads include one year yields minus three months Treasury bill rate (Sp 1, 3) and ten year yields minus five year yields (Sp 10, 5) which are known as policy horizon spreads, whereas, five year yields minus three months Treasury bill rate (Sp 5, 3) and ten year yields minus three months Treasury bill rate (Sp 10, 3) are known as Non-policy horizon spreads. The results are graphically depicted in Fig. 1.1 to Fig.1.5.
IPASJ International Journal of Computer Science(IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org Volume 2, Issue 5, May 2014 ISSN 2321-5992
Volume 2 Issue 5 May 2014 Page 33
In our study, the Original Time Series values (IIP Growth and 4 constructed spreads) are normalized. The normalization step will rescale the orginal time series and will minimize the effect of outliers in the series and the series is reduced to zero mean and unit standard deviation. The following transformation would transform the original input variables x 1 , x 2 , x 3 ,.. x d into normalized variables x 1 ; x 2 ,x 3 , x d . i=1..d
Two approaches are followed. In the first approach (Model 1), the normalized values of IIP Growth and four constructed spreads are fed to the neural network model separately such as : IIP Growth and Sp(1,3) , IIP Growth and Sp(10,5), IIP Growth and Sp(5,3) and IIP Growth and Sp(10,3) respectively. The lagged IIP Growth values which lag by value of 1, 2, 3 and 4 (namely lag1, lag2, lag3 and lag4) and the values of four constructed spreads are fed to the network independently and the observations are recorded for the forecasted values of IIP Growth by neural network model. Neural Network GUI toolbox of Matlab (R2010a) was used for this purpose[7] . The results are graphically depicted in Fig 2.1 to Fig 2.5.
In the second approach (Model 2), the normalized values of IIP Growth and four constructed spreads are de-noised using wavelets before they are fed to the neural network model and same procedure is followed to forecast IIP Growth values as was used in Model 1 after De-noising step.
Fig. Model 2 IPASJ International Journal of Computer Science(IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org Volume 2, Issue 5, May 2014 ISSN 2321-5992
Volume 2 Issue 5 May 2014 Page 34
One-Dimensional Discrete Stationary Wavelet Transform (SWT) was used for De-noising. De-noising requires thresholdingthat discards only the portion of the details that exceeds a certain limit. Using wavelets to remove noise from a signal requires identifying which component or components contain the noise and then reconstructing the signal without those components. The Wavelet GUI Toolbox of Matlab (R2010a) was used for this purpose that includes automatic thresholding[8]. The results are graphically depicted in Fig 3.1 to 3.5.
The de-noising procedure proceeds in three steps:
1. Decomposition by any wavelet. In our case Daubechies wavelet at level 2 was used. 2. Thresholding detail coefficients. For level 2 in our case, fixed formthreshold was chosen and soft thresholding was applied to detail coefficients. 3. Reconstruction. In our case the reconstructed series at level 2 are IIP Growth values and four constructed spreads.
In both the approaches, Feed forward back propagation network was created. The training function TRAINLM is used that updates weight and bias states according to Levenberg-Marquardt optimization which is often the fastest backpropagation algorithmin the toolbox. The adaptive learning function LEARNGDM is used which is the gradient descent with momentumweight and bias learning function. A two layer feed forward backpropagation network was created with 10 number of neurons in the hidden layer which gave the optimumresults in this network. The transfer function used is tansig which is a Hyperbolic tangent sigmoid transfer function which is fastest as well. The performance function used is MSE (Mean Squared Error). The following Regression Analysis were obtained for IIP Growth & 4 constructed spreads with two approaches discussed above. IPASJ International Journal of Computer Science(IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org Volume 2, Issue 5, May 2014 ISSN 2321-5992
1. The spread is a reliable predictor of output growth such as IIP and that the spread provides good recession forecasts, especially up to one year ahead. 2. Since better regression results are obtained from both the approaches which implies that both the approaches can successfully forecast the IIP Growth than other models. 3. In both the approaches, it is evident that ten year yields minus five year yields (Sp 10, 5) which are known as policy horizon spreads remains an important variable in predicting IIP Growth than other spreads. 4. The use of lagged IIP enables greater efficiency for yield spreads to produce better IIP forecasts. The best results are obtained at lag 2 in both the approaches and Sp(10,5); the forecasting results are depicted graphically in Fig 4.1 & Fig 4.2. using Neural Network Approach and Wavelet Neural Network Approach respectively. 5. Using Wavelet Neural Network Approach gives slightly better forecasting results than Neural network Approach.
References [1] Michael P. C, Philip H.N., and Swanson R., Forecasting economic and financial time-series with non-linear models, International Journal of Forecasting 20 169 183 (2004). [2] V. Kodogiannis, I. Petrounias, and J. Lygouras, "Adaptive Wavelet Neural Networks for Non-linear Modelling and Control," NEURAL PARALLEL AND SCIENTIFIC COMPUTATIONS, vol. 15, p. 221, 2007. [3] E A Rying, G L Bilbro and Jye-Chyi Lu, Focused local learning with wavelet neural networks, IEEE Transactions on Neural Networks ,02/2002; vol. 13(2):304-19,2002. [4] Daniel W.C.Ho, Ping-An Zhang , and J inhua Xu, Fuzzy Wavelet Networks form Function Learning , IEEE Transactions on Fuzzy Systems, Vol 9, No. 1,February 2001. [5] Marc Thuillard , A Review of Wavelet Networks, Wavenets, Fuzzy Wavenets and Applications, Advances in Computational Intelligence and Learning , International Series in Intelligent Technologies Volume 18, pp 43-60 , 2002. [6] IIP Growth. Available at http:// www.rbi.org.in. [7] Neural Network Toolbox Users Guide, The MathWorks, Inc. available at http://www.mathworks.com. [8] MATLAB- The Language of Technical Computing, available at : http://www.mathworks.com/products/matlab. IPASJ International Journal of Computer Science(IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm A Publisher for Research Motivation ........ Email: editoriijcs@ipasj.org Volume 2, Issue 5, May 2014 ISSN 2321-5992
Volume 2 Issue 5 May 2014 Page 36
AUTHOR
M. Yasin Pir received M.C.A. (Master of Computer Application) from Kashmir University,India in 2004. During 2005-2011, he was working as Assistant Professor in the Department of Computer Sciences in BGSB University, J&K and is pursuing his research from the same University. He is presently working as Assistant Professor in Higher Education Department, J&K,India.