The Effect of Stock Characteristics on Technical Analysis Profitability: the Portuguese Stock Exchange case

The Effect of Stock Characteristics on Technical Analysis Profitability: the Portuguese Stock Exchange case By Ricardo Gonçalves Fernandes Dissertation Master in Finance Supervisor: Mestre Ricardo Miguel Araújo Cardoso Valente 2015 ii Biographical Note Ricardo Gonçalves Fernandes was born on June 23, 1991, in Aveiro, Portugal. He completed the bachelor degree in Economics by Faculdade de Economia da Universade do Porto in 2013, joining the Master in Finance the same year. Ricardo is currently working for EY - Ernst & Young, on the transaction advisory services, where he is specializing on mergers and acquisitions consultancy. iii Acknowledgments Concluding this dissertation truly represents a milestone in my life. First of all, I would like to deeply thank my supervisor, Mestre Ricardo Valente, for all the support and advices given throughout this dissertation. Without his help and assistance this work wouldn´t be possible. Also a special thankful word to statistics teacher Luis Delfim Santos, that helped me with the statistical part of this dissertation. Secondly, I would like to express my gratitude to my family, my parents, my sister and grandparents, that made the person I am today. Everything I have I owe to them, my success is their success also, and all the words won’t be enough to describe how thankful I am. Lastly, a huge thanks to all my friends. It’s not necessary to name names, all that accompanied me throughout my academic life, living the same adventures as me, sharing the same happiness, the same smiles, the same tears, and that make the years up to this moment the best years I could have asked for. We will have some great stories to tell, one day. Thank you all! iv Abstract Many of the existing literature and recent-made enquiries state that technical analysis, whose goal is to predict the future evolution of asset prices from the observation and the use of their historical prices and trading volumes, is a growing investing technique used by many on the investors/traders community, being in the middle of a debate that attracted the attention and interest of both investors and academics for many decades. This study aims to contribute to the ongoing discussion, making an investigation about technical analysis trading rules profitability in the Portuguese Stock Market. Our principal aim is studying if certain stock characteristics (like size, liquidity, volatility and the stock´s industry) have a relevant influence on its results. Our time frame is between 2003 and 2013 and we’ve considered 38 stocks present in the Portuguese stock market and the PSI 20 index itself. We’ve found that technical analysis based investing models are not always statistically profitable, but, when they are, they produce substantial profits, even after transaction costs. There is some evidence that they are more profitable for less liquid stocks and more volatile stocks. There is no evidence to any industry or size bias in our study. Our conclusions can have enormous practical implications for investors and traders as they can decide to which extend it is advisable, or if it is advisable at all, to rely on technical analysis when making their investment decisions, and in what type of stocks to use it. Similar studies have been done but never for the Portuguese Stock Exchange, nor the selected time frame. Key-words: Financial Markets, Stock Markets, Technical Analysis, Portuguese Stock Market, Stock Characteristics. JEL-Codes: G10; G11; G14; G15; G23 v Resumo A maioria da literatura existente sobre este tema aponta a análise técnica, cujo principal objectivo é a previsão da evolução futura do preço dos activos, através da observação dos seus preços históricos e volumes, é uma técnica de investimento crescentemente usada pelos investidores/traders, estando no meio de um debate que atrai a atenção quer de investidores, quer de académicos por muitas décadas. Este estudo visa contribuir para a corrente discussão, investigando a possibilidade de modelos baseados em regras de análise técnica se verificarem lucrativos no mercado accionista português, com o principal objectivo de averiguar se certas características das acções, como o seu tamanho, liquidez, volatilidade e a indústria em que estão inseridas, influenciam de alguma forma os seus resultados. Estudamos tal questão em 38 acções transaccionadas em Portugal, bem como no índice PSI 20, no tempo compreendido entre 2003 e 2013. Concluímos que modelos de investimento baseados em análise técnica nem sempre demonstram ser estatisticamente lucrativos, mas, quando assim o é, produzem lucros substanciais. Existe evidência de melhores resultados em acções ilíquidas e mais voláteis, não existindo ligação entre os mesmos resultados e a indústria em que a acção se insere e o seu tamanho. As conclusões aqui apresentadas poderão ter enormes implicações para quem pretenda investir no mercado bolsista, visto que permitem decidir se é, e até que ponto é, compensatório seguir modelos de investimento com base em regras de análise técnica, e em que tipo de acções os utilizar. Estudos similares foram realizados, mas nunca para o mercado accionista português, nem para o período temporal considerado. Palavras-chave: Mercados Financeiros, Mercados Accionistas, Análise Técnica, Mercado Bolsista Português, Características de Acções; Códigos JEL: G10; G11; G14; G15; G23 vi Index 1. Introduction ............................................................................................................. 1 2. Literature Review .................................................................................................... 4 2.1 Birth and Origins of Technical Analysis .................................................................... 4 2.2 Efficient Market Hypothesis and Technical Analysis ................................................ 5 2.3 Technical Analysis Profitability ................................................................................. 6 2.4 Behavioral Finance and Technical Analysis ............................................................... 8 2.5 The Portuguese Stock Market and its Efficiency ........................................................ 9 2.6 Similar studies and their conclusions ........................................................................ 10 2.7 Critical analysis of the reviewed literature ............................................................... 12 3. Methodological Aspects ......................................................................................... 13 3.1 Data .......................................................................................................................... 13 3.2 Technical trading rules/model ................................................................................. 15 3.3 Stock characteristics description ............................................................................. 23 3.4 Statistical approach .................................................................................................. 26 4. Empirical Results ................................................................................................... 30 4.1 Investor assumptions ................................................................................................. 30 4.2 Sample descriptive analysis ...................................................................................... 31 4.3 Models application results ........................................................................................ 32 4.4 Transaction costs effect ............................................................................................ 40 4.5 Characteristics influence tests ................................................................................... 45 5. Conclusions and future research suggestions ..................................................... 54 References ...................................................................................................................... 56 Appendix ........................................................................................................................ 61 vii Table Index Table 1 - Stocks that constitute our sample .................................................................... 14 Table 2 - Daily returns are presented for the full PSI 20 sample .................................... 31 Table 3 – Results of the models application: complete sample, 2003-2013 ................... 33 Table 4 - Results of the models application: 1st sample, 2003-July 17 2007 ................. 34 Table 5 - Results of the models application: 2nd sample, July 18 2007-2013 ............... 35 Table 6 - Results of the models application, not having in consideration transaction costs. ............................................................................................................................... 38 Table 7 - Results of the models application, having in consideration the transaction costs: full sample, 2003 – 2013 ...................................................................................... 40 Table 8 - Results of the models application, having in consideration the transaction costs: 1st sample, 2003 – July 17 2007 ........................................................................... 41 Table 9 - Results of the models application, having in consideration the transaction costs: 2nd sample, July 18 2007 – 2013 ......................................................................... 41 Table 10 - Results of the models application, having in consideration transaction costs, for the full samples ......................................................................................................... 