ijfm journal

Upload: traderescort

Post on 03-Jun-2018

228 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/12/2019 IJFM Journal

    1/46

    Chief Editor

    Dr. Balwinder Singh,Associate Professor, Department of Commerce & Business Management, Guru Nanak Dev University, Amritsar, India

    Associate Editors

    Dr. Revti Raman,Victoria University of Wellington, New Zealand

    Dr. Mahesh Joshi,RMIT University, Melbourne, Australia

    Dr. Kapil Gupta,Professor, Management Department, Punjab Technical University, Kapurthala

    Editorial Coordinator

    Ms. Rekha Handa, Assistant Professor, Department of Commerce & Business Management, Guru Nanak Dev University, Amritsar, India

    Editorial Advisory Board

    Dr. C.P.Gupta, Professor, Dept. of Commerce, Delhi University, New Delhi, IndiaDr. Ravinder Vinayak, Professor, Maharishi Dayanand University, Rohtak, Haryana, India

    Dr. Sanjay Rastogi, Indian Institute of Foreign Trade, New Delhi, India

    Dr. H. Venkateshwarlu, Professor, Osmania University, Hyderabad, A.P., India

    Dr. Bala Balachandran, La trobe University, Australia

    Dr. Golaka C Nath, Vice President, Clearing Corporation of India, Mumbai, India

    Dr. N. Panchanatham, Prof., Dept. of Business Administration, Annamalai University, Tamil Nadu

    Dr. Annalisa Prencipe, Universita Bocconi, Milan(Italy)

    Dr. Bhagwan Khanna, Victoria University of Wellington, New Zealand

    Samanthala Hettihewa, University of Ballarat, Ballarat

    Editorial Review Board

    Dr. Ashok Bhardwaj Abbott, Associate Professor, West Virginia University, United States

    Dr. Sourendra Nath Ghosal, Director, Nicco Financial Services Ltd., Kolkata, India

    Dr M.R. Vanitha Mani, Director, MBA Dept., SSK College of Engineering & Tech., Coimbatore, India

    Samuel Gil Martin, Faculty, Universidad Autonoma de San Luis Potos, MexicoDr. Syed Hussain Ashraf, Senior Professor, Dept. of Commerce, Aligarh Muslim University, India

    Dr. Sanjay Jayantilal Bhayani, Associate Prof., Saurashtra University, Rajkot, Gujarat, India

    Mr. Moid U Ahmad, Assistant Professor, Jaipuria Institute of Management, Noida, India

    Mr. Kushankur Dey, Faculty, Institute of Rural Management, Anand, Gujarat, India

    Dr. Shafali Nagpal, Director, UGC-Academic Staff College, BPS Mahila Vishwavidhyalaya, Sonepat

    Dr. Punit Kumar Dwivedi, Assistant Professor, Prestige Institute of Mgt. and Research, Indore, India

    Prof. Amrit Lal Ghosh, Professor, Department of Business Admn., Assam University, Assam, India

    Dr. Ritu Gupta, Assistant Professor, Kamla Lohtia S.D. College, Ludhiana, India

    Prof. S.L. Gupta, Professor, Birla Institute of Technology, Mesra, Ranchi, India

    Mrs. Maithreye Sunil Holeyachi, Assistant Professor, City college, Bangalore, India

    Dr Pawan Jain, Assistant Professor, Institute of Management Technology, Nagpur, India

    Dr. Sudhanshu Joshi, Assistant Professor, School of Management, Doon Univ., Uttrakhand, India

    Badar Alam Iqbal, Professor, Department of Commerce, A.M.U., Aligarh, India

    Mrs. Sonal Gupta, Faculty, CMRIT, Bangalore, India

    Mr Pankaj Varshney, Associate Professor, Apeejay School of Mgt., Dwarka, New Delhi, India

    Dr Jaideep Gulabrao Jadhav, Associate Professor, MIT School of Telecom Management, Pune, India

    Dr. Sri kanth, Associate Professor, PES Institute of Technology, Bangalore, India

    Dr. (Mrs.) Parul Khanna, Associate Professor, Institute of Mgt. & Technology, Faridabad, India

    Mr. Nitin Kulkarni, Faculty, MET Institute of Management, India

    Mr. Anandadeep Mandal, Assistant Professor, KIIT School of Management, Orissa, India

    Ratna Shanker Mishra, Assistant Professor, Banaras Hindu University, Varanasi, India

    Akhil Mishra, Associate Professor, Faculty Of Commerce, BHU, Varanasi, India

    Mr. Debabrata Mitra, Faculty, Department of Commerce, Univ. of North, Bengal, Darjeeling, India

    Dr. Chimun Kumar Nath, Assistant Professor, Dept. of Commerce, Dibrugarh University, India

    Prof. Nikunjkumar Ramnikbhai Patel, Associate Prof., S.V. Institute of Management, Gujarat, India

    Mr. Sudhakar T Paul, Assistant Prof., Dept. of MBA, MVJ College of Engineering, Bangalore, India

    Dr.Raja Ram, Faculty, Kalasalingam University, India

    Prakash shanmugasundaram, Faculty, Department of MBA, Anna University, Coimbatore, India

    Prof Subhash Chander Sharma, Professor, Dept of Commerce and Business Mgt., GNDU, Amritsar,

    Dr Sajeev Surendranath, Senior Lecturer, Institute of Mgt. in Government, Trivandrum, India

    International Journal of Financial Management

    Journal tends to bring about a revolution in the nancial research through its unparalleled quality, undaunted approach and panoptic coverage of

    the research efforts being undertaken all around the globe. The journal intends to provide the super ordinate podium to the researchers to share their

    ndings with the global community after having crossed the quality checks and legitimacy criteria, which in no way promise to be liberal.

  • 8/12/2019 IJFM Journal

    2/46

    Editorial Message

    Greetings for the newly dawned year! Hope it lends us umpteen opportunities to learn and align with the dynamic learn

    ing environment. The advent of New Year has lent me the privilege to add another volume to the growing glory of our

    journal. The International Journal of Financial Management having found a strong foothold with readers, contributor

    and reviewers has taken a betting leap to its next volume. The joy I experience is the shared credit of not only the jour

    nals editorial and reviewing team but also the silent yet strong inputs of our valued critics and suggestion makers.

    The present issue is a ne blend of seven research papers which could take the coveted place here surpassing the

    quality barriers raised through our rigorous and regularly updated review procedures. The rst paper talks about the

    application and development hybrid methodology that combines both ARIMA and Articial neural network model to

    model and predict the stock market index returns. The paper next in line purportedly is the rst exhaustive study of itskind on linkages and the interrelationship between the Asian stock markets and other stock markets during and after

    the crisis employing sophisticated and procedurally sound analytical techniques like Granger Causality test based on

    Vector Error Correction Model and Co-Integration It concludes that the linkages between the Asian and the US stock

    markets are stronger in the post-crisis period. Islamic banking, elimination of gharar, two prominent Islamic banking

    nancing instruments Bay al-Inah and Bay al-Dayn and the legal implications of the presence of gharar on the validity

    of these contracts is the central thought in the next research compilation. Critical investigation into the jurists views to

    examine the revisiting of gharar have been essayed which lend the desired distinctiveness to the work. In line with the

    contemporary talks of intellectual capital and its enormous role in building value the next paper analyses the intellectua

    capital and physical capital of selected companies and their impact on corporate performance using multiple regression

    technique. Performance of both public sector and private sector banks through multivariate analysis has been evaluated

    in the next empirical work using comprehensive measures of performance. Analyzing threadbare the performances ofve major bank in India the study makes signicant contributions in the eld of banking and nance. The literature

    review on a unique research area regarding the role of human capital management in economic value addition of large

    scale organizations also has been made a part of this issue. The culminating paper of this issue relates to construction

    of appropriate benchmark index for mutual funds involving an empirical analysis with specic reference to tax saver

    funds The methodology focuses on estimating the risk adjusted abnormal return generated by the fund that exhibits the

    predictive ability of the fund manager.

    I sincerely hope that the issue and its comprehensive contents meet the quality standards set by our previous issues.

    Being positively receptive to all your valued comments, observations and feedbacks I stand committed to our promises

    of dissemination of quality research in nance.

    Warm wishes,

    Balwinder Singh

    Editor IJFM

    [email protected]

  • 8/12/2019 IJFM Journal

    3/46

    Within Volume 1 Issue 1, January 2012

    ISSN No. 2229-5682(P) ISSN No. 2229-5690(O)Online Access www.publishingindia.com

    1. Stock Index Return Forecasting and Trading Strategy Using

    Hybrid ARIMA-Neural Network Model

    Manish Kumar and M. Thenmozhi 1-14

    2. A Study on the Linkages of Asian and the US Stock Markets

    S.M. Tariq Zafar, D.S. Chaubey and S.R. Sharma 15-32

    3. Revisiting the Principles of Gharar (Uncertainty) in Islamic Banking

    Financing Instruments with Special Reference to Bay Al-Inah and

    Bay Al-Dayn Towards a New Modied Model

    Siti Salwani Razali 33-43

    4. The Role of Intellectual Capital in Creating Value in Indian Companies

    Amitava Mondal and Dr. Santanu Kumar Ghosh 44-54

    5. Performance Evaluation of Public and Private Sector Banks: A

    Multivariate Analysis

    K.V.N. Prasad and D. Maheshwara Reddy 55-62

    6. Role of Human Capital Management in Economic Value Addition of

    Large Scale Organizations: A Literature Review

    Sujata Priyambada Dash and Vijay Agarwal 63-74

    7. Construction of Appropriate Benchmark Index for Mutual Funds:

    Specic Reference to Tax Saver Funds

    Venkatesh Kumar and Ashwini Kumar 73-89

  • 8/12/2019 IJFM Journal

    4/46

    Stock Index Return Forecasting and Trading Strategy Using Hybrid ARIMA-Neural Network Model 1

    Abstrac t

    This study presents the application and development

    of hybrid methodology that combines both ARIMA and

    Articial Neural Network model to take advantageof the unique strengths of both linear and non-linear

    modeling to model and predict the stock market index

    returns. The performance of the hybrid ARIMA Neural

    Network model is compared with the performance of

    ARIMA and Neural Network model. The performance

    of the models are evaluated in terms of widely used

    statistical metrics, correctness of sign and direction

    change and various trading performance measures like

    annualized return, Sharpe ratio, maximum drawdown,

    annualized volatility, average gain/loss ratio, etc. via a

    trading strategy. The ndings of the study reveal that

    the hybrid ARIMA Neural Network model developed is

    the best Forecasting model to achieve greater accuracy

    and yields better trading results.

