1.
ALTERNATIVE NEURAL NETS APPROACHES FOR ENHANCING STOCK PICKING USING EARNINGS FORECASTS Giuseppe Galloppo University of Rome Tor Vergata Mauro Aliano University of Rome Tor VergataABSTRACTIn the last decade, neural networks have drawn noticeable attention from many computer and operationsresearchers. Essentially, a Neural Network is a non-parametric estimation technique. It does not make anydistributional assumption regarding the underlying variable. In various studies Artificial Neural Network(ANN) models, have been proved to be powerful predictive tools, where a variable is explained by a set ofexplanatory variables without assuming any structural or linear relationship among the variables. Somecritical success factors to train ANN are the network architecture, network design algorithm, trainingalgorithm, and stop training conditions.While some previous studies, have found encouraging results with using this artificial intelligence techniqueto predict the movements of established financial markets, there is a lack of studies examining the stockpicking ability of different ANN paradigms, taking into account analyst earning forecasts as input data. Ourapproach is based on the notion, that trading strategies guided by forecasts of stock earnings, may beeffective and lead to higher profits.This paper attempts to enhance the stock selection process by employing ANN to select stocks in the Usstock market. Neural networks are used to identify stocks for the portfolio which are likely to outperform themarket, given the forecast earnings information of stocks. Our purpose is to compare various ANN models,to identify critical predictors to forecast stock prices and to see, which model show the best stock pickingability, increasing in this way investment strategy profitability for the professionals in the market. Thecompetiting models, are examined in terms of various trading performance and economic criteria, likeannualized return, Sharpe ratio, maximum drawdown, annualized volatility, average gain/loss ratio etc via atrading experiment.Empirical experimentation suggests that by using artificial neural networks for nonlinear predictions there ispotential economic value for subsequent portfolio choices. ANN-based investment strategies, once financialanalyst earning forecast are considered, obtain higher returns than other investment strategies examined inthis study. Consequently, we find that the returns obtained from the equally weighted portfolio formed by thestocks selected by neural networks, outperform those generated by the buy and hold strategy, computed witha benchmark index for a given period under investigation. The influences of the length of investment horizonand the commission rate are also considered. The present study does not support efficient market hypotheses.Keywords: Financial Forecasting, Efficient Market hypotheses, Artificial Neural Network,Investment Strategy.JEL Classifications: G12, G14, C451. INTROInterest in financial markets has increased in the last couple of decade, among fund managers, policy makers,investors, borrowers, corporate treasurers, and specialized traders Forecasting the future returns has alwaysbeen a major concern for the players in stock markets and one of the most challenging applications studiedby researchers and practitioners extensively. The prediction of financial market is a very complex task,because the financial time series are inherently noisy and non-stationary and more it is often argued thatfinancial market is very efficient. Fama (1970) defined efficient market hypothesis (EMH) where the idea isa market in which security prices at any time “fully reflect” all available information both for firmsproduction-investment decisions, and investors’ securities selection. More in EMH context no investor is in aposition to make unexploited profit opportunities by forecasting futures prices on the basis of past prices. On
2.
the other hand, a large number of researchers, investors, analysts, practitioners etc uses different techniquesto forecast the stock index and prices. In the last decade, applications associated with artificial neuralnetwork (ANN) have drawn noticeable attention in both academic and corporate research. Neural networkshave flexible nonlinear function mapping capability, where a variable is explained with a set of explanatoryvariables without assuming any structural or linear relationship among the variables and is capable toapproximate any continuous function with arbitrarily desired accuracy. They are also capable of continuouslearning through the new information received. In the financial market, it has proven to be very powerfulpredictive tools and a reliable instrument with good error tolerance, capable of handling large andcomplicated information and achieving satisfactory forecasting results. Due to their success in financialforecasting, neural networks have been adopted as an alternative method in the prediction of stock prices,exchange rates etc.There exist vast literatures which concentrate on the predictability of stock market return. In almost all cases,the performance metrics and the acceptability of the proposed models are measured by the deviations offorecast value from the actual values. The drawback of the previous studies is, none of the studies evaluatedthe effect of nonlinear predictions on portfolio performances driven by analyst and the question whetherthese predictions are economically exploitable has been neglected in the literature. This paper contributes tofill the gap, and we think that the results of this study are significant value addition to the trading decisions inthe stock index futures. This paper intends to test the forecasting ability of ANN models in case of Niftyindex returns when financial analyst information is taken into account. Quarterly time series data of a stockportfolio is analyzed using two layer architecture of the ANN and various input data combinations.Specifically the major contributions of this study are as follows: 1. to demonstrate financial analyst information contribution in forecasting stock return and in asset allocation decision in a non linear context; 2. to compare the out-of-sample one step forecast of various ANN architectures; 3. whether the length of the investment horizon has a significant impact on the quality of the forecasts and on asset allocation performance.The different competiting models are rigorously compared using two approaches. Firstly, the study examinethe out-of-sample forecasts generated by different competing models employing statistical criteria such asgoodness of forecast measures (i.e., root mean squared errors (RMSE)). To provide a more completeevaluation of the models, our comparison is based on not only the performance statistics but also the tradingprofits. Thus, this study develops a set of trading strategies of which the performances are compared withthose generated by the buy and hold strategy, and the investment strategies guided by the forecasts estimatedby a parametric forecasting approaches, namely the random walk. Random walk is also used because it is anatural benchmark that is based of the efficient market hypothesis. Many sophisticated forecasting modelsare not able to outperform the “naive” random walk model. Given the notion that a prediction with littleforecast error does not necessarily translate into capital gain, nevertheless empirical results show that theanalyst information ANN-based investment strategies obtain higher returns than other investment strategiesexamined in this study.The remaining portion of this paper is organized as follows. Beside this introduction, utilization of neuralnetworks and researched and results of similar empirical works are presented. Section 3 describes the datasetand covers details how neural networks have been designed to perform the task of forecasting the indexreturns and how to translate it in an investors portfolio decision. Then, the results of forecasting arepresented and discussed in Section 4. Section 5 describes the proposed index trading strategies which aredriven by the forecasts made by various forecasting models. Finally, section 6 summarizes the main findingswith concluding observations.2. LITERATURE REVIEWA survey of the literature has not revealed any papers which purposes is in testing artificial neural networks(ANN) for making forecast by using as input, information, recommendations and earnings forecast, derivedfrom financial analysts via IBES data.
3.