44 Table 11 - Biggest and smallest stocks, in terms of Market Cap., where Model A revealed to be profitable. ................................................................................................ 47 Table 12 - Results for the size influence test on Model A. ............................................. 47 Table 13 - Biggest and smallest stocks, in terms of Market Cap., where Model B revealed to be profitable. ................................................................................................ 47 Table 14 - Results for the size influence test on Model B. ............................................. 48 Table 15 - Most liquid, and less liquid stocks, in terms of Turnover by Volume, where Model A revealed to be profitable. ................................................................................. 48 Table 16 - Results for the liquidity influence test on Model A. ..................................... 49 Table 17 - Most liquid and less liquid stocks, in terms of Turnover by Volume, where Model B revealed to be profitable. ................................................................................. 49 Table 18 - Results for the liquidity influence test on Model B. ...................................... 49 Table 19 - Most volatile and less volatile stocks, measured by the Standard Deviation, where Model A revealed to be profitable. ...................................................................... 50 Table 20 - Results for the liquidity influence test on Model A. ..................................... 50 Table 21 - Most volatile and less volatile stocks, measured by the Standard Deviation, where Model B revealed to be profitable. ....................................................................... 51 Table 22 - Results for the liquidity influence test on Model B. ...................................... 51 Table 23 - Differences in statistical significance profitability by industry ..................... 53 1 1. Introduction Technical analysis can be seen as “an art”, according to Pring (2002), with the main objective of “identify a trend reversal at a relatively early stage and ride on that trend until the weight of the evidence shows or proves that the trend has reversed”. In fact, technical analysis is a forecasting method of price movements using past prices, volume and, in case of future markets, open interest (Park and Irwin, 2004). Many previous studies show the relevance of this analysis in the real world of investing decisions, with Bray (2002) stating that 35% of all the volume on the New York Stock Exchange (NYSE) is due to technical trading rules, and Talley (2002) suggesting that the doubts that investors have about the integrity of the accounting data that generates fundamental analysis indicators are leading technical analysis to become even more popular. With so many interest showed by investors and traders, the academic community soon started studying technical analysis as a whole investment technique and stock picking tool, with a strong focus on whether it has proven to be a profitable investment strategy. Brock et al. (1992), Bessembinder and Chan (1998) or Coutts and Chen (2000) are all examples of studies that investigate technical analysis profitability in indexes like the Dow-Jones Industrial Average (DIJA), or the Hang Seng Index (HSI). Park and Irwin (2004) made an extensive review on the evidence of profitability of technical analysis, categorizing the analyzed studies in “early” and “modern” groups, according to the characteristics of testing procedures. However, only Marshall et al. (2009) studied the effect that certain stock characteristics, specifically size, liquidity or industry, have on the possible profits gave by the application of technical trading rules, focusing on the US stock market. Our study aims to analyze, not only the technical analysis profitability, but mainly the effect that certain stock characteristics, namely the size (measured by market capitalization), liquidity (measured by the turnover), volatility (measured by the stock standard deviation), and the industry where the stock is inserted, have on technical analysis profits, between the years 2003 and 2013, for the stocks quoted on the Portuguese Stock Exchange. The time frame will be divided into two groups (2003July 2007 and July 2007-2013) in order to see if the fact that the market is in its bull or bear state has any influence in possible profits generated by technical analysis trading rules. 2 In order to reach this goal, we define a sample of the vast majority of stocks listed in the Portuguese Stock Exchange (PSI Geral index), and construct two investing models that includes three main technical rules/indicators each, in order to obtain a full insight on technical analysis profitability, rather than individually testing one simple trading rule. We then perform the necessary statistical tests to see if the chosen technical trading models produced profitable results in the given time interval. Then, we will divide the stocks of our sample into sub-samples according to the chosen characteristics, and test if those same characteristics have any impact in the models profitability. We also compare the model´s results with other strategies, like a “buy-and-hold” one, and test if it produced positive results on the PSI 20 index (the most relevant indicator of the Portuguese stock market) itself. Besides this section, this dissertation proposal is structured as follows: in the next section, section 2, we will present a literature review of the topic. In section 2.1 we will present the birth and origins of technical analysis, in section 2.2 the efficient market hypothesis and its relation with technical analysis, in section 2.3 the studies made on technical analysis profitability. Section 2.4 addresses the relation between technical analysis and behavioral finance, section 2.5 gives a quick overlook on the Portuguese Stock Market and its efficiency, section 2.6 makes a review of the similar studies and their conclusions, and lastly, in section 2.7, a final critical analysis of the reviewed literature will be discussed. In section 3, we will disclose the methodological aspects of this study, specifying in section 3.1 the data that will be used, in section 3.2 the technical trading rules that we will analyze/test and that will constitute our model, in section 3.3 the chosen stock characteristics description and in section 3.4 the statistical tests/approach that we will use. Section 4 will have the empirical results. First, in section 4.1, our hypothetical investor assumptions will be revealed, then, on section 4.2, we will we will present a descriptive analysis of the sample in which we will work upon, followed by section 4.3, in which we will present our models application results, and by section 4.4, which will contain the tests to verify the existence of the transaction costs significance. Lastly, on section 4.5, we will present the tests and assumptions that will lead to our conclusions about the 9 2.5 The Portuguese Stock Market and its Efficiency The Portuguese Stock Exchange had its origins on 1 of January 1769, when the Lisbon Stock Exchange was created. Almost one century later, the Porto Stock Exchanged was born. Both exchanges merger in 1999, becoming the Bolsa de Valores de Lisboa e do Porto, and join the Euronext family in 2007, being now known as the Euronext Lisbon Stock Exchange. Its benchmark index is the PSI 20. The PSI 20 is a free float market capitalization weighted index that reflects the performance of the 20 (18 nowadays) most actively traded shares listed on Euronext Lisbon, and is the most widely used indicator of the Portuguese stock market. It serves also as an underlying for structured products, funds, exchange traded funds and futures. It was introduced on 31 December 1992, and is constituted by stocks from every sector, like utilities, oil and gas, retail or banks. Despite being the main one, other indices exist in the Portuguese Stock Exchange besides the PSI 20. The PSI 20 NR (with NR standing for Net Return, this index is issued since the end of 2010, and is obtained by reinvesting the net dividend, which is equal to the ordinary gross dividend minus the amount of withholding tax), the PSI 20 GR (the GR stands for Gross Return, with this index being published since 1992 and being obtained by reinvesting in the index the ordinary gross dividends declared by the index constituents) and the PSI Geral (also called PSI All-Share) are examples of that. Sectorial indices like the PSI Utilities, PSI Technology or the PSI Financial, which only comprise stocks of the respective sectors, are also available. The Psi Geral index will serve as base for this study. It is a market capitalization weighted price index of the eligible companies listed on the Euronext Lisbon. It includes shares issued by all the companies which are listed on Euronext Lisbon. On the last day of 2013, 58 companies were quoted on that stock exchange. Afonso and Teixeira (1999) were the first to test the Fama´s weak-form of efficiency on the Portuguese stock market, from the time frame between 1990 and 1997. They concluded that the possibility of having abnormal profits in the stock market existed, putting in question the form of efficiency tested. A few years later, Nascimento (2007) tested if the random walk model could be applied to the Portuguese stock market, with data from 1997 to 2007. The study produced mixed results, as the prices followed a random walk (which corroborates the market efficiency hypothesis) in a selected time 10 period, nor following it in the remaining time, which suggest some predictability in prices variation. Recently, Borges (2010) tested the weak-form market efficiency applied to stock market indexes of France, Germany, UK, Greece, Portugal and Spain, from January 1993 to December 2007, using a serial correlation test, a runs test, an augmented Dickey-Fuller test and the multiple variance ratio test, for the hypothesis that the stock market index follow a random walk. The author concluded that the studied hypothesis is rejected for Greece and Portugal, due to serial positive correlation. However, the empirical tests show that these two countries have also been approaching a random walk behavior after 2003, which can indicate an increase in efficiency levels. Based on that work, Vaz (2012) also tested the Fama´s weak-form of market efficiency, applied to the same six European market indexes from January 2007 to January 2012. The author designed a strategy based on the k-NN and on the Neural Network predictions, and concluded that it would be possible to obtain significantly higher earnings by the implementation of this strategy than using a simple buy-and-hold strategy. Being so, this shows that the PSI-20 is not yet totally efficient. Our study will also contribute to this ongoing debate. 