    Keywords: ARIMA, Articial Neural Network,

    Forecasting, Stock market trading

    JEL Codes: C22, C45, C52, E17, G15

    Stock Index Return Forecasting and Trading

    Strategy Using Hybrid ARIMA-Neural

    Network Model

    Manish Kumar*and M. Thenmozhi**

    1. Introducon

    In the last two decades, forecasting nancial time series

    have been attempted using different linear and non-linear

    methods. The most popular and traditional time series

    model is Box-Jenkins or ARIMA model. The ARIMA

    approach is both simple and yields accurate results which

    explains its wide use. Many authors, e.g. Virtanen and

    Paavo (1987), Pagan and Schwert (1990), Leseps and

    Morell (1997), Crawford and Fratantoni (2003), etc.

    have used ARIMA model as proposed by Box-Jenkins to

    forecast different time series such as stock index returns,

    exchange rates, etc. and compared it with different models

    like Markov Switching, Regime Switching GARCH, etc.

    The results show that ARIMA model performed well

    compared to other models.

    However, the major limitation of the ARIMA model is the

    pre-assumed linear form of the model. The approximation

    of linear models to complex real-world nancial time series

    problem is not always satisfactory. Financial time series

    are considered as highly non-linear where the mean and

    variance of the series changes overtime. Grudnitski and

    Osburn (1993) in their study stated that there is noisy non-

    linear process present in the prices. Moreover, Refenes

    et al. (1994) in their study also indicated that traditional

    statistical techniques for forecasting have serious

    limitations with respect to applications with non-linearitiesin the data set such as stock indices. Hence, detecting this

    hidden non-linear relationship and the application of non-

    linear methods may help in improving the forecasting

    *Manish Kumar,Research Scholar, Department of Management Studies, Indian Institute of Technology, Madras, Chennai, India**M. Thenmozhi of Management Studies, Indian Institute of Technology, Madras, Chennai, India

  • 8/12/2019 IJFM Journal

    5/46

    2 International Journal of Financial Management Volume 1 Issue 1 January 2012

    accuracy. Recent developments in the theory of Neural

    Computation provide interesting mathematical tools for

    such a new kind of nancial analysis. One of the popular

    and powerful tools in this area is Articial Neural Network.

    The major advantage of neural networks is their exiblenon-linear modeling capability (Donaldson and Kamstra,

    1996). Neural networks have exible non-linear function

    mapping capability, which can approximate any continuous

    function with arbitrarily desired accuracy. Due to their

    success in nancial forecasting, neural networks have

    been adopted as an alternative method in the prediction of

    stock prices, exchange rates, etc.

    A number of studies (Refenes et al.(1987), Kimoto et al.

    (1990), Takashi et al. (1990), Kryzanowski et al.(1993),

    McCluskey (1993), Bansal and Vishwanatahn (1993),

    Refenes (1994), Donaldson and Kamstra (1996), Zirilli

    (1997)) have investigated Neural Network model for

    predicting the stock market and the results support the

    importance of the model. Castiglianc (2001) and Phua

    et al.(2003) have used Neural Network to forecast stock

    index increment. Yao et al. (2002) used Neural Network

    for forecasting option price; Jasic and Wood (2004)

    examined the daily stock market indices of S&P 500,

    DAX, TOPIX and FTSE for protability of trades based

    on Neural Network prediction. Thus, many studies have

    shown that neural networks are better and can serve as a

    better prediction model that can overcome many of the

    drawbacks associated with the traditional techniques.

    In the Indian context, Thenmozhi (2001) examined

    the feasibility of Neural Network in predicting the

    movement of the daily and weekly returns of BSE

    Index. The architecture used four inputs, which are the

    four consecutive daily returns and one output being

    the prediction of return on the fth day. The study uses

    multiplayer perceptron with backpropagation algorithm.

    The results show that predictive powers of both the models

    (daily and weekly return) were low. Pant and Rao (2003)

    in their work used ANN for estimating the daily return of

    the BSE Sensex using randomized backpropagation. The

    study was based on the daily price time series of BSE

    Sensex. It used four different architectures of three-layer

    Neural Network that consist of three-input parameters and

    one output parameter. Results indicate that ANN based

    forecasting method is superior to the nave strategy of

    holding the stocks. Manish and Thenmozhi (2004) used

    backpropagation neural networks and compared it with

    a linear ARIMA model for forecasting exchange rate like

    INR/USD and the Stock index return. Results indicate

    that ANN based forecasting method is superior to the

    linear ARIMA model.

    The recent researches have focused on using hybrid model

    or combining various models of forecasting to improve

    the forecasting accuracy. The idea behind the model

    combination is to use the unique advantageous features

    of each model to accurately analyse different patterns in

    the data (Reid (1968) and Bates and Granger (1969)).

    The study of Newbold and Granger (1974), Makridakis

    et al.(1982), Makridakis (1989), Clemen, (1989), Palm

    and Zellner (1992) and Makridakis et al.(1993) suggests

    that by combining several different models, forecasting

    accuracy can often be improved. In addition, the combined

    model is more robust and exible with regard to the

    possible structure change in the data.

    There have been some studies suggesting hybrid models,

    combining the ARIMA model and neural networks. An

    important motivation to combine different forecasting is

    that one cannot identify the true process of the time series,

    i.e. the time series under examination is generated from

    a linear or non-linear underlying process. Moreover,

    the time series data often contain both linear and non-

    linear patterns. Hence, different models may be tried in

    approximating the underlying process. However, the nal

    selected model is not necessarily the best for future uses

    due to many potential inuencing factors such as sampling

    variation, model uncertainty, and structure change.Therefore, combining different models can increase

    the chance to capture different patterns in the data with

    increased accuracy and improve forecasting performance

    drastically. By combining different methods, the problem

    of model selection can be eased with little extra effort and

    this can serve as a universal model thus saving time and

    effort (Zhang, 2003).

    Voort et al. (1996) used this combination to forecast

    short-term trafc ow. Their technique used a Kohonen

    self-organizing map as an initial classier; with each class

    having an individually tuned ARIMA model associated

    with it. Su et al.(1997) used the hybrid model to forecast

    a time series of reliability data with growth trend. Their

    results showed that the hybrid model produced better

    forecasts than either the ARIMA model or the Neural

    Network by itself. Wedding and Cios (1996) described

    a combining methodology using radial basis function

    networks and the BoxJenkins models. Luxhoj et

    al. (1996) presented a hybrid econometric and ANN

    approach for sales forecasting. Pelikan et al.(1992) and

  • 8/12/2019 IJFM Journal

    6/46

    Stock Index Return Forecasting and Trading Strategy Using Hybrid ARIMA-Neural Network Model 3

    Ginzburg and Horn (1994) proposed to combine several

    feed-forward neural networks to improve time series

    forecasting accuracy. Zhang (2003) used the hybrid

    methodology to forecast the three well-known data sets

    the Wolfs sunspot data, the Canadian lynx data, and theBritish pound/US dollar exchange rate data. Experimental

    results with real data sets indicate that the combined model

    can be an effective way to improve forecasting accuracy

    achieved by either of the models used separately. Hence,

    there is strong evidence in the literature that hybrid models

    are more robust and are more accurate over the individual

    models.

    Recently, Tugba and Casey (2005) using Zhang (2003)

    approach showed that the combined forecast can

    underperforms signicantly compared to its constituents

    performances. They demonstrated these using ninemonthly time series data sets, Auto Regressive (AR)

    linear and Time Delay Neural Network models (TDNN).

    The last 12 values were reserved for testing, the preceding

    12 values for validation, whilst the rest were used for

    training. For ve of the nine data sets, the linear AR and

    TDNN models outperform the ARIMA Neural Network

    hybrids, albeit with similar levels of performance for

    two of these data sets. They concluded that despite the

    popularity of hybrid models, which rely upon the success

    of their components, single models themselves can be

    sufcient.

    Although, different hybrid ARIMA-ANN model has been

    developed, in earlier studies related to hybrid models,

    auto, regressive terms have been used as input to the

    Neural Network. The residuals of ARIMA model has

    been modeled using Neural Network. Zhang (2003) in his

    study assumed that the non-linear patterns will always be

    present in the residuals of the linear ARIMA model, which

    can be modeled using articial neural networks. Moreover,

    there is an assumption that the relationship between the

    linear and non-linear components is additive and this

    may degrade performance if the relationship is different,

    e.g. multiplicative. Such assumptions are likely to lead

    to unwanted degeneration of performance if the opposite

    situation occurs Tugba and Casey (2005). Clemen (1989)

    and Granger and Ramanathan (1984) in their study states

    that the lack of success using the combination models may

    be attributed to the performance of benchmark models.