Liu and Song (2001) examined analysts forecast of earnings for internet companies surrounding the marketcrash in March 2000. They reported that analysts were more optimistic before than after the March 2000period suggesting that analysts’ optimism may have caused the stock market bubble.O’Brien and Tian (2006) conclude that analysts were more optimistic in their recommendations for Internetcompanies than non-Internet during the 1990 bubble period.2.1 EVIDENCE OF RETURN PREDICTABILITYThere exists considerable evidence showing that stock returns are to some extent predictable. A critical viewon return predictability of risky assets is taken by Valkanov (2003), Ang and Bekaert (2007), and Goyal andWelch (2003; 2008). Most of the research is conducted using data from well-established stock markets suchas the U.S., Western Europe, and Japan. Ferson and Harvey (1993) examine 18 international equity markets,some of which are found in developing economies. The study provides evidence of returns predictability.Harvey (1995) focuses on emerging markets by looking at the returns of more than 800 equities from 20emerging markets including Taiwan. He finds that the degree of predictability in the emerging markets isgreater than that found in the developed markets. In addition, local information plays a much more importantrole in predicting returns in the emerging markets than in the developed markets. This characteristic helpsexplaining the difference in predictability between the two kinds of markets.For the U.S., several studies examine the cross-sectional relationship between stock returns and fundamentalvariables. Variables such as earnings yield, cash flow yield, book-to-market ratio, and size are shown to havesome power in predicting stock returns. In earlier studies, during the 1980s, valuation ratios were used topredict future returns, starting with dividend yields, Banz and Breen (1986), Jaffee, Keim and Westerfield(1989), and Fama and French (1992) are good examples of this group of researchAlso, Campbell and Shiller (1988a; 1988b) found that dividend yields are positively correlated with futurereturns. More recently, Kothari and Shanken (1997), Ponti® and Schall (1998), Lamont (1998), Stambaugh(1999), Lewellen (2004), and Campbell and Yogo (2006) examined the predictability of returns by financialratios. They show that book-to-market ratios and dividend yields have predictive power for subsequent stockmarket returns.2.2 RELATED RESEARCHES IN ANNAlthough a comprehensive review of the literature available on the subject is beyond the scope of this paper,we tried to accommodate the most relevant studies from across the world in respect of the application ofANN models for forecasting index returns. On the whole a number of studies have investigated neuralnetwork model for predicting the stock market and the results support the importance of the model. The firstsignificant application of the concepts of neural network models in stock market context was initiated forquestioning the validity of the efficient market hypothesis by examining the forecasting accuracy of theneural network models on IBM stock’s daily returns (White, 1988). The growing interest in the applicationsof ANN particularly in finance and stock markets started catching the interest of researchers during earlynineties, and eventually became one of the most explored techniques of prediction in stock index returns.Antonio et al. (1996), in their research, used the 240-day trade information of the Santiago Stock Exchangeas samples and the indices and transaction volume of the preceding 10 days as the input data, trying topredict the overnight closing indices of the San Diego Stock Exchange through the neural network approach.The result showed that, through the neural network approach, he achieved an accuracy rate of 63.3% inpredicting directions in the rising range of the stock market; and a 74.7% accuracy rate in the falling range.Another study on forecasting of stock market prices has been done by Ramon Lawrence (1997). This studyreveals the ability of neural network to discover patterns in nonlinear and chaotic system more accuratelythat other current forecasting tools.Tsai et al. (1999), tried to predict the best timing for investment by integrating various technical indices andconstructing a stock forecasting model based on neural networks. The result was that, through the crossutilization of neural network and stop-loss strategies, one can effectively forecast the best timing for stockpurchase and achieve better returns from the in vestment. Fernando et al. (2000) used the Back PropagationNetwork (BPN) to construct his forecasting model for Madrid Stock Exchange General Indices. The result ofhis empirical study also showed that the model is an effective forecasting model for the Madrid StockExchange General Indices and helped to achieve better investment return.
4.
Al-Hindi and Al-Hasan (2002) selected seven Saudi Arabian companies from varying sectors and depictedan efficient prediction ability of the neural networks with 2-5-1 structure. By using technical analysisindicators (momentum MACD, etc.), Diler (2003) studied on predicting the direction of the Istanbul StockExchange (ISE) for the following day. The results of the study presented that the direction of the IMKB-100index could be predicted at a rate of 60.81 per cent.Pant and Rao (2003) in their work used ANN for estimating the daily return of the BSE Sensex usingrandomized back propagation. Wu (2004) adopted the Back Propagation Neural network (BPN) for his stockprice research, based on the transaction volume, trade price and the technical indices.While some studies were focused on measuring forecasting performance of neural network models based onseveral statistical and financial performance measures, there were some other studies which compared theforecasting performances of neural network models with other statistical forecasting methods (Gencay, 1998;Kim and Chun, 1998; Lim and McNeils, 1998; Lam, 2004; Rodriguez et al., 2005).Ma (2003) applied the fuzzy neural network technology in his simulated investment in Taiwan’s stockmarket. In his empirical study, he used the Taiwan’s General Index as input variables. He then compared theresults with the actual results of using merely the 12-day moving average. The discovery was that, byadopting the fuzzy neural network approach, one can avoid the misleading effect of cheat lines, which aremore likely to happen when merely using the moving average approach. The investment return, also, issignificantly better than the return achieved through the buy-hold strategy or the traditional moving averagestrategy.Manish and Thenmozhi (2003, 2004, and 2005) have used back propagation neural networks and comparedit with a linear ARIMA model for forecasting different time series like INR/USD, Stock index return, indexfuture returns etc. Results indicate that ANN based forecasting method is superior to the linear ARIMAmodels.Kyoung-Jae Kim (2006) proposes an advanced genetic Algorithm approach to instance selection in ANNsfor financial data mining. Using this approach the study could avoid the basic limitations of ANNs such asinconsistency, problems in prediction for noisy data, etc. The study produces a satisfactory forecasting in thedirection of change on Korean Stock Price Index (KSPI) using GA based ANN (GANN).Furthermore, Avci (2007) investigated the forecasting performance of the back propagation neural networkmodel for Istanbul Stock Exchange (ISE-100) index with daily frequency. Ince and Trafalis (2007) andBekiros and Georgoutsos (2008) show that neural network models can be successfully implemented forreturn predictability. More recently, Hammad, Ali, and Hall (2009) showed that Artificial Neural Network(ANN) technique provides fast convergence, high precision and strong forecasting ability of real stock price.3. DATA AND METHODOLOGY3.1 DATABASEThe main objective of this study is to determine the predictability of the ANN models in forecasting thereturns of the Nifty index making a comparison between two different ANN architectures and making acomparison between two layer architecture of the ANN and various input data combinations. with particularfocus on financial analyst information contribution in forecasting stock return and in asset allocation decisionin a non linear context. For the stated purpose, basic variable covered in our database pertain to quarterlyaverage Sp500, Nasdaq100 stock price index1. Quarterly average stock price of a portfolio composed by a setof shares quoted in sp500 and NASDAQ market. Other input data are: quarterly average of Barclays UsTreasury 3-5y free risk index, and of the, actual and forecast E/P ratio, financial analyst recommendations,quarterly by quarterly2.For the composition of the sample of stock, we looked at the database Ibes2 at those companies had forecaston average more than 20 per year within the full period (obtaining in this way about 1500 stocks, about onethird from NASDAQ market and the rest from other stock market). Starting from this sample we havedownloaded prices from DataStream.The data set covers the horizon from January 1997 to June 2003 and is divided into two periods: the firstperiod runs from 01/01/1997 to 31/12/2000 and the second period runs from January 2001 to june 2003. The1 Both indexes are used for a variety of purposes such as benchmarking fund portfolios, index based derivatives andindex funds.2 The data from financial analysts used in this study are obtained from the IBES database .