2.6 Similar studies and their conclusions As we have seen in section 2.3, many studies exist on technical analysis profitability (with an even bigger number that is not here referenced), focusing on both Portuguese and international stock markets. However, as mentioned in the introduction, the main research question of this dissertation is a broader one. It aims to analyze, not only the technical analysis profitability, but mainly the effect that certain stock characteristics, such as the size, liquidity and volatility, have on technical analysis profits, for stocks quoted on the Portuguese stock exchange. The only study made with such specifications was Marshall et al. (2009). In it, the authors studied the effect that certain stock characteristics, like size, liquidity or industry where the stock is inserted, have on the possible profits gave by the application of technical trading rules, focusing on US stock markets, namely the New York Stock Exchange (NYSE) and the National Association of Securities Dealers Automated Quotations (NASDAQ), from the years of 1990 to 2004. They picked stocks listed both at the beginning and the end of the time period in order to keep the methodology manageable, with a maximum yearly nontrade 11 percentage of less than 5% over all years. The final sample contained 1065 stocks, 866 quoted on the NYSE and 199 on NASDAQ. The technical trading rules tested were the Variable-length Moving Average (VMA), the Fixed-length Moving Average (FMA) and the Trading Range Break-out rule (TRB). For each type of rule the short term is set at 1 trading day, and the long-term at 50, 100, 150 and 200 trading days. After defining the sample and the rules to be tested, the authors applied a bootstrapping methodology, originated by Efron (1979), using the GARCH-M (Generalized Autoregressive Conditionally Hetroskedastic in the Mean) as the model for modeling the stock price process. Bootstrapping has found relevance in numerous statistical procedures and has become a widely accepted alternative to traditional statistical inference. It is used to address the possible problem of non-normality of returns. Testing statistical significance in this way can assure that the results found are robust. The results were run assuming investment at the closing prices on both the signal day and the following day. The study concluded that the number of statistically significant positive return companies is small for all trading rules over all periods for both markets. On average, the trading rules give positive results on only 2.8% of NYSE stocks and 3% of NASDAQ stocks. After this first conclusion, the authors study more in-depth the stocks that revealed to be profitable, generating t-statistics by comparing the size and liquidity of the stocks which generate positive significant returns to those which do not. The results, for all trading rules, indicate that smaller size and smaller volume stocks are more likely to generate positive statistically significant return than larger size and larger volume stocks. The analysis by industry showed that the number of significantly positive stocks was randomly spread across the different industries, not existing a link between a firm´s industry and the profitability of technical analysis. Lastly, the authors tested the stocks that revealed to be profitable for economic significance, concluding that, even after transaction costs, when technical trading rules do generate profits, the profits are substantial, largely beating a simple buy-and-hold strategy, which can be the reason explaining why practitioners continue to use these trading rules despite them not generating profits on a consistent basis. 12 2.7 Critical analysis of the reviewed literature Since its birth, by the hands of Charles Dow in the late 1800´s, technical analysis has always been looked at with disdain by many scholars and academics, and paradoxically has been used as a main investment technique by many traders and investors. Many have attacked it, using the Efficient Market Hypothesis argument or the random walk of stock prices as an explanation of why technical analysis based investment strategies cannot generate any substantial profits, like Bessembinder and Chan (1998). Numerous studies have been made that show, in fact, that technical trading rules do not produce above-market profits (several studies mentioned on section 2.3), and, in many cases, after transaction costs, are even beat by a simple buy-and-hold strategy. However, countless other studies have been made to prove the contrary, showing that technical analysis, in almost all stock exchanges and during all possible time periods, through the use of some technical rules, generate profits in scenarios where otherwise wouldn´t be possible. Portugal is not an exception to this type of studies, with technical indicators being studied and most of them revealed to be profitable, questioning also the efficiency of the Portuguese Stock Market. The heated debate that surrounds these questions, both the technical analysis profitability and the efficiency of the Portuguese Stock Exchange justifies, per si, the realization of this study. Only one study with the same research topic as the one we are conducting have been made, but focusing on the US stock market, making this topic and consequent underlying questions a research hole in Portugal, fact that arises the need for this new study. 13 3. Methodological Aspects This section intends to expose the data used in this study and the methodology adopted. Along it, we will present the data in which we will work on, the technical trading rules and the stock characteristics that will be tested/analyzed, as well as the methodological steps that we will follow in order to reach our final results, including all the assumptions and tests made. 3.1 Data Our study is centered in the Portuguese stock market – NYSE Euronext Lisbon, more specifically in the stocks that integrate the PSI Geral index. The sample period runs from January 1st 2003, the first trading day of that year, to December 31st 2013, exactly 11 years. We will use the data present at Thompson Reuters Datastream database, consisting of a time series of daily stock prices at closing time. Base on those prices, we will extract information like the daily returns of each stock. For a specific stock to be included in our sample, it must belong to the PSI Geral during the years that we will test, and be active at the end of our sample, which makes a total of 58 stocks. Of those 58 stocks, to keep the methodology manageable and have statistical relevant results, we will exclude those that have less than two years of trading, and those that do not have sufficient trading activity to be of interest to traders in general, i.e., stocks that are not traded on, at least, 50% of the days of our time interval. The final number of included stocks is 38, since there are several stocks in the Portuguese exchange that fail the trading activity criteria. The list of stocks to analyze will then be, presented in alphabetical order: Name Code Name Code ALTRI ALT(P) BANCO COMERCIAL PORTUGUES BCP(P) BANCO BPI BPI(P) CIMENTOS DE PORTUGAL CPR(P) CIPAN CIP(P) COFINA CFNA(P) COMPTA COM(P) CORTICEIRA AMORIM COR(P) EDP ENERGIAS EDP(P) F RAMADA RINT(P) 14 DE PORTUGAL INVESTIMENTOS FUTEBOL CLUBE DO PORTO FCP(P) GALP ENERGIA GES(P) GLINTT GLOBAL GLIN(P) IBERSOL IBE(P) IMPRESA IMPR(P) INAPA INA(P) JERONIMO MARTINS JMT(P) LISGRAFICA LIG(P) MARTIFER MAR(P) MOTA ENGIL EGL(VO) NOS NOS(P) NOVABASE NBA(P) PORTUCEL EMPRESA PTI(P) PORTUGAL TELECOM PTC(P) REDITUS RED(P) REN REN(P) SAG GEST GST(P) SDC INVESTIMENTOS SCO(P) SEMAPA SEM(P) SONAE CAPITAL SNC(P) SONAE COM SNCA(P) SONAE INDUSTRIA SOI(P) SONAE SON(P) SPORT LISBOA E BENFICA FUTEBOL SLB(P) SPORTING CLUBE DE PORTUGAL SCP(P) SUMOL+COMPAL SUCO(P) TEIXEIRA DUARTE TDSA(P) VISTA ALEGRE ATLANTIS VAFK(P) Table 1 - Stocks that constitute our sample and will be under analysis in our study. The code refers to the Thompson Reuters Datastream code. Each included stock can have a different number of observations, as not all started to be traded on the same day. However, the maximum number of observations that any individual stock can have is 2870. We will also test the profitability of our model in the PSI 20 index itself, so we will use the index values for the same time frame, in order to also get data like the daily returns. The use of the above described process to select our sample will minimize problems like the forward-looking bias i.e. information that is available ex-post will not be used, the delisting bias, the situation where a stock is getting delisted during a period under study emerges and the