    The performance of the benchmark models was so much

    weaker than that of the neural network models that it is

    unlikely that combining relatively poor models with an

    otherwise good one will outperform the good model

    alone. Hence, the result of the recent study on the hybrid

    ARIMA-ANN model is mixed.

    Moreover, the other key problems associated with these

    studies are as follows. These studies use simulated

    or articial data set for the analysis and the number of

    observation for training and the test data were very

    low (Zhang (2003). The degree of accuracy and the

    acceptability of certain forecasting models are measured

    by the estimates deviations from the observed values, i.e.

    MAE, RMSE, etc. but turning point forecast capability

    using sign and direction test has not been considered

    ((Zhang (2003), and Tugba and Casey (2005)). Leung

    et al. (2000) in their study suggested that depending on

    the trading strategies adopted by investors, forecasting

    methods based on minimizing forecast error may not beadequate to meet their objectives. In other words, trading

    driven by a certain forecast with a small forecast error

    may not be as protable as trading guided by an accurate

    prediction of the direction or sign of return. Hence, the

    competing models must be evaluated not only in terms of

    MAE, RMSE etc., but also in terms of sign and direction

    test. The other drawback of the previous studies isthat,

    none of the studies evaluated their models based on the

    trading performance. Statistical measures of performance

    are often inappropriate for nancial applications. The

    forecast error may have been minimized during model

    estimation, but model with a small forecast error may

    not be as protable as a model selected using nancial

    criteria such as risk adjusted measure of return Leung et

    al.(2000) Evaluations of models using nancial criteria

    through a trading experiment may be more appropriate.

    Although, there are studies addressing the issue of

    forecasting nancial time series such as stock market

    index most of the empirical ndings are associated with

    the developed nancial markets (UK, USA, and Japan).

    However, few studies exist in the literature which predicts

    the nancial time series of emerging markets. Nowadays,

    many international investment bankers and brokerage

    rms have major stakes in overseas markets. Harvey

    (1995) found emerging market returns are more likely

    to be inuenced by local information than developed

    markets; in fact, emerging market returns are generally

    more predictable than developed market returns. Indian

    stock markets have received relatively little attention until

    recently. Now there is more interest and research on Indian

    market data due to the countrys rapid growth and potential

  • 8/12/2019 IJFM Journal

    7/46

    4 International Journal of Financial Management Volume 2 Issue 1 January 2012

    opportunities for investors. Since the establishment of

    National Stock Exchange (NSE), the nancial markets

    in this Asian country have attracted considerable global

    investments.

    Given this notion, this study examines the applicability of

    hybrid ARIMA-neural network models for predicting the

    daily return of the S&P CNX NIFTY Index and compares

    it with isolated ARIMA and neural network model. The

    study differs from earlier studies in several ways. Firstly,

    the study develops the hybrid ARIMA-ANN models. In the

    rst stage of this study, the ARIMA and an articial neural

    network model is used to forecast the variable of interest. In

    second stage hybrid ARIMA-ANN models are developed.

    The hybrid ARIMA-ANN model is similar to the Zhang

    (2003). Secondly, the different competing models are

    rigorously compared using two approaches. Firstly, thestudy examines the out-of-sample forecasts generated by

    different competing models employing non penalty-based

    performance criteria such as Root Mean Square Error

    (RMSE), Mean Absolute Percentage Error (MAPE) and

    Mean Absolute Error (MAE) and performance criteria

    based on direction and sign change such as Directional

    Symmetry (DS), Correct Up trend (CU) and Correct Down

    trend (CD) goodness of Forecast Measures Thirdly, the

    different competing models are also examined in terms of

    trading performance and economic criteria via a trading

    experiment. For example, the study uses the return forecasts

    from the different models in a simple trading strategy (buy

    when the forecast is positive and sell when forecast is

    negative) and compare pay-offs to determine which model

    can serve as a useful forecasting tool.

    Thus, the major contribution of this study will be (1) to nd

    out the appropriate neural network and ARIMA model for

    NIFTY return series; (2) to nd out the appropriate hybrid

    ARIMA-ANN for NIFTY return series; (3) to demonstrate

    and verify the predictability of S&P CNX NIFTY Index

    return by applying the hybrid ARIMA-neural network

    models; (4) to compare the performance of the hybrid

    model with that of individual ARIMA and neural network

    model in terms of forecasting accuracy using non penalty-

    based performance criteria such as Root Mean Square

    Error (RMSE), Normalized Mean Square Error (NMSE)

    and Mean Absolute Error (MAE) and performance criteria

    based on direction and sign change such as Directional

    Symmetry (DS), Correct Up Trend (CU) and Correct

    Down trend (CD); (5) to evaluate the three models in

    terms of trading performance via a trading experiment.

    The remaining portion of this paper is organized as

    follows. The data used in the study, the details of hybrid

    approach and the benchmark models are introduced in

    Section 2. The empirical results from the real data sets

    are discussed in Section 3. Finally, Section 4 contains theconcluding remarks.

    2. Data and Methodology

    The study is based on the daily closing prices for the S&P

    CNX NIFTY Index. The series span the period from 1st

    January 2000 to 31stMarch 2005 totaling a 1,319 trading

    days. The data is divided into two periods- the rst period

    runs from 1stJanuary, 2000 to 26thDecember, 2003 (1,000

    observations) used for model estimation and is classied

    as in-sample, while the second period runs from 27th

    December, 2003 to 31stMarch, 2005 (319 observations)is reserved for out-of-sample forecasting and evaluation.

    The division amounts to approximately 25 per cent being

    retained for out-of-sample purposes.

    The use of data in levels in the stock market has many

    problems: stock market price movements are generally

    non-stationary and quite random in nature, and therefore

    not very suitable for learning purposes. To overcome

    these problems, the NIFTY series is transformed into

    rates of return. Given the price level P1, P2, , Pt , the

    rate of return at time t is formed by: Rt= (Pt /Pt 1) 1.

    An advantage of using a returns series is that it helpsin making the time series stationary, a useful statistical

    property.

    2.1 Forecasng Methodology

    The premise of this research is to examine the use of hybrid

    models in NIFTY returns forecasting and trading models.

    Their performance is compared with univariate linear

    ARIMA model and a non-linear backpropagation neural

    network. As all of these methods are well-documented in

    the literature, an outline of the methods is given below.

    2.1.1 ARIMA Methodology

    Popularly known as Box-Jenkins (BJ) methodology, but

    technically known as ARIMA methodology, assumes that

    the future values of a time series have a clear and denite

    functional relationship with current, past values and white

    noise. The mixed auto regressive model of order (p,q)

    denoted as ARMA (p,q) is dened as

  • 8/12/2019 IJFM Journal

    8/46

    Stock Index Return Forecasting and Trading Strategy Using Hybrid ARIMA-Neural Network Model 5

    t= q+ F1Zt 1+ F2Zt 2+ FpZtp+ F0at+ F1at 1 + F2at 2+ Fqatq

    Where t is the time series and at is an uncorrelated

    random error term with zero mean and constant variance

    and qrepresents a constant term.

    The time series models are based on the assumption that

    the time series involved are stationary. But many a time

    series are not stationary, that is they are integrated. If

    a time series are integrated of order 1 (i.e., it is I (1)),

    and its rst difference is I (0), it is said to be stationary.

    Similarly, if a time series is I (2), its second difference is I

    (0). In general, if a time series is I (d) after differencing it

    dtimes. Then I (0) series will be obtained.

    If a time series is differenced d times to make it stationary

    and then ARMA (p, q) model is applied to it, then theoriginal time series is ARIMA (p, d, q), that is an

    autoregressive integrated moving average time series.

    The Box-Jenkins models are implemented using

    E-Views 4. The correlogram, which are simply the

    plots of Autocorrelation Functions (ACFs) and Partial

    Autocorrelation Functions (PACFs) against the lag length,

    is used in identifying the signicant ACFs and PACFs.

    The lags of ACF and PACF whose probability value is less

    than 5% are signicant and are identied. The possible

    models are developed from these plots for the NIFTY

    Index returns series. The best model for forecasting isidentied by considering the information criteria, i.e.

    Akaike Information Criteria (AIC) and Schwarz Bayesian

    Information Criteria (SBIC). It is also an accepted

    statistical paradigm that the correctly specied model

    for the historical data will also be the optimal model for

    forecasting. Hence, it is reasonable to compare the hybrid

    model and the best neural network results with those of

    ARIMA models.

    2.1.2 Neural Network Methodology

    In this study, one of the widely used ANN models, thefeed forward neural network is used for nancial time

    series forecasting. Usually, the NN model consists of an

    input layer, an output layer and one or more hidden layers.

    The hidden layers can capture the non-linear relationship

    between variables. Each layer consists of multiple neurons

    that are connected to neurons in adjacent layers.