5.
first period, the in-sample estimation period, is used for model determination (i.e., specifying the modelparameters) and validation. The second period is the reserved out-of-sample evaluation period and is used tocompare the forecasts and trading performances of various models.We chose to use as testing period in this study, a very turbulent period, the so called dot com bubble period,in order to test in a better way the predictive capabilities of the different ANN schemes and to betterhighlight the contribution of financial analysts information that at that time were the subject of very muchattention. The use of data in levels in the stock market has many problems: stock market price movementsare generally non-stationary and quite random in nature, and therefore not very suitable for learningpurposes. To overcome these problems, the stock portfolio series is transformed into rates of return. The dataon return are derived from these basic variables. In this study, one-period stock market return at time point t,say Rt , is simply defined as Rt = log(Pt) – log(Pt-1 ); where Pt is a security price. An advantage of using areturns series is that it helps in making the time series stationary, a useful statistical property.3.2 ARTIFICIAL NEURAL NETWORKS MODELSThe development of ANN model usually encompasses the selection of suitable network topology and thedetermination of several key parameters associated with training. Among these decision variables are thenumber of hidden layers, the number of neurons in each of the hidden layers, the number of training cycles inan epoch, the total number of epochs in the complete training session, the learning rate, and the momentum.The artificial neural networks are non-linear systems formed by neurons (region where information areprocessed) which imitate the processing mechanism of the human brain.The connection type, number of entries, layers, exits and the type of training used are aspects which differthe types of neural networks.The most important characteristic of the neural networks, is the ability of learning with its environment andso improve its performance. This is done through an interactive process of adjustments applied to its weights,what is called training stage. One typical method for training a network is to first separate the data series intotwo disjoint sets: the training set and the test set. The network is trained (e.g., with back propagation) directlyon the training set (i.e. arrive at set of weights between two neurons). The testing set is used to test how wellthe neural network performs on new data after the network is trained.The architecture of the neural network is denoted by X-Y-Z. The X-Y-Z stands for a neural network with Xneurons in input layer, Y neurons in hidden layer, and Z neurons in output layer. This study resorts toexperimentation in the network construction process. The number of input nodes is probably the most criticaldecision variable for a time series-forecasting problem since it contains important information about the data.In this study, the number of input nodes corresponds to: the number of lagged returns observations used to discover the underlying pattern in a time series and to make forecasts for future values; financial analyst forecast; financial analyst recommendations; since this study attempt to forecast the direction of daily price change in the stock price index, technical indicators are used as input variables. Different tools are used in technical analysis, out of which two tools are taken as input parameters 3 as determined by prior research, Kim and Han (2000) and Kim (2003).The network construction process has been evaluated with two different architectures, Recurrent MLP andFeedforward Back Propagation, with five levels of the number of input nodes ranging from 1 to 6. Thecombination of 4 input nodes and two ANN architectures yields a total of 8 different neural network models.These in turn are being considered for each in-sample training set for the portfolio returns.Only one output node is deployed in the output layer since one-step-ahead forecast is made in this study. Thenumber of input nodes and hidden nodes are not specified a priori. This will be selected through experiment.This study uses one hidden layer4. The transfer function used for the output layer was the hyperbolic tangent3 Simple Moving Average of last 3 days closing Nifty values and Relative Strength Index of last 3 observations.4 The number of hidden nodes plays a very important role too. These hidden neurons enable the network to detect thefeature, to capture the pattern in the data, and to perform complicated nonlinear mapping between input and outputvariables.
6.
function. As for the transfer function used for the hidden layer, it was found after some testing that the besteffect can be achieved by using Sigmoid Function as the transfer function.Sigmoid and Hyperbolic tangent function are calculated using the following formulas.Sigmoid Function: The Output value is between 0 and 1 (1)Hyperbolic Tangent Function: Symmetrical with respect of the origin, with an output value of between -1and 1 (2)In this study, the researchers have adopted the rollover estimations (moving average window) to generate the1-step ahead forecast for stock returns for the out of sample period. It means, when a new data is received,the oldest data from the training dataset is dropped and new data is added to the dataset. The advantage of themoving average window is its ability to capture the environmental changes as it utilizes more recent data.Moreover, by utilizing such approach, the forecasting performance of neural network models would beobserved on a continuous manner.To achieve this end we consider the following strategy: the ANN models are initially trained on a subset ofthe in-sample data 29 March 1996 to 31 December 2000. The estimated model is then used to generateforecasts for the remaining in-sample-period 1 January 2001 to 31 December 2001.We adopt stricter criteria for convergence particularly we stop training condition of MSE of 1% to find thebest network during the network training.More as different ranges of value are involved, we need to avoid the situation that the significance ofvariables with a smaller range is obscured by those with a larger range in the neuron. Under thecircumstances, variables with a larger range of value will dominate the network learning and adverselyimpact the neural network training results. To avoid this undesirable situation, we need to normalize therange of value of the variables. This will improve the efficiency of the neural network training. The approachis to execute a “pre-processing” prior to the network input process to ensure that the value will always fallwithin the specified range of 0-1 (Yen, 1999). All data has to be normalized first before being used. Theformula for normalization is as follows: (3)Where x stands for the raw data before normalization; xmin stands for the minimum value of raw data priorto normalization and xmax stands for the maximum value of raw data prior to normalization.RECURRENT MLP MODEL. Although, there are several ANN models have been used for forecastingresearch, multilayer perceptron model are mathematically proved to be universal approximator for anycontinuous function5. Besides, multilayer perceptron model has become a standard forecasting tool in neuralnetwork research, especially in the area of finance and stock markets, as over 80 percent of the research iscarried out using this model (Adya and Collopy, 1998). Other advantages of using this multilayer model arethat, it can handle a very high degree of non-linear problem space very efficiently (Roy and Roy, 2008).Figure 1 presents a multilayer perceptron with multiple inputs and outputs. The lines between the nodesindicate the flow of information from one node to the next. In this particular type of neural network, theinformation flows only from the input to the output (that is, from left-to-right).5 Multilayer perceptron models are non-linear neural network models that can be used to approximate almost anyfunction with a high degree of accuracy (white 1992).