survivorship bias, a situation that arises when a portfolio is constructed on the basis of the companies that have survived the time period of the test 15 and are available for portfolio construction. We will also avoid troubles associated with data snooping, as this study does not search for trading rules proved to be profitable in the past but set the trading rules together with all the parameters in advance instead. We will also break down our time frame into two sub-periods: the first one between January 1st 2003 and July 17th 2007 (with each stock having a maximum of 1185 observations), and the second one between July 18th 2007 and December 31st 2013 (with each included stock having a maximum 1685 observations), in order to see if the fact that the market is in its bull or bear state had any influence in possible profits gave by technical analysis and on the purpose of our study. Bulls and Bears can be seen as two opposite market mentalities. Definitions of bull and bear markets can be different in small details, but all agree in the same assumptions. A market can be considered to be on a bull stage when it is in a sustained period of positive market returns, achieving new highs and being characterized as a period of wealth creation. On the other hand, a bear market is a sustained period of negative market returns resulting in a substantial decline of wealth. Cooper et al (2004) define the market state as being "up" (bull in our case) when the preceding 3-year return has been positive and as "down" (bear) if the reverse is true. We will use July 17th 2007 as the splitting day between the two sub-samples, as it was the day where the PSI 20 index reached a pick of 13702 base points, reaching its higher value since the begging of our time sample, putting an end to the existing bull market phase, and marking the begging of a bear market, which would last until the end of our time sample. 3.2 Technical trading rules/model Many trading rules have been tested before, as above mentioned. However, few studies have considered the impact of technical indicators when combined, i.e., many studies have studies if indicators like different length moving averages, MACD, the relative strength index, trading range breakout rules, etc, have proven to be, or not, profitable, but not many have done it using some of those rules together, to test their effect when combined. That is also a research hole for many stock exchanges across the world, as well for the Portuguese one. In this section we will describe the rules/indicators that will be considered in our model, and that will be tested. 16 3.2.1) Moving Averages Among the most used trading rules by researchers to test market efficiency, the one employed most frequently is the so-called moving average (henceforth MA) rule. Due to the precise signals it generates, this rule is used by many investors as a stand-alone method, particularly in the automated trend following trading systems, and, in contrast to other rules of technical analysis, is mathematically well defined, in Neftci (1991). It has many variations to it, as it is described in Pring (2002). In one of its versions, the two non-centered, moving averages with different length are initially created, a “short” moving average of order n, and “long” moving average of order m (with m>n), from the time series of stock prices or index levels, as defined as follows: MASt represents the relatively shorter MA with length m; MALt represents the relatively longer MA with length n, calculated at time t. In both moving averages Pt is the stock price at time t, θi are non-time varying parameters, and B is the backward shift operator, i.e. B  Pt = Pt-i. The MA rule generates a buy signal when MASt rises above MALt and when this happens, a trend is said to be initiated. On the opposite side, a sell sign is generated when MASt falls below the MALt. The most popular MA rule, as indicated in Brock et al. (1992), is the 1-200, which sets n=1 (so that MASt is just the current level of the stock price or index) and m=200. Although, lots of other rules can be constructed by modifying the values of n and m, such as the also popular 1-50, 1-150 or 2-200 rules. These rules are often modified by introducing a band around the MA, which reduces the number of buy and sell signals by eliminating “whiplash” signals that occur when MASt and MALt are close to each other. In our study, we will use the MA rules both with and without a one per cent band, i.e. a buy (sell) signal is generated when MASt is above 17 (below) MALt by more than one per cent. If MASt is inside the band, no signal is generated. This rule attempts to simulate a strategy where traders go long as MASt moves above MALt and short when it moves below: with a band of zero all days are classified as either buys or sells. Mathematically, buy and sell signals can be defined as following: a) Buy Signals b) Sell Signals In both cases, the initial times τᴮ0 and τˢ0 are set equal to zero and D is the so-called band, above described. All θi parameters will be set equal to one. For the propose of our study, we will include in our model two moving average rules. We will use both short and long “longer averages”. An argument for using shorter averages is that it is more in accordance with the time horizon for real investors. Investors usually measure their performance over a one year period and thus, try to maximize their profits in this period. A shorter moving average also captures the direction of the trend at an earlier stage. That being said, we will use the popular 1-200 rule (with no band), and a shorter “longer” moving average, the 1-50 rule (with an 1% band). 3.2.2) Stochastic This technical indicator was developed by George C. Lane in the late 1950´s. It is basically a momentum indicator that shows the location of the close relative to the highlow range over a predefined number of periods. It compares an asset closing price to its price range over a given time period. Its underlying theory is that in a bullish market the highs are located closer to close values, being the opposite for a bear market true as well, as if the asset price tends to decrease, the lowest values are located near the close 18 values. Hence, stochastic oscillator follows the speed or the momentum of prices instead of volume or prices themselves. Its main lead is the possibility of finding and identifying momentum changes direction before price ones. It can also be used to identify Bull and Bear divergences. A bullish divergence is formed when price records a lower low, but the Stochastic Oscillator forms a higher low. This shows less downside momentum that could foreshadow a bullish reversal. On the other hand, a bearish divergence is formed when price records a higher high, but the Stochastic Oscillator forms a lower high. This shows less upside momentum that could foreshadow a bearish reversal. The oscillator can be used in two ways: firstly, it is supposed to be useful for identifying overbought and oversold levels, and secondly, it can be used to identify bullish and bearish divergences to anticipate the change in price direction, as mentioned above. It consists of two lines which are expressed in the following formulas present on Lane (1984): Where, Close is current close value of the asset, Lowest Low is the lowest intraday value reached by the asset over the look-back period and Highest High is the highest intraday value reached by the asset over the look-back period. The other line exploited in the stochastic oscillator is a three day simple moving average (SMA) of the line above, line known as %D, or the signal line. The default setting for the Stochastic Oscillator is 14 periods, which can be days, weeks, months or an intraday timeframe, depending of the objective of the investor. A 14period %K would use the most recent close, the highest high over the last 14 periods and the lowest low over the last 14 periods. In our study, the 14 periods are used which translates into 14 days, and 3 days SMA of the %K for the %D line. There are three versions of the Stochastic Oscillator: the Fast Stochastic Oscillator, which is based on George Lane's original formulas for %K and %D, described above. However, %K in the fast version that appears rather choppy. In fact, Lane used %D to 25 (1994) conclude that volume plays a role beyond simply being a descriptive parameter of the trading process. In the Portuguese case, for instance, on the last year of our sample, companies like BPI, BCP or Portugal Telecom are very liquid, contrary to companies like Corticeira Amorim, Impresa or Novabase, which are in the less liquid bottom of the market. c) Volatility Volatility, in its simplest way, can be defined as the degree to which prices vary over a certain length of time. It is a statistical measure of the dispersion of returns for a given security (stocks in our case) or market index, and can either be measured by using the standard deviation (of changes in the log of asset prices), or variance between returns from that same security or market index. It is an important measure for quantifying risk: for example, a security with a volatility of 50% is considered having very high risk because it has the potential to increase or decrease up to half its value. Commonly, the higher the volatility, the higher its price variations, and so, the riskier the security. Other possible measure of stock relative volatility to the market is its beta. A beta is a measure of the volatility, or systematic risk, of a security in comparison to the market as a whole. It approximates the overall volatility of a security's returns against the returns of a relevant benchmark (normally, the market where the stock is present is used). For instance, a stock with a beta value of 1.2 has historically