    A neural network can be trained by the historical

    data of a time series in order to capture the non-linear

    characteristics of the specic time series. The model

    parameters (connection weights and node biases) will

    be adjusted iteratively by a process of minimizing the

    forecasting errors. For time series forecasting, the nal

    computational form of the ANN model is as

    Yt= ao+ Y ao w f a w Y t jj

    q

    j ij t i t

    i

    p

    = + + +=

    -=

    1 1

    ( ) e

    where aj(j= 0,1,2, q) is a bias on thejthunit, and wij(i=

    1,2,,p;j= 1,2, q) is the connection weight between

    layers of the model, f (.) is the transfer function of thehidden layer, p is the number of input nodes and q is the

    number of hidden nodes. Actually, the ANN model in (2)

    performs a non-linear functional mapping from the past

    observation (Yt 1, Yt 2,, Ytp) to the future value Yt,

    i.e.,

    Yt= j (Yt 1, Yt 2, Yt 3,, Ytp, n) + xt

    where v is a vector of all parameters and jis a function

    determined by the network structure and connection

    weights. Thus, in some senses, the ANN model is equivalent

    to a Nonlinear Auto Regressive (NAR) model.

    2.1.3 Model Formulaon

    This study employs a three-layer backpropagated neural

    network to forecast NIFTY Index returns. The return

    series of NIFTY Index are fed to the neural network

    model to forecast the next period return in this model.

    For example, the inputs to a 5x1 neural network are

    NXi 4, NXi 3, NXi 2, NXi 1 and NXi while the output

    of the neural network is NXi + 1, the next days NIFTY

    return, where NXi stands for the current days NIFTY

    return. The architecture of the neural network is denoted

    byX-Y-Z. TheX-Y-Zstands for a neural network withX

    neurons in input layer, Y neurons in hidden layer, andZ

    neurons in output layer. Only one output node is deployed

    in the output layer since one-step-ahead forecast is made

    in this study. The number of input nodes and hidden nodes

    are not specied a priori. This will be selected throughexperiment. This study uses tansigmoid function for

    the nodes in the input layer for backpropagated neural

    network, while tansigmoid function and pure linear

    function are used at hidden layers and output layers.

    The number of input nodes is probably the most critical

    decision variable for a time series-forecasting problem

    since it contains important information about the data.

    In this study, the number of input nodes corresponds

  • 8/12/2019 IJFM Journal

    9/46

    6 International Journal of Financial Management Volume 2 Issue 1 January 2012

    to the number of lagged returns observations used to

    discover the underlying pattern in a time series and to

    make forecasts for future values. Currently, there is no

    theory suggesting the appropriate number of input nodes.

    But ideally it would be better to have a small number ofessential nodes, since this can unveil the unique features

    embedded in the data. Too few or too many input nodes

    can affect either the learning or prediction capability of

    the network. This study resorts to experimentation in the

    network construction process. The network construction

    process has been evaluated with six levels of the number

    of input nodes ranging from 1 to 6.

    The number of hidden nodes plays a very important role

    too. These hidden neurons enable the network to detect the

    feature, to capture the pattern in the data, and to perform

    complicated non-linear mapping between input andoutput variables. Hornik et al.(1989) in their theoretical

    work found that single hidden layer is sufcient for the

    network to approximate any complex non-linear function

    with any desired accuracy. Most authors use only one

    hidden layer for forecasting purposes. This study employs

    three-layer BPN to forecast the daily returns of NIFTY

    returns. Five levels of hidden nodes, 1, 2, 3, 4 and 5 have

    been experimented. The combination of six input nodes

    and ve hidden nodes yields a total of 30 different neural

    network architectures. These in turn are being considered

    for each in-sample training set for the NIFTY returns the

    backpropagation neural network models.

    This study uses backpropagation algorithm to train the

    BPN. Backpropagation is the most widely used algorithm

    for supervised learning with neural networks. The study

    uses MATLAB 6.5 to build and train neural network. The

    MATLAB program works with default parameter values

    of weight, assigned by the MATLAB.

    2.1.4 The Hybrid Methodology

    This study develops the hybrid models to forecast the S&P

    CNX NIFTY Index return. The forecasting method usinghybrid models initiates with the basic time series data on

    NIFTY Index return. It may be reasonable to consider

    a time series to be composed of a linear autocorrelation

    structure and a non-linear component. A hybrid model

    comprising a linear and a non-linear component has

    been employed in the experiments (Zhang, 2003): It is

    represented as

    Yt=Lt+Nt

    whereLtdenotes the linear component andNtdenotes the

    non-linear component. These two components have to be

    estimated from the data. These data then enter the rst

    stage of the ARIMA to account for a linear component;

    hence the residuals from the linear model will containonly the non-linear relationship. Let etdenote the residual

    components at time t from the linear model, then

    et= YtLt

    where Lt is the forecast value for time t. Any signicant

    non-linear pattern in the residuals will indicate the

    limitation of the ARIMA. By modeling residuals using

    ANNs, non-linear relationships can be discovered. With n

    input nodes, the ANN model for the residuals will be

    et= (et 1, et - 2,, etn) + et

    where is a non-linear function determined by the neural

    network and etis the random error. Denote the forecast

    from ANN as Nt , the combined forecast will be

    Yt = Lt + Nt

    The proposed methodology of the hybrid system by

    Zhang (2003) consists of two stages. In the rst stage, an

    ARIMA model is tted to the time series data to capture

    the linear part of the problem. In the second stage, an

    appropriate neural network model is developed to forecast

    the residuals from the ARIMA model. The hybrid model

    exploits the unique feature and strength of ARIMA model

    as well as ANN model in determining different patterns.

    So, the above hybrid ARIMA neural network model

    uses the following: (a) forecast residuals Nt (results of

    ARIMA model) of neural network and (b) the forecast Lt

    (results of ARIMA model).

    The optimal architecture of hybrid model that captures

    the non-linear patterns of residuals of ARIMA model

    is formed in the same way as discussed in the model

    formulation of neural network methodology.

    2.2 Forecasng Accuracy and Trading

    Simulaon

    To compare the performance of the models, it is necessary

    to evaluate them on previously unseen data. This situation

    is likely to be the closest to a true forecasting or trading

    situation. To achieve this, all models were maintained

    with an identical out-of-sample period allowing a direct

  • 8/12/2019 IJFM Journal

    10/46

    Stock Index Return Forecasting and Trading Strategy Using Hybrid ARIMA-Neural Network Model 7

    comparison of their forecasting accuracy and trading

    performance.

    2.2.1 Out-of-Sample Forecasng Accuracy

    Measures

    This study uses six widely used statistical metrics such

    as Mean Absolute Percentage Error (MAPE), Mean

    Absolute Error (MAE), Root Mean Square Error (RMSE),

    Directional Symmetry (DS), Correct Up trend (CU) and

    Correct Down trend (CD) to evaluate the forecasting

    capabilities between the three models. RMSE, MAPE and

    MAE measure the deviation between actual and forecast

    value. The smaller the values of MAE, MAPE and RMSE,

    the closer are the predicted time series values to that of the

    actual value. It is observed that RMSE, MAE or MAPEfunctions that are used for nancial forecasting models

    may not make sense in the nancial context. Caldwell

    (1992) gives a general review for the performance metrics.

    Yao et al. (1996) use the correctness of the trend to judge

    the performance of neural network forecasting model.

    So this study uses additional evaluation measures, which

    includes the calculation of correct matching number of

    the actual and predicted values with respect to sign and

    directional change. DS measures correctness in predicted

    directions while CU and CD measure the correctness of

    predicted up and down trend, respectively, in terms of

    percentage. Higher value of these metrics indicates betterdirection and time information. The statistical performance

    measures used to analyze the forecasting techniques are

    presented in Appendix 1.

    This study also uses other measures to test the models

    ability to predict turning points. A correct turning point

    forecast requires that:

    Sign (Yt Yt t

    - ) = Sign (Yt Yt 1)

    Where Ytand Yt represents the actual and predicted value

    at time t.

    The ability of a model to forecast turning points can be

    measured by a fourth evaluation method developed by

    Cumby and Modest (1987). This model denes a forecast

    direction variable Ftand an actual direction variable At

    such that

    At= 1 if DYt> 0 andAt= 0 if DYt0

    Ft= 1 if DYt > 0 and Ft= 0 if DYt

    0

    Where DYt is the amount of change in actual variablesbetween time t1 and t; and DYt is the amount of changein forecasting variables between time t1 and t.

    Cumby and Modest (1987) suggest the following

    regression equation:

    Ft= a0+ a1A1+ et

    where etis error term; and a1is the slope of this linear

    equation. Here, a1 should be positive and signicantly

    different from 0 in order to demonstrate those FtandAt

    have a linear relationship. This reects the ability of a

    forecasting model to capture the turning points of a time

    series.

    2.2.2 Out-of-Sample Trading PerformanceMeasures

    Statistical performance measures are often inappropriate

    for nancial applications. Typically, modeling techniques

    are optimized using a mathematical criterion, but

    ultimately the results are analyzed on a nancial criterion

    upon which it is not optimized. In other words, the forecast

    error may have been minimized during model estimation,

    but the evaluation of the true merit should be based on the

    performance of a trading strategy.

    Hence, this study uses a simple trading strategy toevaluate the performance of different models. The

    operational detail of the trading is as follows. This study

    considered an index in place of a single stock to avoid (or

    average out) the impact of company-specic news on the

    prediction of only one stock, given that the prediction is

    performed by taking into account past prices only. In the

    simulated market set up for experimenting the proposed

    methodology, a virtual trader can buy or sell stock index

    fund on the stock index concerned, and both short and

    long positions can be taken over the index.

