TW200532524A - Method for constructing investment system - Google Patents

Abstract

This invention relates to a method for constructing investment system, using data mining technique to perform analysis on information content of public financial statements in basic analysis for selecting stocks with healthy financial structure and intrinsic values. Regarding technical analysis, artificial intelligence is used to prioritize orders thereby constructing an investment combination to serve as stock purchasing reference for investors. The object of this invention is to establish the basis for recommending investment combination, using decision tree classification, neural network, and gray theory to analyze enterprise financial index and technical analysis index. In summary, this research is summarized to provide the following aspects: (1) searching for key financial ratios that are essential to healthy financial structure and intrinsic values; (2) using decision tree classification to locate the decision attributes for selecting stocks; (3) using gray theory to prioritize stock orders; and (4) establishing standardized investment procedure serving as reference for investors.

Description

200532524 IX. Description of the invention: [Technical field to which the invention belongs] The main purpose of the investment system architecture method of the present invention is to establish investment portfolio recommendations based on the use of decision tree classification, neural network-like, and grey theoretical concepts to analyze corporate financial indicators And technical analysis indicators. To sum up, the present invention can be summarized in the following directions: 1. Through neural network-like analysis, look for the key financial ratios related to physical fitness and intrinsic value. 2. Use decision tree classification to clearly find the decision attributes for screening stocks. Lu 3. Using gray theory to find the priority order of stock trading. 4. Establish a standardized investment procedure for investors as a reference. The general manufacturing industry for Taiwan listed companies is used as a research sample. The research period is from November 2000 to April 2005. The source of financial information is the Taiwan Economic New Newspaper Database and the timely information information winner database. The data belongs to Season information. In order to avoid different financial comparison basis, the financial industry, life insurance industry, securities industry, futures industry, investment and credit industry, investment industry, and venture capital industry are excluded. [Previous technology] The stock market is one of the most important sources of capital for enterprises, and it is also an important financial market for short-term or long-term investment of corporate funds. Therefore, stock investment decisions and financial operations are also an important decision for modern enterprises. The size of Taiwan's stock market has been expanding. By the end of March 1994, there were more than 700 listed companies. It was not easy for companies seeking financial stability and promising prospects in the vast stock market. In the past, scholars proposed many analytical methods to predict the stock market trend, analyze the attributes of various investment targets and the characteristics of technical indicators, try to find out the characteristics of the stock market or investment targets, and establish valuable decision-making models. The past research directions can be roughly divided into two categories, basic analysis and technical analysis. Graham, the godfather of basic analysis, believes that the value of stocks should come from the results of the company's operations. The most direct and quantifiable data of the company's operating results are recorded in the company's financial statements. Therefore, the financial statements and financial ratio data have become an important reference basis for assessing the value of a company. Some people will use the example of Enron, WorldCom, or Boda, Huangtong Company to make false accounts κ 200532524 to fudge the bridal countability of the county. In addition, the financial details are technical analysis, and the spirit department uses statistics. This principle helps investors find the rhythm and context of price fluctuations, and develops technical indicators through various graphics and numbers. In order to choose the right one, take a bigger paste, or reduce the loss. The basic assumption is A history-repeat it again. As long as the deviated ambiguity is changed, the trend of the stock price axis can be compensated, and the profit can be obtained from it. Both of these analysis methods have their advantages and disadvantages. For example, fundamental analysis can judge the reasonable value of an enterprise in essence, but it requires a lot of information and knowledge, and the society is misled by incorrect information of companies or their factors; However, traditional technical analysis lacks a single-standard criterion, and its validity still needs to be strengthened. For example, different indicators may conflict with each other and cause investor interference. The analysis of the above two analysis methods has the same system, but this does not mean that the trend of the stock market is unpredictable. According to the hypothesis of inefficient markets (Fan An, 1999): Investors are completely female, and dragons ca n’t be compared quickly. Obtaining