7.
Figure 1The mathematical expression for the MLP(1,8) drawn in Figure 1 is given by equation (1), where thesubscripts t from the output and input variables are suppressed to ease the exposition. Thus: (4)where f(.) is the activation logistic cumulative distribution function, are the weights for the direct signalsfrom each of the two input variables to the output variable, is the weight for the signal from each of thehidden units to the output variable, and ci,j, are the weights for the signals from each of the various inputvariables combinations to the hidden units. The network interpretation of equation (4) is as follows. Inputvariables, Xj send signals to each of the hidden units. The signal from the j-th input unit to the i-th hiddenunit is weighted by some weight denoted by ci,j, before it reaches the hidden unit number i. All signalsarriving at the hidden units are first summed and then converted to a hidden unit activation by the operationof the hidden unit activation function f(.). The next layer operates similarly with connections sent over to theoutput variable. As before, these signals are amplified by weights bi and summed. Finally, signals aretransmitted directly from the input variables to the output variable with weight aj.FEEDFORWARD MULTILAYER NETWORK. Feedforward network is a collection of interconnectedsimple processing elements. The most popular and successful one is the backpropagation neural network(BPN). A BPN is typically composed of several layers of nodes. The first or the lowest layer is an input layerwhere external information is received. The last or the highest layer is an output layer where the problemsolution is obtained. The input layer and output layer are separated by one or more intermediate layers calledthe hidden layers. The units in the network are connected in a feedforward manner, from the input layer tothe output layer. Every connection in a neural network has a weight attached to it. In backpropagationalgorithm input variables are passed forward to the hidden layer from the input layer and multiplied by theirrespective weights to compute a weighted sum of total input value to a neuron in the hidden layer. Theweighted sum is modified by a transfer function and then sent as input to neurons in the next layer (hidden oroutput). They stand for the signals thus generated from earlier layers to later layers and the signal finallyreaches the output layer. The output layer neuron re-calculates the weighted sum and applies the transferfunction to produce the output value of the signal received by it. Finally, an error signal is backpropagated tothe hidden layer in a sequence opposite to that of the input variable. The error signal is computed as thedifference between the output value of the neural network and the actual output value (also called the target
8.
value of the neural network) The weights that connect two layers are adjusted proportionally according to thecontribution of each neuron to the forecast error. This is done so as to minimize the mean squared error(MSE). This training process continues until an acceptable MSE target that is specified based on requirementis achieved.3.3 FORECASTING ACCURACY AND TRADING SIMULATIONTo compare the performance of the models, it is necessary to evaluate them on previously unseen data. Thissituation is likely to be the closest to a true forecasting or trading situation. To achieve this, all models werecompared for the out-of-sample forecasts using two different approaches, namely an out-of-sampleforecasting accuracy measures and an out-of-sample trading performance measures.This study uses root mean squared errors (RMSE), to evaluate the forecasting capabilities between thevarious ANN models. RMSE measure the deviation between actual and forecast value. The smaller thevalues of RMSE, the closer are the predicted time series values to that of the actual value.Statistical performance measures are often inappropriate for financial applications. In other words, theforecast error may have been minimized during model estimation, but the evaluation of the true merit shouldbe based on the performance of a trading strategy.We formulate a set of trading rules guided by the returns predicted by various ANN models and then wecompare results with B&H and random walk models. The empirical testing takes the form of a tradingsimulation which closely mimics the timely investment decisions faced by investors in the marketplace. Thistrading simulation also allows us to evaluate the relative economic profit of the proposed investmentstrategies. Essentially, the trading simulation investigates the influence of three experimental factors: lengthof the investment horizon, architecture and input data. The length of investment horizon is the period of timein which the portfolio stock returns are realized. This is practically the same as the horizon lengths associatedwith the predicted stock returns. Thus, three month, and twelve month investment horizons are used toimplement the forecasts.We now describe the operational details of the trading simulation.In this study, we adopted two approaches. The first concerns the choice between a risky and a risk-free asset.By the second approach an investor instead, is able to go long or short on a portfolio of assets depending onmarket prediction about future performance of the portfolio of stocks.The trading strategy is to go long on stock portfolio when the model predicts that the average stock portfolioprice will rise i.e. the forecast is positive and a sell otherwise. Then the stock portfolio will be held at handuntil the next turning point that the model predicts.The trading performance measures used to analyze the forecasting techniques are:. mean annualized return,Standard Deviation of return, and Sharpe ratio, maximum drawdown and average gain/loss ratio. The Sharperatio is a risk-adjusted measure of return, with higher ratios preferred to those that are lower.4. EMPIRICAL RESULTS4.1 FORECAST ACCURACYIn the first part of this section we will consider the results of the forecast performance of the portfolio takinginto account: the time horizon, the network architecture used and the contribution of analysts.As forecast accuracy criterion having estimated the two empirical models discussed above and obtained 1-step ahead out-of-sample forecasts for the out-of-sample period, we proceed to evaluating their relativeforecast performance comparing the out-of-sample data by calculating the root mean squared errors (RMSE)of out-of-sample predictions. The RMSE is calculated as: (5)In the second part of this section we will discuss the results of certain trading models, constructed using theestimates of the returns provided by the neural network structures, with the results of the models Buy andHold (starting from Nasdaq100 SP500 stock indexes) and Random Walk model.
9.