    Assume that a certain amount of seed money is used in thistrading experiment. The seed money is used to buy stock

    index funds when the prediction shows a rise in the stock

    index price. To calculate the prot, the stock index funds

    are bought or sold at the same time. It should be noted that

    the price of the stock index fund is directly proportional to

    the index level so that the virtual investor can gain from

    both a fall and rise of the stock index price. The trading

    strategy is to go long when the model predicts that the

  • 8/12/2019 IJFM Journal

    11/46

    8 International Journal of Financial Management Volume 2 Issue 1 January 2012

    stock index price will rise, i.e. the forecast is positive and

    a sell otherwise. Then the stock index funds will be held at

    hand until the next turning point that the model predicts.

    For many traders and analysts, market direction is more

    important than the value of the forecast itself, as in

    nancial markets money can be made simply by knowing

    the direction the series will move. The trading performance

    measures used to analyze the forecasting techniques are

    presented in Appendix 2. Some of the more important

    measures include the Annualized return, Annualized

    volatility, Sharpe ratio, maximum drawdown and average

    gain/loss ratio. The Sharpe ratio is a risk-adjusted measure

    of return, with higher ratios preferred to those that are

    lower; the maximum drawdown is a measure of downside

    risk and the average gain/loss ratio is a measure of overall

    gain, for which a value above one is preferred.

    The application of these measures may be a better standard

    for determining the quality of the forecasts. After all, the

    nancial gain from a given strategy depends on trading

    performance, not on forecast accuracy.

    3. Results

    3.1 Summary Stascs

    The mean, median, standard deviation, skewness, and

    kurtosis for the NIFTY Index return are given in Table

    1. The analysis shows that the sample mean of dailyreturns of NIFTY returns is not statistically different from

    zero. The measure of skewness and kurtosis indicates

    that the distributions of the return series are different

    from the standard normal distributions. They reveal a

    slight skewness and high kurtosis, which is common in

    nancial time series data

    Table 1 Descriptive Statistics of NIFTY Returns

    Mean Median Std. Dev Skewness Kurtosis Observations

    .000306 .001202 .015392 .621756 8.409610 1319

    3.2 Staonarity Test

    The Augmented Dickey Fuller test and Philip Perrons test

    statistics as given in Table 2 indicate that the rate of return

    of the NIFTY Index is stationary as the absolute value of

    statistics is greater than the critical value and thus, the

    time series is suitable for modeling.

    Table 2 Unit Root Test for the NIFTY Return

    Series

    Augmented Dickey Fuller Test Phillip Perron Test

    Statistic Critical Value Statistic Critical Value

    15.64312 3.4382 32.62406 3.4382

    3.3 ARIMA Model

    The correlogram is used to identify the number of

    signicant spikes of ACF or PACF of the NIFTY

    Index return series. The lags of ACF and PACF whose

    probability value is less than 5% are signicant and are

    identied. Several ARMA specications were tried out.

    After considering all possible models and looking at AIC

    and SBIC as given in Appendix 3, the ARIMA (1 1 2)model are identied for NIFTY return.

    In order to verify the adequacy of ARIMA model, the

    study uses one of the popular diagnostics test known as

    Breusch-Godfrey LM Test. Here the test is used to check

    the presence of serial correlation in the residuals. It allows

    us to examine the relationship between residuals and

    several of its lagged values at the same time. The Null

    Hypotheses to be tested is there is no serial correlation.

    If the predictability value is greater than 5% then we can

    accept the Null Hypotheses (at 95% condence levels)

    which means there is no serial correlation in the series.The Breusch-Godfrey LM Test for serial correlation of

    residuals as shown in Table 3 suggests that, in case of

    NIFTY return the ARIMA model captures the entire

    serial correlation and the residual do not exhibit any serial

    correlation.

    Table 3 Breusch-Godfrey Serial Correlation LM

    Test for the NIFTY return

    F-Statistics Probability Obs*R Square Probability

    1.004433 0.366624 2.016921 0.364780

    3.4 Neural Network Model

    The combination of six input nodes and ve hidden nodes

    yields a total of 30 different neural network architecture

    which are being considered for each in-sample training

    set for NIFTY return and the residuals of ARIMA model.

    The best network architecture thus obtained from this

  • 8/12/2019 IJFM Journal

    12/46

    Stock Index Return Forecasting and Trading Strategy Using Hybrid ARIMA-Neural Network Model 9

    experiment for NIFTY return and residuals of ARIMA

    model on the basis of least error (MSE) associated with

    the model is 3-2-1, i.e. three input nodes in input layer,

    two nodes in hidden layer and one node in output layer.

    The Neural network model 3 2 1 provides better t tothe NIFTY returns and ARIMA residuals series.

    3.5 Hybrid Model

    One hidden layer is used to develop the hybrid models.

    The study experimented with different nodes or neurons

    in hidden layer, which varies from one to ve for the

    different hybrid models. The output layer has one neuron

    or node, which is the forecast value. The study uses

    MSE to select the architecture. The hybrid model which

    uses forecast results of ARIMA and the forecast residual

    (results of ARIMA) of neural network has three hidden

    neurons in hidden layer and three input neurons in input

    layer.

    3.6 Forecast Evaluaon

    Out-of-Sample Forecasng Accuracy Results

    For the NIFTY return series one-period-ahead forecast

    were produced by the three models namely hybrid modes,

    ARIMA and neural networks. The predictive performance

    of the three models is summarized in Table 4.

    it is observed that hybrid model outperforms the other two

    models. MAE and RMSE achieved by the hybrid model

    is quite low indicating that there is a smaller deviation

    between the actual and predicted values in hybrid model.

    Between neural network and ARIMA models, the former

    performs better in terms of the three most commonly

    used criteria i.e. MAE, RMSE and NMSE. The results

    of the hybrid model show that by combining two models

    together, the overall forecasting errors can be reduced

    considerably.

    In terms of other performance metrics like correct up

    (CU) and correct down (CD), hybrid models yields

    better performance than the other models. It is really the

    directional symmetry (DS) measure that singles out the

    neural network model as the best performer, predicting

    most accurately 52.38 per cent of the time. These three

    criteria provide a good measure of the consistency in

    prediction of the time series direction.

    Between hybrid model, neural network and ARIMA

    models, the latter performs worst almost all of the times

    in terms of performance metrics like direction sign and

    change and non penalty based measure like MAE, NMSE

    and RMSE. A majority decision rule would therefore

    select the hybrid model as the overall best model.

    Turning Point Evaluaon

    Table 4 Out-of-Sample Prediction Accuracy

    Model Performance Metrics

    MAE RMSE NMSE DS CU CD

    Hybrid Model 0.011063 0.015916 0.937088 0.514285 0.540541 0.481928

    Neural Network 0.011107 0.016153 0.965219 0.523809 0.527027 0.481928

    ARIMA 0.011125 0.016275 0.979853 0.425396 0.405405 0.439759

    The main purpose of any nancial time series modelingis to determine how well forecasts from estimated

    models perform based on the non penalty based measure

    of performance such as MAE, RMSE and NMSE. The

    forecasting accuracy statistics provide very conclusive

    results. A glance at these values shows the superiority of

    hybrid model over the two other models. Comparing the

    forecasting performance of the three models in terms of

    MAE, NMSE and RMSE for the NIFTY return time series;

    The turning point evaluation method using Cumby andModest (1987) regression equation is shown in Table 5

    for all the models.

    The tratio of the slope coefcient a1of all the models

    shows that it is statistically different from zero for the

    NIFTY Index return time series. This implies that all

    models have good turning point forecasting ability.

  • 8/12/2019 IJFM Journal

    13/46

  • 8/12/2019 IJFM Journal

    14/46

    Stock Index Return Forecasting and Trading Strategy Using Hybrid ARIMA-Neural Network Model 11

    The hybrid model has the highest number of transactions

    at 213, while the ARIMA model has the lowest at 183.

    In addition, the hybrid model has the highest average

    gain/loss ratio at 1.43, highest maximum daily prot at

    12.24 per cent and lowest maximum daily loss at 3.94per cent, while the ARIMA model has the lowest average

    gain/loss at 1.11 and highest maximum daily loss at 8.29

    per cent. A simple neural network model outperforms the

    hybrid model and ARIMA models in terms of percentage

    of winning and per cent of losing trades with a value of

    55.44 per cent and 44.55 per cent respectively. As with

    statistical performance measures, nancial criteria clearly

    single out the hybrid model as the one with the most

    consistent performance: it is therefore considered the

    best model for this particular application.

    Zhang (2003) in their study found that hybrid ARIMA-NN model outperform the individual neural network

    and ARIMA model. However, the studies uses only non

    penalty based criteria (MAE, RMSE etc) to evaluate the

    forecast model. The turning point forecast capability test

    has not been considered. Moreover, these studies does

    not evaluated their models based on the performance of

    trading. The present study generally supports the nding

    of the Zhang (2003) and contradicts the ndings of Tugba

    and Casey (2005). The results validate the ndings with

    real nancial time series data and also using by evaluating

    the performance of models using a trading strategy.

    4. Conclusion

    This study reports an empirical work which investigates the

    usefulness of hybrid (ARIMA and neural network) model in

    forecasting and trading the S&P CNX NIFTY Index return.

    The linear ARIMA model and the non-linear ANN model

    are used in combination, aiming to capture different forms

    of relationship in the time series data. The performance of

    the hybrid model was measured statistically and nancially

    via a trading simulation. The logic behind the trading

    simulation is that, if prot from a trading simulation is

    compared solely on the basis of statistical measures, the

    optimum model from a nancial perspective would rarely

    be chosen. The hybrid model was benchmarked against

    traditional forecasting techniques such as ARIMA and

    non-linear technique like neural network to determine any

    added value to the forecasting process.