information must bear additional information costs, and the power of a small number of investors is sufficient to affect changes in stock prices. There may be many ways in the efficiency market to predict future changes in stock prices. [Summary of the Invention] The main purpose of this county is to provide a design of a unified architecture method. After discussing related articles, it is difficult to ride a dragon with a ship, and analyze the _, and construct an investment based on the analysis results. The decisive formula and process, the entire invention steps are as follows: ★ Literature exploration · Collect and review the city's traffic and weeding papers, in order to understand and plan the Internet and gray theory, and how to properly take stock market transactions into consideration The analysis of the data is easy. It takes all the manufacturing companies of Taiwan listed companies as the target, and its financial statements of the stocks in the past four years. 3. Find out the analysis attributes, and define all the variables and belongings in detail. 200532524 IV. Decision tree, grey theory and neural network-like theoretical analysis: Construct analysis flata, and construct a standard process that can be used for investment practice. Mouth five. Simulated investment: Use historical data to verify the return on investment of the constructed investment model. Whether it is better than the market performance. The techniques used include statistical analysis methods, sequence correlation analysis, artificial intelligence, neural network and machine learning, including quantitative analysis. And qualitative analysis. With the improvement of algorithm ^ program, the data mining technology can perform real-time online analysis and share the result f through the network to achieve the current mobile trend. I [Embodiment] The present invention The system is an investment system architecture method. It uses neural-like networks to solve unstructured problems, and uses its sensitivity analysis to find key financial ratios of good or not in financial ratios, and then uses the advantages of decision tree classification. The financial ratio found by sensitivity analysis is trained by a decision tree to form a decision to screen stocks. Finally, gray correlation is used to analyze the technology. Gray correlation can be used to find them between factor sequences with a small amount of information. To determine the final investment portfolio. Each algorithm will be described in detail below, please refer to the attached figure 1: The method of the research method is shown in the diagram, the data exploration is divided into six steps (Fayyar (J · M ·, iggg) ··· Data selection, data pre-processing, data conversion, extraction of meaningful information, interpretation of information, and information expression. ^ Is to provide a method to solve non-structural problems. General structural problems, after finding the corresponding mathematical formula, can be used by computers to solve a large number of and repeated calculations. In addition, Breiman proposed an algorithm in 1984 , Which is called CART (Classification and Regression Tree). Like neural networks, it can also find the rules between independent variables and dependent variables. Because it will represent all independent variables as a clever Tree Come out, so we can use simple rif-then ”method to show its rules, provide these rules to relevant experts to determine whether these rules are reasonable, and finally the experts make correct decisions based on these rules. And use factor analysis and The data exploration technology of the decision tree classification technology 'established a financial early warning model of its financial ratios, and its research findings found that in the case of the economic crisis 200532524' the main impact variable of the bankruptcy early warning model was "fixed assets to shareholder rights" leverage ratio ", "Gross margin", "Current ratio", "Quick ratio", "Inventory turnover ratio" , "Net margin", "net margin on long-term funding ratio", "cash flow rate and" long-term debt to cash flow ratio. " Grey correlation analysis: Li Junming (2002) used the segmentation and fuzzification of Edwanis-Bell-GhlsGn to evaluate the EPS (earnings per share) forecast value & growth rate in a face-to-face method (a value of 4 cents). Propose active fuzzy model and conservative fuzzy model. Artificial neural net-like (artificial neural netwrks) is a kind of button that uses computers to imitate biological neural chain patterns to process information. It can be used to assist large and complex calculations. The advantages of human neuron chains can process large amounts of data. Handle practically complex problems, and only need to use the calculation mode of the appeal form to complete the process. When the signal passes through the neuron, the original signal size can be changed due to the weighting effect of the neuron. The expression of the artificial neuron, and the relationship between the input and output values of this artificial neuron can be expressed by the following formula: η ~

Yif) = f ΣΜ; ΐχχι (ΐ) -〇 ............ ... η χ, -_1 ............ Feet: Neuron weighted value imitating biological nerve cells I: Input signal received Θ: Imitation The threshold value of the nucleus (meaning partial weight value) of the biological nerve cell and the input signal φ of the conduction must be greater than this value before being transmitted down.