According to the Efficient Market Hypothesis (EMH), asset prices will follow a random walk asnews is instantaneously incorporated into prices. The random walk model is therefore a naturaltheoretical benchmark. Random walk has been used as a benchmark for forecasting ability bynumerous studies over the years. The random walk model is a very simple model to use and is oftentermed the “naïve model” because it does not involve much technical skills to implement. However,it has been shown to outperform, in terms of forecasting, many sophisticated methods. Therefore, itis a norm in the financial forecasting area to use it as a benchmark. The argument is that any newmodel that involves the implementation of advanced techniques should at least outperform the random walkmodel. Otherwise, the random walk model will be preferred since it does not involve much effort. For thosereasons, we compare the performance of our ANN models with that of the random walk model. The randomwalk model assumes that the best forecast is equal to the most recently observable observation.The results concerning the prediction of performance are set taking into account three elements: the first isthe type of architecture used (architecture type A6 or B7), the second refers to the time horizon used forpredicting the performance of portfolio (yearly or quarterly), the third relates to the contribution made byanalysts to estimate stock returns. The analysts contribution is measured by two variables: earnings per shareand recommendations8.To better articulate this part of the paper, it was considered appropriate to adopt the following logicalscheme: the results will be presenting to horizon estimation, in which we will analyze the best results for thetwo network structures used and beyond the contribution made by analysts. Finally, we examine the effectresulting of the time horizon, or in other words we will discuss the best network structures for different timehorizons.The statistical measure used to test the networks ability to predict portfolio returns is given by the RMSE(Root Mean Square Error). The RMSE is the root square of the squared sum of the differences betweenestimated and actual return. In reference to this measure it ha been computes also its standard deviation.Table 1 shows the empirical results for the quarterly time horizon. As first analysis we can observe atendency of decreasing RMSE, particularly for the architecture of type A, according to an increasing amountof information used by the network as input for the estimation of returns.In particular, the introduction of recommendations and earnings’ forecast as ANN inputs leads to a reductionin both the RMSE (RMSE Arch A column and RMSE Arch B column) and the standard deviation of theRMSE itself (Dev.st Arch A column and Dev.st Arch B Column) than the models lagged and lagged +. at(that stands for lagged variables or technical indicators as input in ANN models). That is, the forecast andrecommendations allow, given the forecast quarterly time horizon, to reduce the error in the prediction ofreturn.Regarding the comparison between architectures, not just dwelling on the best models for the structure(shown in green), one can see how the structure A allow to generate an average RMSE lower than the Bstructure, this observation is supported not only by the value average reported in the last row of Table 1 butalso by the values in column "Diff Mean" who, being almost all negative confirming the point made above.Finally, by selecting combinations of inputs that have a lower RMSE for the two architectures, we observehow the "AT + lagged + Recc" is for architecture A that has both the lowest RMSE and the lowest standarddeviation, with regard to all other models. The model with the combination of input "lagged + forecast9 + at+ recc " appears instead to be the best for architecture B.It should be noted that while for the architecture B is clear that the introduction of joint recommendationsand financial analyst forecast leads to a reduction of error of the estimate of return, it is not as clear thecontribution that this two input jointly provide for the architecture A. For the latter it would seem that, unlikethe architecture B, the combined effect recommendations and forecast does not lead to a reduction of RMSEthan using them separate. However, even if taken together allow a reduction of RMSE than the models"lagged" and "lagged + at ".6 Recurrent MLP model.7 Feedforward multilayer network.8 For the quarterly horizon were also used two indicators of technical analysis on returns namely Relative StreinghtIndex (3) and Moving Average (3).9 This variable stands for financial analyst earning forecasts as input in ANN models.
10.
In conclusion, for the quarterly time horizon the presence of recommendations and of financial analysts’earning forecasts allows, regardless of network used, involves a certain reduction of the error in estimatingthe return of stock portfolio.Table 1 Results for the prediction of returns. Quarterly horizon. RMSE Arch A Dev St Arch A RMSE arch B Dev St arch B Diff Mean Diff Dev st lagged 15.6406% 3.0715% 20.0654% 4.0075% -4.4248% -0.9360% lagged+at 5.4584% 0.6917% 4.9212% 0.3082% 0.5373% 0.3836% lagged+at+forecast 3.6456% 0.1636% 5.3848% 0.3179% -1.7392% -0.1543% lagged+at +rec 3.1669% 0.0809% 5.3357% 0.1551% -2.1688% -0.0742% lagged+at +forecast+recc 4.1232% 0.1632% 4.7788% 0.2393% -0.6556% -0.0761% mean 6.4069% 0.8342% 8.0972% 1.0056% -1.6902% -0.1714%Table 1 shows a clear contribution from the use of joint recommendations and earning forecasts to theforecast accuracy of ANN models in order to predict the future returns of the equity portfolio. However, it isappropriate to make distinctions, and indeed for the structure A the use of disjoint recomendations andearning forecasts leads to increased forecast accuracy of models than the "lagged" input models (RMSE ArchA column). For architecture B the use of disjoint these two input does not lead to a reduction both of RMSEand the standard deviation.Regarding the comparison between architectures, excluding models with just lagged input variables, the Aarchitecture presents RMSE and standard deviation lower than B architecture not only on average (last rowof Table 1) but also for individual model (excluding the first row of Table 1 let you have a look to thecolumns "Mean Diff" and "Diff Dev St"). In this regard the negative differential terms indicate the ability ofmodels with A architecture to have RMSE and standard deviations less high in value than the models witharchitecture B.In the comparison between accurate models according to the architecture used is clear, for both architectures,that one can reach better performance by using networks that consider both together recommendations andearning forecast; in this context prevails the model: “lagged + forecasts + recc" with A architecture whichshow a RMSE and a standard deviation lower than the model with architecture B.Finally, in reference to the models related to the A architecture of the neural networks is clear that therecommendations and forecasting, used either separately or together, allow to improve the model "lagged" interms of lower RMSE and in terms of lower deviation standard associated with it with respect to ANNmodels not considering these two input variables. This observation is not valid for models built with thearchitecture B, for which only the joint use of recommendations and earning forecast, reduces the RMSE andits standard deviation once compared to model with just lagged variables,.Table 2 Results for the prediction of returns. Yearly horizon. RMSE Arch A Dev St Arch A RMSE arch B Dev St arch B Diff Mean Diff Dev st lagged 9.5240% 1.1072% 6.2128% 0.6017% 5.3112% 1.0055% lagged+forecast 3.9895% 0.2053% 7.4014% 0.7131% -3.4119% -0.5078% lagged+recc 7.0480% 0.5701% 8.5655% 0.8879% -1.5175% -0.3178% lagged+forecast+recc 3.3625% 0.1434% 5.1578% 0.3711% -1.7954% -0.2277% mean 5.9810% 0.5065% 6.3344% 0.5185% -0.3534% -0.0120%By comparing the model results for the two different time horizons some considerations emerge. First of all,the architecture A with quarterly holding period has identified, on average, a number of models with RMSEand a standard deviation lower than the model with yearly time horizons. The more accurate model wascreated for the yearly horizon, with combination: "lagged + forecasts + recc". Between this two ANNarchitectures, architecture B gave rise to models with a higher RMSE on average.
11.