moved 120% for every 100% move in the benchmark, based on price level, as, on the other hand, a stock with a beta of 0.7 has historically moved 70% for every 100% move in the underlying index. For the purpose of our study, we will measure stock volatility based on the standard deviation of its historical daily prices. d) Industry In our industry analysis, we will use the same classification made by the Industry Classification Benchmark (ICB). The ICB is an industry classification taxonomy launched by Dow Jones and the FTSE in 2005 which divides the market into increasingly specific categories, allowing investors to compare industry trends between well-defined subsectors. It uses a system of 10 industries, partitioned into 19 26 supersectors, which are further divided into 41 sectors, which then contain 114 subsectors. In our study, we will insert our stocks into the 19 supersectors, which are: Oil & Gas, Chemicals, Basic Resources, Construction & Materials, Industrial Goods & Services, Automobiles & Parts, Food & Beverage, Personal & Household Goods, Health Care, Retail, Media, Travel & Leisure, Telecommunications, Utilities, Banks, Insurance, Real Estate, Financial Services, and Technology. We will also add an Other category, in case any of the stocks does not include itself in any of the above mentioned supersectors. After we conclude if the stocks are profitable or nor under technical analysis rules, we will try to find any existing link between a firm´s industry and the profitability of technical analysis trading rules, by seeing if the number of significantly positive stocks are randomly spread across different industries or not. 3.4 Statistical approach Our first step will be testing the profitability of the model, using conventional statistical tests, traditional t-tests. Our model provide us with buy and sell signals, above described, signals that we will use, as well as data like stock returns. In order to obtain a comprehensive answer to the question of technical analysis profitability, the following hypotheses are formulated and will be tested: 1) The returns are significantly larger/smaller than zero; For our model to be revealed profitable, we expect returns on buy periods to be bigger than zero and returns on sell periods to be less than zero. We will compare the mean daily returns earned during buy or sell periods with corresponding t-statistics. The procedure for testing is as stated below. 1.1) For buy days: 𝑯𝟎:𝑿𝟏=𝟎 𝑯𝟏:𝑿𝟏>0 27 With the corresponding t-statistics being the following: 𝒕= 𝑿𝟏 𝒔𝟏 √𝒏𝟏 Where, 𝑿𝟏 is the mean return on buy days; 𝒔𝟏 is the standard error estimated from the buy days sample; 𝒏𝟏 is the number of returns on buy days. 1.2) For sell days: 𝑯𝟎:𝑿𝟐=𝟎 𝑯𝟏:𝑿𝟐<0 With the corresponding t-statistics being the following: 𝒕= 𝑿𝟐 𝒔𝟐 √𝒏𝟐 Where, 𝑿𝟐 is the mean return on sell days; 𝒔𝟐 is the standard error estimated from the sell days sample; 𝒏𝟐 is the number of returns on sell days. 2) The returns on buy and sell days differ appreciably We will verify the previous results by comparing returns on buy and sell periods. In order for our model to be profitable, both returns should significantly differ. We will compare the mean daily returns earned during buy and sell periods with corresponding 28 t-statistics following two-tailed distribution. Under the null hypothesis, these returns will not differ. The procedure for testing is as stated below. 𝑯𝟎:𝑿𝟏−𝑿𝟐=𝟎 𝑯𝟏:𝑿𝟏−𝑿𝟐≠𝟎 With the corresponding t-statistics being the following, as used in Brock et al. (1992): 𝒕= 𝑿𝟏−𝑿𝟐 (√𝒔𝟏 𝟐 𝒏𝟏+𝒔𝟐 𝟐 𝒏𝟐) 3) The returns on buy days exceed returns on a buy and hold strategy Lastly, we will compare returns obtained using our model to a simple “buy-and-hold strategy”. We will assess whether the respective returns are equal to the unconditional mean daily returns with corresponding t-statistics. Under the null hypothesis, the returns earned using technical trading rules on buy days do not differ significantly from returns earned according to a “buy-and-hold” strategy. The procedure for testing is also stated below. 𝑯𝟎:𝑿𝟏−𝑿 =𝟎 𝑯𝟏:𝑿𝟏−𝑿 ≠𝟎 With the corresponding t-statistics being the following, also as used in Brock et al. (1992): 𝒕= 𝑿𝟏−𝑿  (√𝑺 𝑵+𝒔𝟏 𝟐 𝒏𝟏) Where, 𝑿  is the unconditional mean daily returns; 29 N is the number of observations; S is the standard error estimated from the entire sample. For every hypothesis tested, if we cannot reject the null hypotheses formulated in this section, we cannot find evidence of our model profitability. Considering the size of our sample, and some statistical tests below described, critical values corresponding to a normal distribution will be used instead of those corresponding to a student distribution. The critical values are 1.282, 1.645 and 2.326 (for levels of significance α 10 %, 5 % and 1 %, respectively) and their negative equivalents will be used when needed. For two tailed distribution related to the second hypothesis the values 1.645, 1.96, 2.576 (for levels of significance 10 %, 5 % and 1 %, respectively) will be used. Hence, for a certain α level of significance, if the resulting T is bigger than the critical value, we will reject the null hypothesis. After assessing the model profitability per si, and comparing its results with a “buy-andhold” strategy, we will compare the size, volatility and liquidity of the stocks which generate positive statistically significant returns to those which do not, in order to see if those characteristics have any effect on the results, which is our main research question. The procedure for these tests will be explained on section 4.5. 30 4. Empirical Results In this section, we will present the empirical results of the test statistics presented above. We will divide this section into 5 subsections, as described in the introduction. 4.1 Investor assumptions To proper conduct our study, we will need to establish some considerations about our hypothetical investor. Here are the main ones: a) Our investor will be relatively risk averse and will not incur on short-selling techniques. He or she will start a trade after our model produces a buy signal, and will keep the trade open until the model produces a sell sign. After that sign, the investor will liquidate its positions, to re-open them after the model’s next buy signal. When the investor isn´t on the market, it will simply keep its funds in liquidity/cash form; b) Our investor will not be an intraday investor, which means that the enters and the exists of the market will be made at the day’s closing price, when our model produces its signs; c) Since our study will be about the Portuguese Stock Exchange, we will use to open and close our investor’s trades a Portuguese based broker; d) On the last step of our study, we will compare the model´s results with a simple “Buy-and-Hold” strategy. In that strategy, we will assume that our investor buys the assets in the first day of our sample and liquidates its positions on the last one, making no changes in its portfolio, regardless of the price fluctuations in the market; e) At the end of our eleven year time sample, our investor will exit any open trade, liquidating its positions; f) Our investor won’t set any stop losses or take profit orders. Him or her will simply liquidate its positions or enter a new one once a signal appears; 31 g) Our investor won’t have a risk management and exposure strategy. Whenever the model tells to enter in a trade, him or her will open the position with the full amount of capital, divesting also the full amount when the model produces a sell signal; h) As mentioned above, we will perform our study with and without transaction costs, in order to assess its effects in our model. To compute the value to use, we analyzed the prices practiced by all brokers acting in Portugal, to get to an average cost per operation for every market order above 10.000€, using the internet as the only negotiation channel. These prices and calculations can be found on Appendix 1. We reached a value of 0,168%, value which we will use. 4.2 Sample descriptive analysis To start off this section, we will present an analysis of the daily returns of the PSI 20 index, with the main aspects presented in the table below. 2003-2013 2003 – July 17 2007 July 18 2007 - 2013 Number of days 2870 1185 1685 Mean 0.00413% (0.000041362) 0.07218% (0.000721885) -0.04372% (-0.000437226) Std. Deviation 1.162% (0.011627333) 0.628% (0.006282741) 1.421% (0.014213413) Skewness -0.146524803 -0.100478460 -0.049309604 Kurtosis 12.37 5.83 9.27 𝝆(𝟏) 0.058 0.057 0.055 𝝆(𝟐) -0.006 0.035 -0.015 𝝆(𝟑) -0.027 0.029 -0.038 𝝆(𝟒) 0.002 0.058 -0.009 𝝆(𝟓) -0.040 -0.038 -0.043 JB Normality Test 10504.66 398.3821 2759.124 Probability 0.00 0.00 0.00 Table 2 - Daily returns are presented for the full PSI 20 sample and for two nonoverlapping subsamples. Daily returns are measured as log differences of the level of the index: r_t=ln⁡(x_t/x_(t-1) ). ρ(i) is the 32 estimated autocorrelation at lag i of r_t. ‘JB Normality Test’ stands for the Jarque-Bera Test. Probability is the probability that a Jarque-Bera statistic exceeds (in absolute value) the observed value under the null hypothesis - a small probability value leads to the rejection of the null hypothesis of a normal distribution. Table 2 contains the summary statistics for daily returns for the complete PSI 20 sample and for two nonoverlapping subsamples, corresponding to the periods between January 1st 2003 and July 17th 2007 and July 18th 2007 and December 31st 2013, having the first sub-period 1185 observations and the second one 1685. Returns are calculated as log differences of the PSI 20 level. Returns are leptokurtic (fat tailed) and negatively skewed for the entire sample and both subsamples. Volatility is larger for the second sample, in which the market was in its