Input layer artificial neuron output layer 200532524 In the class of neural networks, it is divided into supervised learning and unsupervised learning. The former requires trained input variables and output variables to build a complete network, but the latter requires only trained input variables. Yes, there is no need to output variables. The reverse transmission neural network used in the present invention belongs to the former, which is one of the neural networks commonly used in the present invention, and is widely used in prediction, classification, sample identification, noise processing, and the like. The basic principle of the inverse transfer network is to use the steepest slope drop method to minimize the error function. Because there are more hidden layers with the basic perceptron network, the interaction between input layer units can be expressed in the network. A smooth and differentiable transfer function is used, which makes it possible to derive the network's weights using the steepest slope-descent method described earlier. Please refer to Figure 2: The basic architecture of the inverted transmission network. The inverted transmission network is mainly composed of three layers of neural networks. The two layers are the input layer, the hidden layer, and the output layer. The following is the main explanation of these three layers. The round-robin layer is mainly used to represent the input variables of the network, the number of which depends on the topic, and the input function is usually represented by a linear function. Hidden layer: It is used to indicate the interaction between input units. There is no certain conclusion on the number of this layer. It is usually determined after testing. The number of network layers can be zero to multiple layers, and the conversion functions used are mostly non-zero. Linear transfer function, but linear transfer functions can also be used. # Round-out layer: It is used to represent the output variables of the network. The number of processing units is also determined by the topic. The output functions are mostly non-linear conversion functions. Since the input unit value processed by the n-th layer is provided by the value of the pre-layer round-out unit, the calculation formula of the network can be expressed as follows: V i J ........... .............. The output value of the j-th processing unit in the nth layer /: layer-to-layer conversion function% ·: the j-th processing unit in the n-th layer το And the weight of the (n-1) th soil processing unit 200532524

A: Bias weight for the jth processing unit. Unit = The main purpose of learning is to reduce the network inference output unit and its target output value U. Quality will generally use an energy function (or error function) to Represents the suppression of network science s, and the function (error function) is expressed as follows: E-Da (Tk ~ Yky ............ ....................................... A 1st The target output value of the output unit 4: the inferred output value of the k-th output unit. The main purpose of this energy function (error function) is to modify the weight π of the association on the network. The function "quote" is the most expensive, that is, the inferred output value of the network approaches the actual target output value. In his research, the method of reading the steepest descent method by females has minimized the Ruuri function (error function). The weight correction value is shown in the following formula: Na Er-η Man Bu ... ............................... (4) ^: Learning rate (l_ing rate), which is used to control the magnitude of the correction when the weights are iterated. For the process of its correction, between the output layer and the hidden layer unit, and the hidden field and its weight correction formula. , _ _ Each miscellaneous ^ ^ layer unit to do-explain. Equilibrium weight and partial weight value 1. Derivation of weight correction formula between output layer processing unit and hidden layer processing unit. PiF · (3)

AW jk II—η

dE Using the chain rule, the right half of Equation 5 can be changed to: dE ..... (5)

dW jk < Tk-Yk) xf \ netk) xYj: weight value between the kth processing unit in the output layer and the jth processing unit in the hidden layer Fy: the output value of the jth processing unit in the hidden layer is defined as the output The error amount of the k-th output processing unit of the layer: π (6) 200532524 h two ih-Yk) xf'—k) ............. ....................... The network output layer processing unit and hidden layer processing The weight correction between units is: β = —η ^ ΓηχδΐζχΥ] .. The correction amount of the partial weight of the output unit is:. · ⑻ Called: IW, — * ............... (9) Office: bias of the kth processing unit in the output layer value