    The empirical results with the NIFTY returns clearly

    suggest that the hybrid model is able to outperform each

    component model used in isolation. A neural network

    architecture of 3-2-1 and ARIMA (1 1 2) is the best

    identied model for forecasting the returns of NIFTY

    Index. With the prediction, signicant prots were

    obtained for a chosen testing period.

    The results show that useful prediction could be made

    for NIFTY without the use of extensive market data or

    knowledge. It also shows how an 81.40 per cent annual

    return and a Sharpe ratio of 3.15 could be achieved by

    using the hybrid model. The present results indicate that

    the hybrid ARIMA neural network models is important

    in out-of-sample forecasting and trading performance,

    and are in line with Wedding and Cios (1996) and Zhang

    (2003) who found that hybrid model works well and

    found to outperform the isolated models. The results are

    in contrary with the results of Tugba and Casey (2005).

    Thus, the study shows that hybrid ARIMA-neural network

    model outperforms in forecasting stock index returns both

    in terms of forecasting accuracy and in generating trading

    returns. Probably this type of hybrid model could be used

    by policy makers in forecasting nancial and economic

    data, apart from traders, borrowers, FIIs and arbitrageurs

    developing trading models that leads to better investment

    decision and returns.

    References

    Bansal, R.., and Viswanathan, S. (1993), No Arbitrage

    and Arbitrage Pricing: A new Approach,Journal of

    Finance. 48(4), 1231-1262.

    Bates, J. M., and Granger, C. W. J., (1969), The

    Combination of Forecasts, Operation Research

    quarterly, 20, 451468.

    Caldwell, R.B., (1995), Performances Metrics for Neural

    Network-based Trading System Development,

    NeuroVest Journal, 3 (2), 22-26.

    Castiglianc, F. (2001) Forecasting Price Increments

    Using an Articial Neural Network, Advances in

    Complex Systems, 4 (1), 45-56.

    Clemen, R., (1989), Combining Forecasts: A Review

    and Annotated Bibliography with Discussion,

    International Journal of Forecasting, 5, 559608.

    Crawford G.W., and Fratantoni M.C., (2003), Assessing

    the Forecasting Performance of Regime-Switching,

    ARIMA and GARCH Models of House Prices,

    Real Estate Economics, 31 (2), 223-243.

    Donaldons, R. G., and Kamstra, M. (1996), A New

    Dividend Forecasting Procedure Rejects Bubbles in

  • 8/12/2019 IJFM Journal

    15/46

  • 8/12/2019 IJFM Journal

    16/46

    Stock Index Return Forecasting and Trading Strategy Using Hybrid ARIMA-Neural Network Model 13

    Terui, N., and van Dijk, H., (2002), Combined Forecasts

    from Linear and Non-linear Time Series Models,

    International Journal of Forecasting, 18, 421438.

    Thenmozhi, M., (2001), Predictability of BSE Returns

    using Neural Networks: An exploratory study,International Conference on Industrial Mathematics,

    IIT Madras, India.

    Tugba T. T, and Casey, M. C., (2005), A Comparative

    Study of Autoregressive Neural Network Hybrids,

    Neural Networks, 18, 781789.

    Virtanen. I., and Paavo Yli-Olli., (1987), Forecasting

    Stock Market Prices in a Thin Security Market,

    OMEGA: International Journal of Management

    Science, 15 (2), 145-155.

    Voort, V.D., Dougherty, M., Watson, M., (1996),

    Combining Kohonen Maps with ARIMA Time

    Series Models to Forecast Trafc Flow, Transport.Research Circular, 4C (5), 307318.

    Wedding, D. K. II., and Cios, K. J., (1996), Time

    Series Forecasting by Combining RBF Networks,

    Certainty Factors, and the BoxJenkins Model,

    Neurocomputing, 10, 149168.

    Yao, J.T., and Poh, H.L., (1996), Equity Forecasting: a

    Case Study on the KLSE Index, Neural Networks

    in Financial Engineering, Proceedings of 3rd

    International Conference On Neural Networks in

    the Capital Markets, Oct 1995, London, A.-P N.

    Refenes, Y. Abu-Mostafa, J. Moody and A. Weigend

    (Eds.), World Scientic , 341-353.Yao, J.T., Li, Y., and Tan, C.L., (2002) Option Price

    Forecasting Using Neural Networks Omega, 28,

    455466.

    Zhang, G.P., (2003), Time Series Forecasting Using

    a Hybrid ARIMA and Neural Network Model,

    Neurocomputing, 50, 159175.

    Zirilli, J., (1997), Financial Prediction Using Neural

    Network, London, International Thompson

    Computer Press.

    Appendix 1

    1. NMSE =( )

    ( )

    Y Y

    Y Y

    t t

    t t

    -

    -

    2

    2

    2. MAE =| |Y Y

    N

    t t-

    3. RMSE =S ( )Y Y

    N

    t t- 2

    4. DS =100

    Ndt

    t , dt=

    1

    0

    1 1if

    Otherwise

    ( ) ( )Y Y Y Y t t t t- -

    - -

    5. CU =

    d

    k

    t

    t

    tt

    dt=1 0 0

    01 1 1If

    Otherwise

    ( ) ; ( ) ( )Y Y Y Y Y Y t t t t t t - > - -

    - - - ,

    kt=1 0

    01If

    Otherwise

    ( )Y Yt t- >

    -

    6. CD =

    d

    k

    tt

    tt

    dt=1 0 0

    01 1 1If

    Otherwise

    ( ) ; ( ) ( )Y Y Y Y Y Y t t t t t t - < - -

    - - - ,

    kt=1 0

    01If

    Otherwise

    ( )Y Yt t-

  • 8/12/2019 IJFM Journal

    17/46

    14 International Journal of Financial Management Volume 2 Issue 1 January 2012

    MD= min ( max( ))t N

    tc

    tc

    i t

    R R= =

    -1 1

    8. % Winning trades : WT= 100

    F

    NT

    tt

    N

    =

    1

    where Ft= 1 transaction prott> 0

    9. % Losing trades :LT= 100

    G

    NT

    tt

    N

    =

    1

    where Gt= 1 transaction prott< 0

    10. Number of up periods : Nup = number ofRt> 0

    11. Number of down periods : Ndown = number ofRt< 0

    12. Number of transaction :NT= Ltt

    N

    =

    1

    whereLt= 1 if trading signalt= trading signalt 1

    13. Total trading days : Number ofRts

    14. Avg. gain in up periods : AG = (sum of allRt> 0)/

    Nup

    15. Avg. loss in down periods : AL = (sum of allRt< 0)/

    Ndown

    16. Prot T-statistics : T-statistics = N RA

    As

    17. Number of periods daily returns : NPR = Qtt

    N

    =

    1

    rise

    where Qt= 1 if Yt> 0 else Qt= 0

    18. Number of periods daily returns : NPF = Stt

    N

    =

    1

    falls

    where St= 1 if Y

    t< 0 else S

    t= 0

    19. Number of winning up periods :NWU= Btt

    N

    =

    1

    whereBt= 1 ifRt> 0 and Yt> 0 elseBt= 0

    20. Number of winning down periods:NWD= Ett

    N

    =

    1

    whereEt= 1 ifRt> 0 and Yt< 0 elseEt= 0

    21. Winning up periods (%) : WUP = 100*(NWU/

    NPR)

    22. Winning down periods (%) : WDP = 100 * (NWD/

    NPF)

    Appendix 3

    Model AIC SBIC

    ARIMA (1 1 1) 5.567444 5.552709

    ARIMA (2 1 1) 5.571542 5.551880

    ARIMA (1 1 2) 5.572746 5.553099

    ARIMA (2 1 2) 5.571903 5.547325

    ARIMA (1 1 3) 5.570930 5.546371

    ARIMA (2 1 3) 5.571832 5.552339

    ** ARIMA (1 1 2) has the lowest AIC and SBIC value...

  • 8/12/2019 IJFM Journal

    18/46

    Abstrac t

    In the current unpredictable and volatile economic

    environment, the investment avenues have been

    changing rapidly. The stock market is one of them.There are multiple unpredictable factors which affect

    the performance of the global market time to time. In

    recent years, we have observed an unprecedented

    growth in the complexity of instruments for trading

    and risk management in international market and thus

    issues of international stock market linkages and the

    relationship between the Asian stock markets and

    others stock markets deserves to be investigated

    to justify the risk and return factor after the Asian

    Financial Crisis. This is the first exhaustive study of its

    kind on linkages and the interrelationship between the

    Asian stock markets and others stock markets namely,

    Malaysia (Kuala), Singapore (Strait), Philippines (Pse),

    Indonesia (Jakarta), China (Shanghai), Japan (Nikkie),

    Korea (Kospi), and the US (Dow) and reveal that stock

    markets of Thailand, Japan and China are exogenous

    before, during and after the crisis respectively. For the

    purpose of study composite sample consisting of all the

    stocks based on weekly stock indexes is been used to

    construct panels and for the same the total samples

    are separated into three sub periods January 2005

    to December 2007, January 2008 to December 2008,

    January 2009 to December 2009. The first part of

    paper gives an insight about the Asian and US stock

    markets and its various aspects. The second partconsists of data and their analysis, collected from the

    various websites and manuals. The short-term linkage

    was tested through granger causality test based on

    A Study on the Linkages of Asian and the

    US Stock Markets

    S.M. Tariq Zafar*, D.S. Chaubey**and S.R. Sharma***

    1. Introducon

    Over the past decade, business has continued to grow with

    pace and became more globalised than ever and resulting

    demand for finance in many folds. With the growing

    global trade the needs to communicate across the borders

    has correspondingly multiplied, consequently there is

    globalization of capital markets which became integral

    part of economy and also custodian of socio-economic

    integrity and play instrumental role in global economic

    growth and have a deep impact on overall capitalemployment. Company in one country is borrowing in the

    capital market of another country. In an open economic

    competition and in the era of globalization performance

    * S.M. Tariq Zafar, Director, Charak Institute of Business Management, Lucknow, India. ** Dr D.S. Chaubey, Director, UIBS, Dehradun, India.*** Dr S.R. SharmaDr. S.R.Sharma, Director, MIT, Dehradun, India