bear state. However, the mean daily return is larger foe the first period. Serial correlations are generally small, with a few negative values. For the whole sample and both of the subsamples, we can reject the null hypothesis of a normal distribution, that the Jarque-Bera statistic is distributed as 𝜒2 with 2 degrees of freedom, proving so that the samples are normally distributed. As we can see, a simple “Buy-and-Hold” strategy for the whole eleven years sample produced mean daily returns of approximately 0,00414%, which is about 1,08% at an annual rate (the annual average return is calculated by multiplying the daily mean return by 261, which is the average number of trading days in a year). The results, however, are different for the two sub-periods. The first one, that corresponds to the bull market, as we above mentioned, produced daily returns of about 0,0722%, almost 18,84% at an annual rate, as, on the other hand, the second sample, corresponding to the bear market, produce negative daily returns, of about -0,0437%, meaning almost -11,41% at an annual rate. 4.3 Models application results After looking at the daily returns of the PSI 20 index, which characterize a simple “Buy-and-Hold” strategy, we will analyze the results of our models for the complete eleven year sample and for the two sub-periods, which are present in tables X-Y. The columns labeled “Buy” and “Sell” present the mean returns obtained during buy and sell periods, respectively: as mentioned on previous sections, a buy period, for example, is defined as the period after a buy signal up to the next sell signal, with the 33 sell period being the other way around. The column labeled “Buy-Sell” presents the difference between these two means. The complete sample results are presented below, on table 3. Model N(Buy) N(Sell) Buy Sell Buy > 0 Sell > 0 Buy-Sell Model A 1641 1212 0.0912 (0.0003) (0.0032) -0.1146 (0.0008) 0.545 0.477 0.2058 (1.6E-05) Model B 1670 1183 0.0963 (0.0004) (0.0019) -0.1270 (0.0008) 0.549 0.470 0.2233 (3.5E-06) Average 0.09375 -0.1208 0.21455 Table 3 – Results of the models application: complete sample, 2003-2013. ‘N(Buy)’ and ‘N(Sell)’ are the number of buy and sell days generated by the model. ‘Buy’ and ‘Sell’ are the daily mean returns on the buy and sell periods. ‘Buy > 0’ and ‘Sell > 0’ are the fraction of buy and sell returns greater than zero. Numbers in parenthesis are the p-values, for a confidence level of 95%, testing the difference between the mean buy returns and zero, between the mean sell returns and zero and between buy-sell and zero. The second number in parenthesis in the ‘Buy’ column is the p-value, for a confidence level of 95%, testing the difference between the mean buy returns and the unconditional mean returns. The mean returns are scaled by a factor of 100 for easier interpretation. For the complete sample, the differences between the mean buy returns and zero, the mean sell returns and zero, the buy – sell and zero and between the mean buy returns and the unconditional mean returns are significantly positive for both models. As we can see, for both models, the mean buy returns are positive, with Model A producing 0,0912% daily returns (about 23,8% at an annual rate), and Model B producing 0,0963% daily returns (approximately 25,1% annually), both models with results way superior to the unconditional mean 1,08% annual rate returns given by a “Buy-andHold” strategy, and that can be found characterized on the section above. For both models, sell mean returns are negative, with Model A producing an average daily return of -0,1146% and Model B of -0,1270%. The average return of the models of the buy periods is 24.5% at an annual rate. The “Buy > 0” and “Sell > 0” columns present the fraction of buy and sell returns that are greater than zero. The buy fraction is, for both models, greater than fifty percent, while that for sells is considerably less, being in the region of forty-seven percent. Under the null hypothesis that technical trading rules do 34 not produce useful signals, these fractions should be the same: a binomial test shows that these differences are highly significant and the null of equality can be rejected. Tables 4 and 5 report our findings for the two sub-periods investigated here. Model N(Buy) N(Sell) Buy Sell Buy > 0 Sell > 0 Buy-Sell Model A 861 307 0.1183 (0.0004) (0.088) -0.0572 (0.0008) 0.580 0.404 0.1755 (7.5E-05) Model B 794 374 0.1315 (0.0004) (0.033) -0.0538 (0.0007) 0.591 0.412 0.1853 (4.0E-06) Average 0.1249 -0.0555 0.1804 Table 4 - Results of the models application: 1st sample, 2003-July 17 2007. Meanings are the same as in table 3. Table 4 gives us the results for the first sample, in which a strong bull market occurred, and in which volatility decreased and the mean daily returns increased when compared to the overall sample. We can see that both models give us positive daily mean buy returns, with Model A producing mean daily returns of 0,1183%, about 30,9% at an annual rate, and Model B mean daily returns of 0,1315%, translating into approximately 34,3% annually. Both models, for the first sample, give us results superior to the unconditional mean returns, of approximately 18,84% at an annual rate. However, Model A fails to give statistical significant results when testing the differences between the mean buy returns and the mean unconditional returns, for a confidence level of 95%. That can be explained by the relatively low number of observations of this sample, as we would have significant results if we lowered the confidence level to, for instance, 90% (σ = 10%). All the other differences between means are statistical significant. Sell mean returns are negative for both models, with Model A producing an average daily return of -0.0572 % and Model B of -0.0538%. Also for this first sample, the buy fraction is, for both models, greater than fifty percent, while that for sells is considerably less, being in the region of forty to forty-one percent, with the same conclusions lay above. 41 producing mean daily returns for the full sample of 0,0817%, about 21,3% at an annual rate, a decrease of approximately 2,5% at an annual level due to transaction costs. The buy returns were also reduced in Model B, passing from the original 25,1% annually to about 22,8%. Although transaction costs did reduce our models profits, we can see that those reductions did not put in question the overall profitability of the model, and won’t affect our main conclusions. On tables 8 and 9, presented below, the models applications with the transaction costs are displayed. We can prove that Model B still produced positive results when comparing them to a ‘buy-and-hold’ strategy, whether the market is on a bull or bear state. As for Model A, it failed to produce statistically significant results, as the mean buy returns and the unconditional mean returns are not statistically different from each other, for both samples. Model N(Buy) N(Sell) N.Op Buy Sell Buy > 0 Sell > 0 Buy-Sell Model A 861 307 19 0.1121 (0.0004) (0.137) -0.0572 (0.0008) 0.580 0.404 0.1693 (0.0001) Model B 794 374 13 0.1270 (0.0004) (0.049) -0.0538 (0.0007) 0.591 0.412 0.1808 (7.8E-05) Average 0.1196 -0.0555 0.1751 Table 8 - Results of the models application, having in consideration the transaction costs: 1st sample, 2003 – July 17 2007. ‘N.Op’ is the number of operations that our models produce. Other meanings are the same as in table 3. Model N(Buy) N(Sell) N.Op Buy Sell Buy > 0 Sell > 0 Buy-Sell Model A 780 905 43 0.0492 (0.0007) (0.056) -0.1341 (0.0010) 0.508 0.473 0.1833 (0.006) Model B 876 806 45 0.0530 (0.0006) (0.041) -0.1608 (0.0012) 0.511 0.465 0.2138 (0.002) Average 0.0511 -0.1475 0.1986 Table 9 - Results of the models application, having in consideration the transaction costs: 2nd sample, July 18 2007 – 2013. ‘N.Op’ is the number of operations that our models produce. Other meanings are the same as in table 3. 42 Just like for the models results without transaction costs, we will replicate this same procedure for all the individual stocks that constituted our sample, and that can be found in the previous sections. Below, table 10 synthetizes the main results for each stock. Just like for the previous analysis, all stocks have proven to have mean buy and sell returns statistically significantly different than zero, as well as buy – sell returns also different from zero. Stock Full Sample 1st Sample 2nd Sample B-a-H M. A M. B B-a-H M. A M. B B-a-H M. A M.B ALTRI ¹ 25.0% 41.4% (0.283) 46.4% (0.157) 106.9% 109.3% (0.926) 109.6% (0.905) -5.1% 16.9% (0.197) 23.4% (0.089) BCP - 17.1% 17.4% (0.002) 28.7% (0.001) 15.1% 24.2% (0.484) 31.9% (0.203) - 39.8% 12.7% (0.009) 27.4% (0.002) BPI -3.9% 18.2% (0.072) 24.0% (0.025) 24.9% 23.6% (0.930) 23.2% (0.902) - 24.1% 14.6% (0.040) 24.5% (0.011) CIMPOR -1.7% 13.3% (0.119) 14.6% (0.098) 18.5% 16.1% (0.826) 17.2% (0.902) - 15.9% 11.2% (0.070) 12.7% (0.063) CIPAN - 15.4% 4.8% (0.724) -4.0% (0.862) 16.2% 35.2% (0.801) 20.7% (0.912) - 37.7% -19.2% (0.812) -23.1% (0.897) COFINA -5.8% 19.8% (0.084) 28.7% (0.016) 15.11% 19.3% (0.734) 23.4% (0.491) - 20.5% 20.3% (0.086) 36.5% (0.020) COMPTA - 25.9% 20.1% (0.111) 5.5% (0.323) -32.4% 9.0% (0.282) 5.9% (0.399) - 21.3% 25.9% (0.401) 3.3% (0.617) CORTICEIRA AMORIM 9.2% 24.0% (0.183) 26.3% (0.119) 21.3% 30.7% (0.471) 32.1% (0.399) 0.79% 19.3% (0.260) 22.4% (0.187) EDP 5.1% 12.0% (0.450) 10.7% (0.535) 22.4% 25.1% (0.824) 23.7% (0.894) -7.1% 2.9% (0.442) 0.01% (0.507) F RAMADA ² 5.5% 23.7% (0.490) 35.8% (0.276) - - - - - - F.C. PORTO - 19.4% 2.9% (0.312) 9.4% (0.201) -7.8% 5.0% (0.312) 4.9% (0.314) - 27.6% 0.2% (0.424) 12.9% (0.199) GALP ³ 9.5% 21.7% (0.454) 30.5% (0.194) 82.0% 84.0% (0.952) 84.0% (0.952) 1.25% 15.1% (0.430) 25.0% (0.178) GLINTT - 21.6% 44.6% 4.4% 12.7% 45.9% - 27.6% 43.5% 43 