2. Derivation of the weight correction formula between the processing unit of the hidden layer and the processing unit of the input layer: _ dE ij-71 ^^ .............. ........................................ . (10)

Using the chain rule, the right half of Equation 10 can be changed to: (λ — ^ f \ netk) xYi ............... ... cyvU Kk J

%: Weight K between the jth processing unit of the hidden layer and the ith processing unit of the input layer. The output value of the ith processing unit of the input layer lets A be defined as the error amount of the jth round out processing unit of the hidden layer: ~. 舍 ，] ", (,) ............................ ........... Then the weight correction amount between the network hidden layer processing unit and the round-robin processing unit is: ......... .............................. (13) Same It can be proved that the partial weight correction of the losing dj unit is: 峋 If yin ............. VHide "Layer No." · Partial weight of each processing unit (14) From the above results, we can get the correction value of the weight value between each layer and the correction amount Δθ of each partial weight value 200532524, but we did not mention the conversion above. In the form of a function, if the conversion function is a sigmoid function, the conversion function can be expressed by the following formula ... f \ X) ..... and Δ% can be rewritten as the following Formula: ....... (15) (16) ^ Wjk ^ 1 {Tk ^ Yk) xYkx {\ ^ Yk) xYj and (17) -Wl) FT: initial weight of network training for class By secret green, the post of sub-Wei shot and thread, but should (iv) will be selected on the network weights. For example, the sensitivity analysis of Karnin's weight for the target = output value in the year of the year, and the sensitivity analysis to understand which weight link is more important for this project. Then delete the less sensitive link weight and hiding Layer unit element, the purpose of adaptive modification of the inbound and outbound transmission network. The sensitivity 3 is expressed as follows: Built to promote (18): the weight at the end of the network training

: The error value of the output layer when the weight is W. The error value of the output layer when the weight is W. When the weight is, the molecular part in Equation 18 can be written as

........... (19) Bringing formula 19 into formula 18 can be modified to become $, and s is shown in formula 20 below: Ν-Λ

^-Σ ^ -Wa ^ W

(20) 200532524 In addition, we know Δ% from the above, so we can modify the formula 20 as follows:

A can calculate the sensitivity of each weight after obtaining S, and temporarily freeze or delete the weight with lower sensitivity. The general class> Bayesian algorithm often uses Decision Tree, which first divides the original data into two parts: training data and test data. Then, each piece of training data is used to establish a decision tree according to its attributes and classification attributes. Finally, test data is used to verify the correctness of the decision tree. Basically, a decision tree can be regarded as a Boolean function. The input value of the function is a set of properties of a certain situation, and the output value is a decision value of "yes" or "no". Each internal node in the decision tree represents a possible value for this attribute; at the end, each leaf node corresponds to a target class. The following figure is a famous example in the decision tree analysis method; its target category is WillWait, which means "customers are willing to wait when they are full," and the "patrns" that appear in the figure represent "old customers". , WaitEstimate represents "estimated waiting time", Alternate indicates "there are other restaurants to replace" and Hungry indicates "is hungry" * The above nodes are rewards for these attributes, other values are "Yes" or " The nodes of "No" are both leaf nodes and represent a decision value. Please refer to the attached figure 3. The decision tree classification method shows that from the implied sentences mentioned above, each implied sentence analyzed by the decision tree only refers to one object. Therefore, the expressive ability of the decision tree is the same as that of zero-order logic (pr〇p〇siti〇nal 1〇gic); therefore, the decision tree cannot describe the situation where two objects are referenced in a test at the same time. By analogy, please refer to Figure 4. Property Patrons are shown in the training set of the test. Information theory was proposed by Shannon in 1949. It was first used to deal with communication-related problems. Quinlan proposed the ID3 decision tree induction algorithm using information theory as the basis for selecting test attributes in 1979. This method is based on the assumption that an event has n types of results. The probability of these n types of results is 200532524 (ΜχΛ / Ο "). These probabilities are known in advance and the information obtained is: When this event occurs, via This event (22) / (P (v〇5KK3P (vJ), | > p (Vj) 10g2P (v) The above formula (22) represents the average of the events required to express the meaning of the event in the binary format. Number of bits. There are many ways to choose attributes in information theory. Χ / 戍 'information gain