    Vector Error Correction Model (VECM), and the co

    integration or long-term linkage was through Engle-

    Granger co integration test. The empirical results show

    that the number of significant cointegrating vector is

    higher during the crisis periods compared to otherperiods and concludes that the linkages between the

    Asian and the US stock markets are stronger in the

    post-crisis period

    Keywords: VECM, Unit Root, DF test, ADF test,

    Shanghai, Nikkie, Kospi, Dow, Stock market

    JEl Codes: COI, C22, C51, C53, G12, G14, G15,

    H83, F3

  • 8/12/2019 IJFM Journal

    19/46

    16 International Journal of Financial Management Volume 2 Issue 1 January 2012

    of organization changes day by day, investment avenues

    became global and expanded gradually with continuous

    strength. With growing capital market and introduction

    of high breed financial instrument for common benefit

    it became important to understand the concept of globalstock markets its investment, risk and return significance

    in economic development.

    To achieve complete efficiency in stock market may not be

    possible because of difference in the economic, political,

    legal and cultural environment which cast there shadow

    on shareholders return. In general perception, investment

    is regarded as a sacrifice of certain present value of the

    uncertain future reward or allocation of funds to assets

    that are expected to yield some gain over a period of time

    which exclusively involves strategic decisions like, where

    to invest, when to invest and how to invest. Since every

    investor have different behavior with common appetite to

    invest in those investment policies which may generate

    maximum return with minimum risk. To have return

    with safe growth and investment in unpredictable global

    capital / financial market, investors have to establish some

    diversified policies and procedures to shed the risk and

    equate the invested expectations through global portfolio

    which is an appropriate selection and collection of

    investments held by institutions or a private individual.

    In order to attract the investors globally, market reforms

    are inevitable which may fuel competition in financial

    market with safe return to investors and produce positive

    financial vibration which explores capital market efficiency

    and zenith the growth. Balance financial employment and

    motivating returns on investment need healthy and vibrant

    capital market. Positive stock markets encourage common

    investors to invest in security market and maximize the

    wealth. In order to cater the global economic competition

    and varied requirements of savers and investors wide

    spectrum of financial intermediaries with high breed

    investments offerings both in money market and capital

    market with nations central banks as the apex body have

    come into existence across the globe. Further in effortsto manage unexplored challenges and capitalize the

    global investment opportunities to the fullest, pursuance

    of nations policy of liberalization, privatization, and

    globalization has fueled overall competitiveness in

    global economies and respective stock markets. Global

    economies offering tremendous opportunities to stock

    market and other financial organizations to explore

    expand and diversify their product range and operators

    besides improving their operational efficiency. Effective

    execution of strategy is contingent upon adoption of new

    high breed global financial instrument, technology better

    possess of credit and risk appraisal, fund management,

    product diversification, responsive structure, internal

    control and availability of talented man power.

    2. Causes of Recession and Its Impact

    The financial crisis which in respect recognized as great

    recession was sponsored by greedy uncontrolled fall

    of world trade currency the US dollars by 40% which

    chopped investors interest earning on their assets by more

    than 80% during 2001-2008. It happens due to Federal

    Reserve Banks secret mission with no accountability. The

    over emphasis upon debt instead of income through hyper

    inflating the money supply US dollar began to fall in value

    and touched historic low against the major competitivecurrencies. The Federal Reserve Bank printed too many

    dollars to cover mounting budget deficit which was $164.7

    billion in the third quarter as to avoid a recession which

    has created an imbalance between income and assets and

    caused drastic inflation a hidden tax which was 4.3% in

    2007 and was 1% higher than its GDP. It has been noted

    that inflation sabotaged US growth and swallowed 15% of

    its economy and its Infection spread and created insecurity

    globally. The IMF sees inflation rate nearly doubling by

    almost 12 % in emerging and developing world due to

    recession impact.

    To control inflation you need to control money supply but

    in USA money supply exploded astonishingly 3,000%

    from $302 billion in 1959 to $11.5 trillion in 2006 and thus

    dollars purchasing power collapsed almost by 85% during

    the period resulting America becoming largest debtor

    on earth which mounted to & 53 trillion. Once largest

    creditor nation in the world US became largest debtor

    nation and its overall debt observation was considered in

    between 70 to 100 trillion dollars with increasing trend

    of more than $7.4 billion per day. In the previous year of

    recession, its total debt grown almost by & 4.3 trillion,

    comparatively 5.5 times larger than US, GDP. Interest onforeign debt rose almost by $2.2 trillion, Business and

    financial sector debt grown by 7-11%, Almost 80% of

    American debts which stood to around $42 trillion were

    created since 1990. Highest debt ratio in world history,

    thats $175,154 per man, woman and child or $700,616

    per family of four.

    In May 2007 trade surplus of US recorded historic

    negative trade deficit of $827 billion. Since 1985, its

  • 8/12/2019 IJFM Journal

    20/46

    A Study on the Linkages of Asian and the US Stock Markets 17

    international deficit is approximately 35% larger than

    social security spending, almost 50% larger than all

    defense spending, and 2.5 times higher than Medicare. Its

    merchandise ratio to national income has grown to 18%.

    Its overall merchandise trade defi

    cit of $815 billion in2007 was due to its trade performance and created history

    by establishing second largest negative trade balance. Its

    cumulative deficit mounted to 6.6 trillion dollars which

    caused negative international net worth of $5 trillion and

    its core manufacturing base reduced by 60%,

    Americas 27% of the economy depends on international

    trade in goods; foreign interests have more control over

    the US economy than Americans and interest on foreign

    debt rose almost by $2.2 trillion. They own about $9

    trillion of US financial assets, including 13% of all stock,

    13 % of agencies, and 27 % of corporate bonds. Its

    foreign reserve and universal reserve fell from 50% of

    the worlds total to 2.4% ratio, a 95% drop. As 2006 SDR

    data revealed that USA has $69 billion as compared to

    China $1.04 trillion, Japan $882 billion reserves. During

    the period China and Japan together own 40% of the total

    world international reserve ($5 Trillion) and US share is

    just 1% with tremendously growing international debts in

    comparison to it mare $69 billion international reserve.

    It is also revealed that 80% of worlds official foreign

    exchange reserves which is about $2.2 trillion dollars are

    held by Asian central banks.

    Since 1990s with declining US manufacturing base, its per

    capita energy consumption have increased heavily. Each

    day the world oil market consumes 76 million barrels,

    America, with 5% of the total world population consumes

    three times more oil than its total productivity, and it has

    consumption of 20 million barrels per day (26 % of world

    total oil production). The difference between Americans

    production and consumption during the period was almost

    75% which produced deficit gap of 15.5 million per day

    and collectively 5.7 billion per year. Further, its population

    increased 70 million and in its comparison in last 30 years

    its oil reserves declined by 42% and it produces only 20billion barrels oil.

    In comparison to economic growth its spending ratios and

    its employee number increases faster than its population.

    Federal government spending ratios reached almost

    25% of its total national income which was 10 times

    more growth in government spending than its economy

    growth since 1930 and consumed 15% of its economy.

    During the period education productivity dropped by

    71% and unemployment rate increases to it all time

    high. There was very unusual situation. The economy

    grew at a 0.6% annual rate over the last two quarters, the

    slowest pace since 2001 recession. In 2007, US housing

    market worsened and were one of the major causes forthe subprime crises that were witnessed and resulted

    in collapse of large financial institutions, the bailout of

    banks by national government and downturns in stock

    markets around the world. Years of easy liquidity and

    relatively low interest rate regime fuelled an economic

    boom across the globe, driven largely by credit expansion

    and magnificent rise in asset prices. Default and losses on

    other loan types also increased significantly as the crises

    expanded from the housing market to other parts of the

    economy. However, the obscure problem of plenty began

    to surface in the US mortgage economy, with disastrous

    parameter. Mortgage prices in US declined 40% in lessthan a year and impacted the economy of US in large.

    Policymakers did not recognize the increasingly important

    role played by financial institutions such as investment

    banks and hedge funds, also known as the shadow banking

    system which resulted number of established and leading

    banks collapsed as some of them were not commercial

    banks but was connected with commercial banks through

    derivatives transactions. Large number of the banks

    was heavily dependent for short term funds on money

    market mutual fund that provided wholesale fund, fled

    the market. In fact banks were not felled losses on theirsubprime loans. They were felled by losses on mortgage

    backed securities caused by drying up of liquidity and

    by the loss of nerve of market participant, confidence.