From the analysis previously made and reported in Tables 1 and 2, one can see how the introduction ofrecommendations and earning forecast between inputs, in general, can upgrade the error estimates of themodels. In particular, for the time horizon yearly to the joint use as input nodes of earning forecast andrecommendations allows for greater accuracy of forecasting models used. Instead, for the quarterly holdingperiod, the combined effect has a better result only for architecture B, while for the A architecture theapplication not contemporary of the earning forecast and recommendations as input node achieve modelswith lower estimation errors on average.In conclusion, for all time horizons, models that have the lowest RMSE use, jointly or severally, therecommendations and the forecast of financial analysts. Furthermore, considering the overall model andwanting to compare the results based on the architecture used, it is evident the greater accuracy offered bythe architecture A.4.2 TRADING SIMULATIONThe first simulation experiment assumes that, at the beginning of each monthly period, the investor makes anasset allocation decision of whether to shift his liquid assets into the riskfree bonds or into the stock portfoliofund (Equity-Bond10 model) . Liquid assets are defined as money that is currently not invested in either theriskfree bonds or the stock portfolio. Further, it is assumed that the money that has been invested in eitherriskfree bonds or the stock portfolio becomes illiquid and will not become liquid until the end of theinvestors chosen investment horizon. In other words, the invested money will become available after theselected investment horizon reaches its maturity. For example, suppose the investor has decided to use aninvestment horizon of three months. The money that he has invested into either riskfree bonds or the stockportfolio in the last three months is considered to have been "locked up" in asset allocation. Hence, the assetwill not be available for another round of investment decision before the security or portfolio matures.Second simulation experiment invests exclusively in the stock market with the possibility of short selling onthe basket of shares comprising in the stock portfolio (Long-Short Model).The Equity-Bond model were compiled by the following logic, if >0, then it is investing in the EquityPortfolio otherwise it invests in the bond market by buying the free risk index. represents the estimatedreturn of the Stock Portfolio made at time t-1. This forecast Questa is provided by various neural networkmodel according to the architecture, combinations of input data and forecast horizon (that correspond also tothe holding period) (eg Arch A + delayed + at + recc, quarterly). For comparison purpose the same forecastis provided also by a random walk model for which the estimated return at time t correspond to the returnobservation at time t-1.Long short model, instead, was constructed as follows, if >0, then the Equity Portfolio is invested in longposition, otherwise it sells short the whole portfolio. As with the Equity Bond model represents theestimated return of the Stock Portfolio made at time t-1. This forecast can be provided, in order to implementthe trading rules, both from ANN models and Random walks or Buy and hold strategies.The testing period runs from January 2001 to June 2003 for a total of 10 quarterly of out of sampleobservations. In the trading experiment, it is assumed that, during the initiation period, an investor will invest$1 at the beginning of each month in either risk free bonds or the stock index fund depending on his choseninvestment strategy.Table 3 shows the results of trading patterns not only by analyzing the average return on an annual basis(Annualized Mean) and standard deviation on an annual basis (Standard Dev. Annualized) but also a Riskadjusted performance measure like Sharpe Ratio. This indicator allows a comparison of investment strategieswith diversified risk and return.Almost all trading models have driven by the estimates of neural networks have been a better result whencompared to the random walk models, not only with respect to the average return but also for the SharpeRatio, for all time horizons to estimate analyzed. In particular, strategies that use the best models of neuralnetworks and a quarterly time horizon have recorded very high Sharpe Ratio when compared to the resultsachieved by the buy and hold portfolio that is composed of 50% from the indexs Standard & Poors 500 andthe remaining 50% from the index BARCLAYS TREASURY U.S. 3-5Y. The strategies employing neural10 BARCLAYS US TREASURY 3-5Y has been taken into account.
12.
networks with a yearly time horizon of estimate showed a higher Sharpe Ratio compared to the Standards &Poors 500 Index.The high performance obtained from the quarterly data models are mainly due to their ability to predict thesize and sign of future performance (one step ahead). Furthermore, quarterly data models since they have ashorter time horizon with respect to models feed with yearly data, they show for a same out of sample period,a greater number of estimates. It appears that, in almost equal accuracy in return forecasting, the greaternumber of estimates made during the period 2000-2003, by quarterly data models, make true the opportunityto capture a greater number of upward and downward movements of the return portfolio, and in this waymake it possible a remarkable performance.Table 3 Trading Model Results - Quarterly horizon. Horizont Period Type Model Mean Annualizzed Standard Dev. Annualizzed Sharpe Ratio quarterly long/short Arch A Lagged+At +Recc 39.67% 22.57% 158.06% quarterly equity/bond Arch A Lagged+At +Recc 25.44% 23.81% 90.03% quarterly equity/bond Arch B Lagged+At+recc+forecast 25.44% 23.81% 90.03% quarterly long/short Arch B Lagged+At+recc+forecast 19.20% 29.54% 51.46% yearly equity/bond Lagged+rec+forecast 3.77% 0.58% 26.21% yearly long/short Lagged+rec+forecast 5.16% 2.54% 8.12% yearly equity/bond Lagged+rec+forecast -1.80% 7.30% -0.99% yearly equity/bond Random Walk (3-5 years) -1.80% 7.30% -0.99% yearly long/short Lagged+rec+forecast -1.80% 7.30% -0.99% yearly long/short Random Walk (3-5 years) -1.80% 7.30% -0.99% quarterly equity/bond Random Walk (3-5 years) -6.31% 25.40% -40.57% quarterly long/short Random Walk (3-5 years) -17.54% 43.61% -49.39% S&P 500 Index -18.39% 14.51% -5.07% Buy and Hold -8.73% 4.96% -7.04%5. CONCLUSIONFinance is a promising area for applying the ANN models to forecasting prices, returns, and indices.Generally speaking the success of the ANN models depends to a great extent on the selection of explanatoryinput variables which have a structural and corresponding with the output variable. In the present study,attempts have been made to test the contribution of information related to financial analysts like earnings’forecasts and recommendations and the effectiveness of different ANN architectures. It has also taken intoaccount of the forecasting horizon. To pursue this goal the study investigated the effectiveness of variousneural network models in prediction of stock returns in the case of a stock portfolio of sp500 and nasdaq100indexes shares. For the purpose, quarterly data, have been obtained from January 1997 to June 2003, and theneural networks are trained with varying sets of input data. Once the training of the neural networks getsover, the network has been used to predict the portfolio stock returns for one quarterly ahead. Theperformance of the various nonlinear models and the linear model were measured statistically and financiallyvia a trading experiment. On the whole the results are quite impressive, in fact the findings of our studysupport, to a great extent, the effectiveness of the neural network models in stock portfolio return forecasting,when as data input is also considered the contribution of financial analysts, namely earnings’ forecasts andstock recommendations. Furthermore between various ANN schemes used, the Recurrent MLP model it hasbeen showed as particularly useful in predicting the portfolio returns. More the trading experiment showsthat the ANN-guided trading strategies, and with particular reference to the quarterly horizon forecasting,obtain higher profits than the other investment strategies namely B&H and random walk trading strategy.
13.