Quinlan proposed in 1979 'Daggers to measure the amount of information to collect more information to calculate the amount of information in each category', and then calculated the average amount of information (also known as the hearing and chat) to express the secrets of the collection. . I advise __ 4 axis county office tree category C々U3K, the number of data is expressed as ~ = 〇), the number of all data in the shape of bismuth's, this remuneration is simple d {Lai rate can be blessed According to the message theory, the information of each aspect is -10 g2 (^^). The training set contains training breeding materials for each category. From the amount of information in each category, the average amount of information in the training set can be calculated ( That is the degree of disorder). The sum of the information volume of all categories is multiplied by the information volume of each category and the outgoing probability of each category of data is: inf〇 (^), ^ ί ^ 1χ0% 2 (^ 1 ^) .. ... (23) According to the calculation method, when the set S is divided according to a certain attribute A When it is divided into multiple sub-sets, the amount of information it occupies after division is equal to the sum of the information amount of each sub-set multiplied by the proportion of each sub-set: ^ 〇A (S) ^-± ^ xMo (Si) .. ....................................... ( 24) μ rl 200532524 Therefore, the amount of information obtained by the set s after attribute division is the amount of information before the division minus the amount of information after the division, which is expressed as: gain {A) ^ Mo {S) ^ mi〇A { S) ............... ........... (25) The ID3 learning method is to select all the miscellaneous postal values and select the largest gain value as the classification attribute. The decision tree uses this attribute Attributes are split into multiple training sub-sets to form multiple numbers. Repeat the above steps for each sub-tree.

Finding the largest gain value as a classification attribute, and splitting it into multiple subtrees until it can no longer be divided. (B) the gain ratio attribute selection method

In 1986, Quinlan modified the attribute selection method in the ID3 decision tree induction algorithm to normalize information gain to test attribute information (n〇rfflalizati〇n), which is called gain ratio. The following formula (26) (26) is (and especially) a gain ratio, / ⑺-4 ^, Z) is the information gain of the attribute ^, and / F (々) is the information of the attribute. Generally, attributes with a lot of values have relatively inferred professionalism. Similarly, their information is also relatively high. Therefore, the normalized action can reduce the deviation caused by nf〇rmati㈤ gain in this respect. Similarly, the gain ratio attribute selection method has its problems. (1) After testing with test attributes, if there is only one instance in a subset, the attribute information will be zero. Therefore, the denominator of the above formula (26) may be zero, which means that this formula will be undefined. (2) The denominator / f (4) in the above formula (26) exists mainly to compensate for the deviation of the information gain. Therefore, basically, the attribute with a large inf〇nnati〇ngain is selected. However, if the information gain is not large and its / Γ (4) is small, it may make the gain ratio very large. Therefore, it is a wrong 200532524 situation to use this property. (2) Distance-based attribute selection method In order to solve the two problems of the gainratio attribute selection method mentioned above, Mantaras proposed the distance-based attribute selection in 1991. law. A group of subsets separated by a certain attribute test is called a partition (partition), and a group of subsets separated by a category is called a correct partition. Among all the segmentations, the segment with the smallest normalized distance from the correct segmentation, and the corresponding attribute is the attribute we choose. ,,

The following formula ⑼ is the definition of dist (dist fiber), the mismatch between the correction and &

The test of sexual A is then performed on Wei B, and the remaining information is tested on Mi A and then on attribute B. ........ (27) The following formula ⑽ is the formal definition of the normalized distance. The normalized distance between the segmentation ^ and the heart is the distance between the segmentation heart and & divided by the distraction and & intersection. Information.