    Questions regarding banks solvency, declines in credit

    availability, and damaged investors confidence had an

    impact on global stock markets, where securities suffered

    large losses during late 2008 and early 2009.

    The US tried to maintain and expands consumption

    rather than producing real goods and savings thus facing

    uncontrollable debt in relation to size of its economy. In

    order to develop stability in overall market it providedfunds to encourage lending and restore faith in commercial

    paper markets in addition it also bailout integral financial

    institutions and implemented economic stimulus programs

    to promote and protect market confidence. The (FED)

    chairmen acknowledged that the central bank faced

    increasingly contradictory pressure of slowing growth

    and rising consumer prices. In past 1 % decline in US

    growth impacted growth in emerging economies by 0.5%

    to 1 % depending on trade and financial links to US and

  • 8/12/2019 IJFM Journal

    21/46

    18 International Journal of Financial Management Volume 2 Issue 1 January 2012

    Week dollars impacted Asian exports in particular.

    Since 1997, Asia attracted almost 50% of the total

    capital inflow. It is due to large population which make

    Asia darling of investor. The economies of Southeast

    Asia handsomely maintained a high interest rate whichattracted investors who are found of high return. With the

    support of World Bank, IMF, regional economies of Asia

    experienced high growth rates but the recession of US

    impacted global market to a large extent and market with

    close interrelation suffered in multiple way. In addition to

    recession, Asian countries weak domesticfinancial system,

    free international capitals flows, fluctuating market and

    investors sentiment and hype hazard economic policies

    also played supportive role.

    3. Literature ReviewLiterature review is an organized and scientific approach

    of study which require collection and systematic analysis

    of literatures in the selected area of the researchers in

    which they have limited or no exposure. A deep survey of

    literature exposed the truth that large number of researchers,

    independent and professional research institutions and

    academicians has carried out extensive research in the

    field of international stock market linkages and many

    are concerned about the relationship between the Asian

    stock markets and others after the Asian Financial Crisis,

    Sharpe and Cooper, Basu, Irala,Brown, S.V. RamanaRao, Zafar S.M, Tariq, Naliniprava Tripathy,M. Kabir

    Hassan, Neal C. Maroney Hassan Monir El-Sadyand

    Ahmad Telfah, Barman and Smanta, Myong Jae Lezand

    SooCheong (Shawn) Jang,and they produced important

    findings which pave multiple dimensions and set ultimate

    standard. It has been noted that large number of the studies

    has been carried out in developed economy or developed

    countries stock market, few studies in this context is been

    carried out in developing economy or countries stock

    market. Further outcome of these studies reflects that

    studies which are carried out in developing nations are not

    scientific and lack authenticity and validity thus keepingdeveloping nations in mind this paper initiate humble

    beginning in this, respects.

    Ayshanapalli et al. (1995) in his study he examine the

    existence of a common stochastic trend between the US

    and the Asian stock market movements during pre- and

    post-October 1987 periods. For the study he took data

    for the time period January 1, 1986 to May 12, 1992

    from Singapore, Thailand, Malaysia, Philippines, Hong

    Kong Japan and United States. He used cointegration

    and error-correction model for his study and found that

    influence of the US stock market innovations were in

    excess during the post-October 1987 period. The study

    concluded with the fact that Asian stock markets are lessintegrated with Japans stock market than they are with

    the US stock market, Masih and Masih (1999) in his

    study examined the long-term and short-term dynamic

    linkages between the international and Asian emerging

    stock markets, and tried to quantify the extent of the

    Asian stock market fluctuations which are explained

    by intra-regional contagion effect. The finding of study

    at the global level, confirm the widely flouting doctrine

    of united State leadership in short and long term stock

    market and its existing relationship between OECD stock

    markets along with the emerging Asian stock markets.

    At Southeast Asia, the results of the study confirm thedominating role of Hong Kong stock market which was

    even expected too, Malliaropulos and Priestly (1999)

    in their study investigated the predictable component of

    Southeast Asia Stock market. For the purpose data from

    January 1988 to December 1995 at weakly frequency

    basis was taken as sample and the study findings were

    assessed by adjusting stocks returns for potential time

    varying expected returns and partial integration of this

    emerging market into world capital market. The study

    clearly indicated the danger of testing market efficiency

    without sufficiently adjusting the stock returns with care,

    especially time variation in the expected return and partial

    integration of local markets into world class, Sheng and

    Tu (2000)in their study they investigated linkages among

    national stock markets before and during the period of

    their Asian Financial crisis through co-integration and

    variance decomposition analysis. For the study they

    took data from July 1, 1996 to June 30, 1998 on daily

    closing prices basis of the New York S&P 500 and the 11

    major Asia-Pacific equity market indexes. The outcomes

    of the study reveals that Southeast Asian countries have

    strong relationship in comparison to the Northeast Asian

    countries and thefi

    ndings also indicate that there were nocointegrational relationship prior to Asianfinancial crises.

    Further the forecast error variance decomposition finds

    that the degree of exogeneity for the stock markets has

    been reduced to an extent, Manning (2002)in his study

    tried to investigate the co- movement of stock markets in

    South Asia, concurrently taking the United States to be

    an external stock market. For the study they took the data

    compromised weekly and quarterly information on stock

    indexes and US dollars series for the US, Hong Kong,

  • 8/12/2019 IJFM Journal

    22/46

  • 8/12/2019 IJFM Journal

    23/46

    20 International Journal of Financial Management Volume 2 Issue 1 January 2012

    4.1 Methodology

    The study is done with special reference to relationship

    between the Asian stock markets and others after the Asian

    Financial Crisis. For the purpose, data from January 2005

    to December 2009 from Malaysia (Kuala), Singapore

    (Strait), Philippines (Pse), Indonesia (Jakarta), China

    (Shanghai), Japan (Nikkie), Korea (Kospi), and the US

    (Dow) stock markets were mainly extracted. Three panels

    namely Panel A, Panel B and panel C containing different

    equity prices and their variances listed on the selected stock

    exchanges have been drawn. Simple random technique has

    been used, analytical and descriptive research design is

    been adopted which based on the secondary data collected

    from the websites, annual reports and journals, published

    periodicals, stock exchange and various other related sites.

    A composite sample, consisting of all the stocks is beenused to construct panels and for the same the total samples

    are separated into three sub periods. First period is pre-

    crises period spanning from January 2005 to December

    2007 denoted as panel A. Second period is during crises

    period spanning from January 2008 to December denoted

    by panel B and third period is post-crises period spanning

    from January 2009 to December 2009 denoted by Panel

    C. To interpreting the results Dickey-Fuller (ADF) unit

    root test, Phillips- Perron (PP) test and Granger- causality

    based on Vector Error Correction Model (VECM) are

    used.

    4.2 Tools Used for Analysis

    In this study, for interpreting the results and to determine

    stationarity of the data series the statistical and econometric

    tools are been used. The very first step is to examine the

    stationary of the variables. Further unit root test is applied

    to check the stationary of the series by using the Dickey-

    Fuller (DF) test, Augmented Dickey-Fuller (ADF) test.

    For robustness of unit root test results the series is also

    tested by using the Phillip- Perron (PP) test. Then the

    cointegration test (Engle-Granger cointegration test) is

    used to estimate the long run equilibrium relationshipamong the variables. Finally, Granger causality test is

    applied to test the short-run relationship between the

    stationary series which deals with financial time series

    data.

    4.3 Unit Root

    Unit Root test is applied to check the stationary of the

    series (Gujarati, 2003; and Enders, 2005). The stationary

    condition here has been tested using the Dickey Fuller,

    Augmented Dickey Fuller and Philip-Peron unit root

    tests.

    4.4 DickeyFuller Unit Root Test (DF Test)

    In statistics, the DickeyFuller test tests whether a unit

    root is present in an autoregressive model. It is named

    after the statisticians D. A. Dickey and W. A. Fuller, who

    developed the test in 1979.).

    A simple AR (1) model is

    yt= ryt 1+ ut

    Where yt is the variable of interest, t is the time index,

    ris a coefficient, and utis the error term. A unit root is

    present if |r| = 1. The model would be non-stationary inthis case.

    The regression model can be written as

    yt= (r 1)yt 1+ ut= dyt 1+ ut

    Where is the first difference operator

    4.5 Augmented Dickey Fuller Test (ADF Test)

    In statistics and econometrics, an Augmented Dickey

    Fuller test (ADF) is a test for a unit root in a time seriessample. It constructs a parameter correction for higher

    order correlation, by adding lag difference of the time

    series. It is consists of a regression of the first difference of

    the series against the series lagged once, lagged difference

    terms, and optionally, a constant and tie trend.

    Dyt= a+ bt+ gyt 1+ d1Dyt 1+ + dpDytp+ et,

    Consequently, there are three main versions of the test

    which are as commonly used in Dickey Fuller unit root

    test (DF test) and in Augmented Dickey Fuller test (ADF

    test). Each version of the test has its own critical value

    which depends on the size of the sample. They are as: 1. Test for a unit root:

    yt= dyt 1+ ut

    2. Test for a unit root with drift:

    yt= a0+ dyt 1+ ut

    3. Test for a unit root with drift and deterministic time

    trend:

    yt= a0+ a1t+ dyt 1+ ut

  • 8/12/2019 IJFM Journal

    24/46

    A Study on the Linkages of Asian and the US Stock Markets 21

    4.6 Philip-Peron Test

    In statistics, the Phillips- Perron test is a unit root test. It is

    used in time series analysis to test the null hypothesis that

    a time series is I