GLOBAL 19.1% (0.049) (0.005) (0.764) (0.169) 35.7% (0.041) (0.013) IBERSOL 5.2% 14.8% (0.284) 25.1% (0.049) 25.9% 23.9% (0.899) 24.2% (0.922) -9.4% 10.9% (0.195) 25.1% (0.041) IMPRESA 1.9% 35.4% (0.037) 40.9% (0.017) 27.5% 36.2% (0.511) 34.4% (0.600) - 16.1% 34.9% (0.048) 45.4% (0.019) INAPA - 21.5% 17.4% (0.014) 25.0% (0.004) -12.3% 5.7% (0.144) 10.9% (0.059) - 28.1% 25.5% (0.040) 34.8% (0.017) JERONIMO MARTINS 22.3% 26.3% (0.708) 31.5% (0.408) 26.9% 28.1% (0.922) 34.1% (0.572) 19.0% 25.3% (0.708) 29.6% (0.524) LISGRAFICA - 18.3% 2.1% (0.799) 30.3% (0.401) 25.7% 47.4% (0.613) 52.9% (0.527) - 49.2% -29.1% (0.871) 17.2% (0.579) MARTIFER 4 - 41.7% 0.0003% (0.018) -4.47% (0.030) - - - - - - MOTA ENGIL 10.1% 26.8% (0.156) 32.3% (0.061) 34.8% 33.1% (0.907) 35.4% (0.951) -7.4% 22.4% (0.091) 30.2% (0.037) NOS 0.7% 16.9% (0.109) 20.6% (0.049) 18.9% 19.5% (0.945) 21.4% (0.836) - 12.1% 15.3% (0.073) 20.1% (0.036) NOVABASE -7.3% 11.4% (0.040) 16.4% (0.010) -3.0% 8.3% (0.233) 10.1% (0.174) - 10.2% 13.8% (0.089) 21.1% (0.027) PORTUCEL 8.4% 20.9% (0.186) 21.2% (0.137) 21.9% 24.4% (0.883) 25.2% (0.765) -1.1% 17.5% (0.140) 18.4% (0.118) PORTUGAL TELECOM -5.0% 15.4% (0.037) 11.1% (0.111) 10.1% 15.4% (0.649) 12.8% (0.822) - 15.6% 15.4% (0.032) 9.9% (0.089) REDITUS 1.5% 20.9% (0.357) 28.9% (0.191) 37.0% 40.9% (0.829) 45.1% (0.711) - 23.4% 9.7% (0.327) 14.5% (0.320) REN 5 -3.2% 0.02% (0.623) 9.4% (0.249) - - - - - - SAG GEST - 12.9% 15.7% (0.046) 27.6% (0.005) 7.6% 15.9% (0.496) 20.1% (0.306) - 27.2% 15.4% (0.060) 33.1% (0.009) SDC INV. -1.5% 34.3% (0.061) 44.6% (0.016) 42.6% 51.1% (0.708) 55.1% (0.581) - 32.4% 22.6% (0.049) 37.5% (0.014) SEMAPA 8.2% 25.3% (0.055) 24.2% (0.074) 30.9% 36.8% (0.640) 37.0% (0.630) -7.7% 17.1% (0.042) 15.2% (0.059) SONAE - 27.4% 42.8% - - - - - - 44 CAPITAL 6 24.9% (0.047) (0.012) SONAE COM 4.5% 20.3% (0.164) 25.4% (0.065) 24.6% 25.9% (0.933) 32.7% (0.623) -9.7% 16.4% (0.091) 20.4% (0.052) SONAE INDUSTRIA - 18.3% 12.3% (0.036) 19.8% (0.011) 20.0% 28.8% (0.602) 33.0% (0.451) - 45.2% 2.1% (0.037) 9.89% (0.014) SONAE 12.8% 29.3% (0.161) 34.1% (0.049) 44.3% 41.5% (0.870) 45.6% (0.936) -9.3% 20.6% (0.061) 26.0% (0.027) S.L.B. FUTEBOL 7 - 29.7% 35.8% (0.036) 60.0% (0.006) - 317.4% 0% (0.437) 0% (0.437) - 22.5% 36.2% (0.046) 61.3% (0.009) SPORTING C.P. - 11.7% 11.5% (0.391) 11.9% (0.408) -2.4% 10.1% (0.483) 10.7% (0.492) - 18.3% 12.1% (0.523) 11.9% (0.568) SUMOL + COMPAL -2.9% 5.9% (0.489) 7.9% (0.398) 5.4% 7.8% (0.624) 13.2% (0.451) -8.7% 4.1% (0.397) 4.3% (0.379) TEIXEIRA DUARTE 8 -4.9% 37.3% (0.268) 54.0% (0.131) - - - - - - V.A. ATLANTIS 9 - 17.7% 4.8% (0.601) 32.8% (0.177) -23.8% 0.1% (0.474) 15.1% (0.284) - 13.7% 6.8% (0.699) 47.1% (0.227) Table 10 - Results of the models application, having in consideration transaction costs, for the full sample, and for the two nonoverlapping subsamples. Meanings are the same as in table 6. Both models, once again, fail in most cases to produce mean buy returns statistically different than unconditional mean returns, for a confidence level of 95%. In this analysis, as well, we will just consider stocks where all the assumptions present in section 3.4 are met, and the models reveal to be statistically significantly profitable. The overall analysis is quite similar to the one without transaction costs. Model A produced statistically positive results in 11 out of 38 stocks tested for the full sample. Once again, it performed better in Sport Lisboa e Benfica Futebol stocks, 65.5% better annually than a ‘buy-and-hold’ strategy, the same stock as Model B, which produced statistically significant results in 16 out of 38 stocks tested, producing results 89.7% better than the unconditional mean returns. For the first sample, Model A failed to produce significant statistical results in any of the tested stocks, happening the exact same with Model B. 45 As for the second sample, Model A was statistically profitable also in 10 of the 38 considered stocks, also performing better on Glintt Global stocks, 63.3% annually better than the unconditional mean returns (27.6% vs -35.7%). Model B was statistically profitable in 15 of the 38 considered stocks, performing better on Sport Lisboa e Benfica Futebol stocks, 83.8% better when compared to a simple ‘buy-and-hold strategy’ (61.3% vs -22.5%). The same conclusion regarding the performance of the models in bull or bear markets, and the performance of the models against each other, taken in the previous section, also occurred when taking in consideration transaction costs. We can also conclude that, despite the fact that they reduce the profitability of the models, and make them produce results overall less statistically significant, they doesn’t represent a great burden for the investors, and that they won´t affect our study’s conclusions, and, since that is the case, we will use the returns without transaction costs, when needed, in the next section. 4.5 Characteristics influence tests We will now test whether the stock’s characteristics, described in section 3.3, have any impact on our model’s profitability, comparing the returns of stocks with different profiles in terms of those characteristics. For this analysis, when needed, we will use the full sample returns, without having in consideration the transaction costs. The values for the market capitalization, needed to differentiate the stocks by their size, and for the turnover by value, needed to separate companies by their liquidity, will be the values of the last year of our sample, 2013, in order to keep our conclusions as relevant and actual as possible. The standard deviation, needed to measure each company volatility, will be the full sample one. As we want to test the effect of this characteristics in the model’s profitability, we will only consider the stock where our models revealed themselves statistically profitable, ignoring the other ones. In order to reach the conclusions that we want, we will separate the stocks that constitute our sample into two groups, one group with a big amount of a certain characteristic to be tested, and other group with a little amount of that same characteristic, “big caps” and “small caps”, or stocks with great volatility and stocks with little volatility levels, for instance. Each group will be constituted by 3 stocks. We 46 will then test if the mean daily return of both of these groups is equal, or statistically different. If it is statistically different, we will be able to conclude that that characteristic in test has in fact influence in the profitability of our models and in technical analysis in general. Our hypothesis will be defined as below described: 𝑯𝟎:𝑿𝟏−𝑿𝟐=𝟎 𝑯𝟏:𝑿𝟏−𝑿𝟐≠𝟎 Where, 𝑿𝟏 is the mean daily returns on buy days for the group 1; 𝑿𝟐 is the mean daily returns on buy days for the group 2; And the t-statistic being the following, with the same meanings as described in section 3.4: 𝒕= 𝑿𝟏−𝑿𝟐 (√𝒔𝟏 𝟐 𝒏𝟏+𝒔𝟐 𝟐 𝒏𝟐) We will perform these tests for both models. 4.5.1) Size The criteria we will use to distinguish the stocks in our samples by their size will be their market capitalization. So, in this case, our 𝑿𝟏 will be the mean daily returns on buy days for ‘Big Cap.’ stocks, and 𝑿𝟐 will be the mean daily returns on buy days for ‘Small Cap.’ stocks. a) Model A According to the used database, the companies with bigger and smaller market capitalization values in which Model A proved to be statistically profitable and that will constitute both our groups are, as follows: 47 Biggest Market Capitalization Values Smallest Market Capitalization Values BANCO COMERCIAL PORTUGUES GLINTT GLOBAL PORTUGAL TELECOM INAPA SEMAPA SDC INVESTIMENTOS Table 11 - Biggest and smallest stocks, in terms of Market Cap., where Model A revealed to be profitable. Our test results are the following: ‘Big Cap.’ Stocks ‘Small Cap. Stocks’ Daily mean buy returns 0.00079091 0.000982434 Number of Observations 8559 8559 Freedom Degrees 12994 T Critical (two tailed) 1.960146568 P-value 0.440979867 Table 12 - Results for the size influence test on Model A. Despite the fact that ‘Small Cap.’ stocks produce much bigger daily returns that the ‘Big Cap.’ ones, the p-value for this test is bigger than 0.05, and we cannot reject 𝑯𝟎 with a confidence level of 95%, thus, we cannot prove that the mean daily returns for both groups are statistically different, and, so, for Model A, we can say that the size of the stocks have no influence on its results. b) Model B According to the used database, the companies with bigger and smaller market capitalization values in which Model B proved to be statistically profitable and that will constitute both our groups are, as follows: Biggest Market Capitalization Values Smallest Market Capitalization Values BANCO COMERCIAL PORTUGUES GLINTT GLOBAL NOS INAPA SONAE COFINA Table 13 - Biggest and smallest stocks, in terms of Market Cap., where Model B revealed to be profitable. Our test results are the following: 48 ‘Big Cap.’ Stocks ‘Small Cap. Stocks’ Daily mean buy returns 0.00111632 0.001343324 Number of Observations 8544 8559 Freedom Degrees 14441 T Critical (two tailed) 1.960128271 P-value 0.392761754 Table 14 - Results for the size influence test on Model B. Once again, just like for the Model A analysis, ‘Small Cap.’ stocks produce much bigger daily returns that the ‘Big Cap.’ ones. However, the p-value for this test is bigger than 0.05, and we cannot reject 𝑯𝟎 with a confidence level of 95%, thus, we cannot prove that the mean daily returns for both groups are statistically different, and, so, for Model B, we can say that the size of the stocks have no influence on its results. 