REFERENCESM .Adya, F. Collopy (1998), “How Effective are Neural Networks at Forecasting and Prediction? A Reviewand Evaluation”, Journal of Forecasting, Vol. 17, Issue 5-6, pp.481-495.A. Ang, G. Bekaert (2007), “Stock return predictability: is it there?”, Review of Financial Studies, 20:pp.651-707.G.B. Antonio, O.U. Claudio, M.S. Manuel, O.M. Nelson, Y.H. Li, (1996), “Stock market indices in Santiagode Chile: forecasting using neural networks＂IEEE International Conference on Neural Networks, Vol.4,1996, pp.2172-2175.E. Avci (2007), “Forecasting Daily and Seasonal Returns of the ISE-100 Index with Neural NetworkModels”, Dogus University Journal, Vol. 8, No. 2, pp. 128-142.R. Banz, .W.Breen (1986), “Sample-dependent results using accounting and market data: some evidence”,Journal of Finance, 41, pp. 779–793.S. D. Bekiros., D. A. Georgoutsos (2008), “Direction-of-change forecasting using a volatility based recurrentneural network”, Journal of Forecasting ,27: 407-417.P. O’Brien, Y.Tian, (2006), “Financial analysts’ role in the 1996-2000 internet bubble”, Working Paper,University of Waterloo.J. Y. Campbell, R. J. Shiller (1988), “The dividend-price ratio and expectations of future dividends anddiscount factors”, Review of Financial Studies, 1: pp.195-228.J. Y. Campbell, R. J. Shiller (1988), “Stock prices, earnings, and expected dividends”, Journal of Finance,43: pp.661-676.J. Y. Campbell, M. Yogo (2006), “Efficient tests of stock return predictability”, Journal of FinancialEconomic, 81: pp.27-60.S S. Debasish, B. Das (2008), “Forecasting Movement of Stock Index Use Of Spread between E/P Ratio andInterest Rate”, Asian Social Science, vol.4 N. 1, pp. 67-78.A. I. Diler (2003), “ISE National-100 Index of the Direction of the Neural Network Errors EstimationMethod with Propagation Backward”, ISE Review, Vol.-25-26, pp. 65-81.E. Fama (1970), “Efficient Capital Market: A review of Theory and Empirical” Work. Journal of Finance,Volume 25, page 383-417.E .Fama, K. French (1992), “The cross-section of expected stock returns”, Journal of Finance, 47, pp.427–65.F.R. Fernando, G.M. Christian, S.R. Simon (2000), “On the profitability of technical trading rules based onartificial neural networks: Evidence from the Madrid stock market”, Economics Letters, Vol.69, No.1, pp.89-94.Al-Hindi, and Zayed F. Al-Hasan (2002), “Forecasting Stock Returns with the Neural Network Models”,Journal of King Saud University, Vol. 14, No. 1, pp. 65-81.W. E. Ferson, C. R Harvey (1993), “The risk and predictability of international equity returns”, Review ofFinancial Studies, 6, pp.527–66.Gencay, Ramazan (1998), “Optimization of Technical Trading Strategies and the Profitability in the StockMarkets”, Economic Letters, Vol. 59, No. 2, pp. 249-254.A. Goyal, I. Welch (2003), “Predicting the equity premium with dividend ratios”, Management Science, 49:pp.639-654.A. Goyal, I. Welch, (2008), “A comprehensive look at the empirical performance of equity premiumprediction”, Review of Financial Studies, 21: pp.1455-1508.A.A.B. Hammad, , S.M.A. Ali, E.L. Hall, (2009),”Forecasting the Jordanian Stock Price using ArtificialNeural Network”.C. R. Harvey (1995), “Predictable risk and returns in emerging markets”, The Review of Financial Studies,8, pp. 773–816.J.Jaffe, D. Keim, R.Westerfield (1989), “Earnings yields, market values and stock returns”, Journal ofFinance, 44, pp. 135–48.S.H. Kim, S.H. Chun (1998), “Graded Forecasting Using Array of Bipolar Predictions: Application ofProbabilistic Neural Networks to a Stock Market Index”, International Journal of Forecasting, Vol. 14, No. 3,pp. 323-337.K. J. Kim, I. Han (2000), “Genetic algorithms approach to feature discretization in artificial neural networksfor the prediction of stock price index”, Expert System Application, 19 (2), pp. 125–132.
14.
K. J. Kim, (2003), “Financial time series forecasting using support vector machines”, Neurocomputing, 55,pp.307 – 319K. Kim (2006), “Artificial Neural Networks with Evolutionary Instance Selection for Financial Forecasting”,Expert System Applications, Vol.30, No. 3, pp.519-526.S. Kothari, J. Shanken (1997), “Book-to-market, dividend yield, and expected market returns: A time seriesanalysis”, Journal of Financial Economics, 44: pp.169-203.H. Ince, T. B. Trafalis (2007), “Kernel principal component analysis and support vector machines for stockprice prediction”, IEE Transactions 39: 629-637.M. Lam, (2004), “Neural Network Techniques for Financial Performance Prediction: IntegratingFundamental and Technical Analysis”, Decision Support Systems, Vol.37, No.4, pp.567-581.O. Lamont (1998), “Earnings and expected returns”, Journal of Finance, 53: pp.1563-1587. R. Stambaugh(1999), “Predicitve regressions”, Journal of Financial Economics, 54: pp.375-421.J. Lewellen (2004), “Predicting returns with Financial ratios”, Journal of Financial Economic, 74: pp.209-235.R.Lawrence (1997), “Using Neural Network to Forecast Stock Market Prices”, University of Manitobaworking paper.G.C. Lim, P.D. McNelis (1998), “The Effect of the Nikkei and the S&P on the All-Ordinaries: AComparision of Three Models”, International Journal of Finance and Economics, Vol.3, No.3, pp.217-228.Q.Liu, F Song. (2001), “The rise and fall of Internet stocks: should financial analysts be blamed?”, Workingpaper, School of Economics, University of Hong Kong.T.H. Ma, (2003), “The Application of Neuron－Fuzzy to Emulate the Investment in TAIEX”, Mastersthesis, department of finance, Chaoyang University of Technology.B. Pant, , K. S. S. Rao, (2003), “Forecasting Daily Returns of Stock Index: An Application of ArtificialNeural Network”, The ICFAI Journal of Applied Finance, Vol. 9(2).K. Manish, M. Thenmozhi (2003), “Forecasting Daily Returns of Exchange Rates Using Artificial NeuralNetwork and ARIMA Model”, ICFAI Journal of Applied Finance.K. Manish, M. Thenmozhi (2005), “A Comparison Study of Selected Training Algorithms for NeuralNetworks to Financial Time Series Prediction”, Proceedings of 7th International Conference on Operationsand Quantitative Management.K. Manish, M. Thenmozhin (2004), “Forecasting Nifty Index Futures Returns using Neural Network andARIMA Models”, ACTA Press, ISBN (CD): 0-88986-417-9.J. K. Mantri, P.Gahan, B.B.Nayak (2010), “Artificial neural networks an application to stock marketvolatility”, International Journal of Engineering Science and Technology, Vol. 2(5), pp. 1451-1460.Mark T. Leung, An-Sing Chen Hazem Daouk (1998) ,”Application of neural networks to an emergingfinancial market: forecasting and trading the Taiwan stock index”, Forthcoming: Computers & OperationsResearch.J. Pontiiff, L. D. Schall (1998), “Book-to-market ratios as predictors of market returns”, Journal of FinancialEconomics, 43: pp.141-160.V Renu., C. Abhijeet (2010), “Predicting Stock Returns in Nifty Index: An Application of Artificial NeuralNetwork”, International Research Journal of Finance and Economics, Issue 49 (2010), pp.15-23.P. Roy, A. Roy (2008), “Forecasting Daily Returns of Nifty Index – Using the Method of Artificial NeuralNetwork”, FFMI 2008 Conference Proceedings, pp. 787-800.J.V Rodriguez,. et al. (2005), “STAR and ANN Models: Forecasting Performance on Spanish Ibex-35 StockIndex”, Journal of Empirical Finance, Vol.12, No.3, pp.490-509.R. Stambaugh, (1999), “Predictive regressions”, Journal of Financial Economics, Vol.54, pp. 375 421.Y.C. Tsai, T.M. Chen, T.Y Yang, C.Y. Wang (1999), “Neural network used in the investment strategy ofstock research”, Web Journal of Chinese Management Review, Vol.2, No.5, pp.25-48.R. Valkanov (2003), “Long-horizon regressions: theoretical results and applications”, Journal of FinancialEconomics, 68: pp.201-232.H. White, (1988) ,"Economic Prediction Using Neural Networks: The Case of IBM Daily Stock Returns",Proceedings of the Second Annual IEEE Conference on Neural Networks, II: 451-458.S.H. Wu, (2004), “Applying Technical Analysis of Stock Trends to Trading Strategy of Dynamic PortfolioInsurance”, Masters thesis, Department Of Department of Computer Science & Information Management,National Chiao Tung University.Y.C. Yen (1999), “Application of neural network models with the implementation”, Taipei: Scholars Books.
15.
Proceedings of Business And Information http://bai-conference.org Volume 8, Issue 1, 2011 ISSN: 1729-9322 Published by Academy of Taiwan Information Systems Research ContentsOrganizers’ Message .................................................................................................................. iSummary of Schedule ............................................................................................................... iiBusiness and Information 2011 Agenda ....................................................................................1International Symposium on Finance and Accounting 2011 ...................................................67Recipient of the BAI2011 Contribution Award........................................................................73Recipient of the BAI2011 Best Paper Award ...........................................................................76BAI2011 Officers and Organizing Committees .......................................................................82Hotel Transportation Information ............................................................................................87Layout of Session .....................................................................................................................88Guide to Presenters and Session Chair ....................................................................................90Author/Chair Schedule Index...................................................................................................91
16.
Business and Information 2011(Bangkok, July 4-6)Session [I5] 15:20 – 16:30 Landmark 5 (7F)Accounting and FinanceSession Chair: Fernando Zanella United Arab Emirates UniversityAlternative Neural Nets Approaches for Enhancing Stock Picking using Earnings ForecastsGiuseppe Galloppo University of Rome Tor VergataMauro Aliano University of Rome Tor VergataThe Technical Efficiency of Savings Groups in Hill Tribe CommunityKrisda Bhackdee Maejo UniversityAree Cheamuangphan Maejo UniversityRoengchai Tansuchat Maejo UniversityMontri Singhavara Maejo UniversityChanita Panmanee Maejo UniversityDoes Quality Certification Pays-Off? A Country Case StudyFernando Zanella United Arab Emirates UniversityThe Adjustment of Capital Structure across the Shifts in Macroeconomic Conditions: Evidencefrom the Electronics Industry of TaiwanHsien-Hung H. Yeh National Pingtung University of Science and TechnologyIndividual versus Team Measures of Auditor Expertise and Effects on RestatementKuei-Fu Li Ming Chuan UniversityYun-Shan Chen National Taiwan University 48
17.
AUTHER / CHAIR SCHEDULE INDEXA Bahari, Ahamad Zaidi [G5]Abasiz, Tezcan [H4] Bahari, Ahamad Zaidi [H6]Abd Rahman, Azmawani [G4] Bang, Yong Tae [H2]Abd. Rahman, Hardayanna [B1] Barajas Figueroa, Marco Antonio [I3]Abdullah, Mohammad Nayeem [E6 ] Barrento, Manuel Pedro [B2]Abdullah, Nor Liza [C1] Barroero, Thiago [D5]Abu-Jarad, Ismael [G2] Bau, Cho Tscan [G1]Abu-Shanab, Emad [C3] Beheshti Zavareh, Farnaz [F2]Acosta Valenzuela, Aurora Guadalupe [ I 3 ] Beneke, Justin [B6]Adenso-Díaz, Belarmino [B5] Berkman, Henk [D1]Agrawal, Kalyan Prasad [I2] Bhackdee, Krisda [I5]Agrawal, Kalyan Prasad [J4] Bin Hamidi, Muhammad Faisal [E1 ]Agus, Arawati [B3] Bin Ismail, Mohammad Shariff [E1 ]Ahmed, Roohi [D4] Bora, Chaytanya [F1]Akinci, Serkan [F4] Brangier, Eric [D2]Aksoy, Safak [F4] Brewer, Paul [H4]Albayrak, Tahir [F2]Albayrak, Tahir [F4] CAli, Noor Azman [G5] Caber, Meltem [F4]Aliano, Mauro [I5] Caber, Meltem [H2]Alm, Håkan [I2] Cao, Ting-Yun [F5]Aloufi, Khalid [C3] Castillo, Jose L [D5]Anantadjaya, Samuel PD [E3] Chalermkanjana, Kotchakorn [H1]Ang, David S [G4] Chan, Alvin M [A2]Anugerah, Rita [F6] Chan, Chung-Han [PO]Apibunyopas, Preeyanuch [C4] Chandra, Dicky [D2]Arshad, Rasidah [E4 ] Chang, Ann-Chen [J3]Asai, Tatsuo [D5] Chang, Bae-Muu [I2]Asgarkhani, Mehdi [G1] Chang, Cheng-Ying [J4]Asgarkhani, Mehdi [I2] Chang, Chia-Hsiang [F1]Ashourian, Mohsen [F2] Chang, Chia-I [PO]Asvial, Muhamad [D2] Chang, Chiao-Feng [C2]Athimethphat, Maythapolnun [J3] Chang, Chia-Yuan [E5 ]Awang, Abd Hair [B3] Chang, Chien-Wei [A5]Azmin, Adi Anuar Bin [E1 ] Chang, Chih-Hsiang [PO] Chang, Chih-Hung [J5]B Chang, Chin-Chih [PO]Bahari, Ahamad Zaidi [F2] Chang, Ching-Ter [PO]
Be the first to comment