4.5.2) Liquidity The stock’s Turnover by Volume will be the criteria used to distinguish the stocks in our samples by their liquidity. So, in this case, our 𝑿𝟏 will be the mean daily returns on buy days for the most liquid stocks, and 𝑿𝟐 will be the mean daily returns on buy days for the less liquid stocks. a) Model A According to the used database, the companies with higher and lower turnover by volume values in which Model A proved to be statistically profitable and that will constitute both our groups are, as follows: Most Liquid Stocks Less Liquid Stocks BANCO COMERCIAL PORTUGUES S.L.B. FUTEBOL PORTUGAL TELECOM NOVABASE SONAE INDUSTRIA IMPRESA Table 15 - Most liquid, and less liquid stocks, in terms of Turnover by Volume, where Model A revealed to be profitable. Our test results are the following: 49 Most Liquid Stocks Less Liquid Stocks Daily mean buy returns 0.000666187 0.001065816 Number of Observations 8559 7417 Freedom Degrees 13092 T Critical (two tailed) 1.960145201 P-value 0.04725822 Table 16 - Results for the liquidity influence test on Model A. As we can see, less liquid stock produce daily mean returns much higher than the most liquid ones, and the test made can confirm that. As the p-value is smaller than 0.05, we are able to reject 𝑯𝟎 with a confidence level of 95%, and we can prove that, for the application of Model A, the liquidity of the stocks has in fact influence on its profitability, as it achieves better results when applied to less liquid stocks. b) Model B According to the used database, the companies with higher and lower turnover by volume values in which Model B proved to be statistically profitable and that will constitute both our groups are, as follows: Most Liquid Stocks Less Liquid Stocks BANCO COMERCIAL PORTUGUES IBERSOL BANCO BPI S.L.B. FUTEBOL NOS NOVABASE Table 17 - Most liquid and less liquid stocks, in terms of Turnover by Volume, where Model B revealed to be profitable. Our test results are the following: Most Liquid Stocks Less Liquid Stocks Daily mean buy returns 0.00098554 0.001207486 Number of Observations 8559 7416 Freedom Degrees 14491 T Critical (two tailed) 1.960127705 P-value 0.173728602 Table 18 - Results for the liquidity influence test on Model B. 50 Just like for Model A analysis, the less liquid stocks perform better that the most liquid ones. However, in this case, the p-value of our test is bigger than 0.05, and we cannot reject 𝑯𝟎 with a confidence level of 95%, thus, we cannot prove that the mean daily returns for both groups are statistically different, and, so, for Model B, we can say that the liquidity of the stocks have no influence on its results. 4.5.3) Volatility The stock’s Standard Deviation will be the criteria used to distinguish the stocks in our groups by their volatility. So, in this case, our 𝑿𝟏 will be the mean daily returns on buy days for the most volatile stocks, and 𝑿𝟐 will be the mean daily returns on buy days for the less volatile stocks. a) Model A According to the used database, the companies with higher and lower standard deviation values in which Model A proved to be statistically profitable and that will constitute both our groups are, as follows: Most Volatile Stocks Less Volatile Stocks S.L.B. FUTEBOL SEMAPA GLINTT GLOBAL NOVABASE SONAE CAPITAL PORTUGAL TELECOM Table 19 - Most volatile and less volatile stocks, measured by the Standard Deviation, where Model A revealed to be profitable. Our test results are the following: Most Volatile Stocks Less Volatile Stocks Daily mean buy returns 0.001101102 0.00071508 Number of Observations 6094 8559 Freedom Degrees 7852 T Critical (two tailed) 1.960266154 P-value 0.098432935 Table 20 - Results for the liquidity influence test on Model A. 57 Coutts, J. A. and Cheng, K. C. (2000), Trading rules and stock returns: some preliminary short-run evidence from Hang Seng 1985-1997, Applied Financial Economics, 10, 579-86. Coval, J. D., and Shumway, T. (2002), Do behavioral biases affect prices?, Harvard University working paper, September 2002. Efron, B. (1979), Bootstrap methods: another look at the jackknife, The Annals of Statistics, 7, 1-26 Fama, E. F. (1970), Efficient Capital Markets: A Review of Theory and Empirical Work, The Journal of Finance, 25, 383-417. Grinblatt, M., and Keloharju, M. (2001), How distance, language, and culture influence stockholdings and trade, Journal of Finance, 56, 1053-1074. Hoffmann, A. O. I. and Shefrin, H, (2014) Technical Analysis and Individual Investors, Journal of Economic Behavior and Organization, Forthcoming. Hrušová, I (2011), How Rewarding Is Technical Analysis? Evidence from Central and Eastern European Stock Markets, Faculty of Social Sciences, Charles University in Prague. Hudson, R., Dempsey, M. and Keasey, K. (1996), A note on weak form efficiency of capital markets: the application of simple technical trading rules to UK prices 19351994, Journal of Banking and Finance, 20. 1121-32. Isakov, D. and Hollistein, M. (1998), Application of Simple Technical Trading Rules to Swiss Stock Prices: Is It Profitable?, Papers 98.2, Ecole des Hautes Etudes Commerciales, Universite de Geneve. Jensen, M. C. (1978), Some Anomalous Evidence Regarding Market Efficiency, Journal of Financial Economics, 6, 95-101. Kirkpatrick, C. and Dahlquist, J. (2011), Technical Analysis: The Complete Resource for Financial Market Technicians, Financial Times Press, 2nd Edition. 58 Kyle, A., and Wang, F. (1997), Speculation duopoly with agreement to disagree: Can overconfidence survive the market test?, Journal of Finance, 52, pp 2073-2090. Lane, George C. (1984), Lane´s Stochastics, Stocks & Commodities, V. 2:3, 87-90. Lento, C. (2007), Tests of technical trading rules in the Asian-Pacific equity markets: a bootstrap approach, Academy of Financial and Accounting Studies Journal, 11, No. 2. Lewellen, W. G., Lease, R. C. and Schlarbaum, G. G. (1980), Portfolio Design and Portfolio Performance: The Individual Investor, Journal of Economics and Business, 32(3), 185-97. Marshall, B. R., Qian, S. and Young, M. (2009), Is technical analysis profitable on US stocks with certain size, liquidity or industry characteristics?, Applied Financial Economics, 19, 1213-1221. Marshall, B. R., Cahan, R. H. and Cahan, J. M. (2010), Technical analysis around the world, Working Papers Series, Available at SSRN: http://ssrn.com/abstract=1181367. Moskowitz, T. and Grinblatt, M. (1999), Do Industries explain momentum?, Journal of Finance, 54, 1249-90. Nascimento, V. (2007), Eficiência informacional do mercado de acções: o caso Português, Faculdade de Economia da Universidade do Porto. Neftci, S. N. (1991), Naive trading rules in financial markets and Wiener-Kolmogorov prediction theory: a study of technical analysis, Journal of Business, 64 (4), 549571. Nison, S. (1991), Japanese Candlestick Charting Techniques. New York Institute of Finance, New York. Odean, T. (1998), Are investors reluctant to realize their losses?, Journal of Finance, 53, 1775-1798. Odean, T. (1999), Do investors trade too much?, American Economics Review, 89,5, 1279-1298. 59 Osório, J. (2010), A análise técnica e o mercado Português: MACD e RSI, Caderno do Mercado de Valores Mobiliários, 35, 42-61. Park, C.-H. and Irwin, S. H. (2004), The profitability of technical analysis: a review, AgMAS Research Report. Pereira, N. R. S. (2011), A utilidade da análise técnica e a eficiência dos mercados: aplicação ao mercado accionista Português, Faculdade de Economia da Universidade do Porto. Pirisi, F. and Vasquez, A. (2000), Simple technical trading rules of stock returns: evidence from 1987 to 1998 in Chile, Emerging Markets Review, 1, 152-64. Pring, M. J. (2002), Technical Analysis Explained: The Successful Investor´s Guide to Spotting Investment Trends and Turning Points, New York: McGraw-Hill, 4th Edition. Ray, S. (2012), Revisiting the Strength of Dow Theory in Assessing Stock Price Movement, Advances in Applied Economics and Finance, Vol. 3, No. 3, pp. 591-598 Rhea, R. (1932), The Dow Theory. New York: Barron’s. Russell (2013), U.S. equity indexes: construction and methodology, Russell Investments Shiller, R. J. (2003), From Efficient Markets Theory to Behavioral Finance, Journal of Economic Perspectives, 17, 83-104. Silva, N. (2001), Eficácia da análise técnica no mercado accionista Português, Grupo de estudos Monetários e Financeiros – Faculdade de Economia da Universidade de Coimbra, 9. Talley, K. (2002), Stock forecasters are having a technical rise in popularity, Wall Street Journal, 26 June 2002. Vasiliou, D., Eriotis N., Papathanasiou S. (2008), Incorporating Technical of Analysis into Behavioral Finance: A field experiment in the large capitalization firms of the Athens Stock Exchange, International Research Journal of Finance and Economics, 9, 14, 100-112. 60 Vaz, S. M. O. (2012), How Efficient Is The Portuguese Stock Market?, Instituto Superior de Economia e Gestão. Wilder, J. Welles (1978), New Concepts in Technical Trading Systems, Trend Research 61 Appendix Appendix 1) Transaction costs calculations Broker Transaction cost (per operation) Banco ActivoBank, SA 0% Banco BPI, SA 0% Banco Comercial Português, SA 0,104% Banco de Investimento Global, SA 0% Banco LJ Carregosa, SA 0,026% Banco Popular Portugal, SA 0,125% Banco Português de Investimento, SA 0,099% Banco Santander Totta, SA 0,104% Banif - Banco de Investimento, SA 0% Banif - Banco Internacional do Funchal, SA 0,318% Barclays Bank PLC - Sucursal em Portugal 0,083% BEST - Banco Electrónico de Serviço Total, SA 0% Caixa - Banco de Investimento, SA 0,099% Caixa Central de Crédito Agrícola Mútuo, Crl 0,13% Caixa Económica Montepio Geral 0,208% Caixa Geral de Depósitos, SA 0,104% Dif Broker - Sociedade Corretora, SA 0,2% Fincor - Sociedade Corretora, SA 0,2% Golden Broker - Sociedade Corretora, SA 0,5% Montepio Investimento, SA 0,234% Novo Banco, SA 0,135% Novo Banco dos Açores, SA 0,135% Orey Financial - Instituição Financeira de Crédito, SA 0,05% Average 0,168% Source: Comissão do Mercado de Valores Mobiliários (CMVM)