Dn {Pa ^ b)-wM) ... …… .. (28). The substituting of mathematical formulae in Beixun's theory changes the normalized distance to a formula with lnf_tiGn gain, which becomes the following formula ⑽. In the distance-based attribute selection method 'look for the shortest distance to the correct dividend regularization distance, and change the angle. 0 ^ W ^^^ (29) gain ral ^ The formula is very similar' because both are correct inf_ti〇n fine regularization.

Pc 'Pv ι (ρ〇) ~ ι \ G ^ x)-^ ........................ ........ — surface,] ............ ........... First of all, 'Because ADR will be zero, there will be no undefined 19 200532524 in the formula of the normalized distance. Furthermore, because the following formula (31) always holds Gain {AKx) < l (pvr ^ Pc) .............. .............. So there will be no prejudice in gainratio because the selected attributes are information Gain is the greater. In addition, the designer of the algorithm, Mantaras, also used the same training set as the gain ratio method to prove that the distance-based attribute selection method can produce a smaller decision tree than the gain ratio attribute selection method. ★ In the case of ambiguity of the system model and incomplete information, system correlation analysis and model construction are performed to solve the problem that the traditional probability statistics need a large number of sample hypotheses in practice, so that researchers can get a small amount of funding. Under certain conditions, after specific data processing, find the ridges and youngsters among the sequence of factors of Lai. Based on this, the non-green state of the research object can be estimated and inferred because it is not limited to the condition that the number of samples must be large. It is becoming more and more widespread, because this method can "white" the "grey" relationship between the factors of the gray system, so it is called gray correlation analysis. The following introduces Deng Julong's gray correlation degree formula. Deng Julong made use of the first grey correlation formula that satisfies four, which is mainly divided into two major items. 1. The grey correlation coefficient is: where + ........ ....... (32) ^ 1) / = 1,2,3, Λ, m. K-1? 2,3A y € / (2) is the reference sequence and x is a specific comparison sequence. The absolute value of the first difference between ⑺ ~ and '(modulo: Nom). ⑷ ,,, △ Ka = v7: M) _⑽) (5) C ·· Identification coefficient, "[〇 ,!]. 70 200532524 Second, the gray correlation degree: Take the average value of the gray correlation coefficient. · 1 17 r〇 / =-Tr (x〇 (kiXifk)))) ........................... .. (33) □ Graphic obtained through mathematical operation

The graph of the traditional gray correlation coefficient, it can be clearly seen that the graph of the gray correlation coefficient is -hyperplane, and is non-linear, so the disadvantage of the above equation is that it can only be analyzed qualitatively, but not quantitatively. . _There is no such thing as the silk on the sleeve, so this plan, if adopted, is the first invention. The above is the description and research operation of the investment system of the present invention. The present invention uses the Top to Bottom analysis method to first determine the short position of the market and determine the strategy of « Freshen the procedural stocks and form the investment surplus. Please refer to Figure 5. The investment system architecture table is shown in detail. The investment system architecture step 1: The long and short stock price index is the leading indicator of the overall economy, and the company ’s breadth is a good indicator of the operating conditions of all listed companies. In the embodiment, the present invention uses the single quarterly quarterly report of all listed companies, and uses the post-tax shareholder return rate (臓) to be greater than 8 as a threshold to measure the overall status of all listed companies. 200532524 R * = good company ^ l ± m ^) xl0 ° Again, let's compare □ R 木 〇 □ 财 ^ / Compared with the stock price index at the time of the quarterly quarterly announcement; assuming that it represents the operating status of all listed companies — The period is good, so the stock price = the trend of the current quarter F to the previous season; the price of the fake secret coffee is higher than _, but the broader market at (Mw)) has a high consolidation or increase. The trend of this market is inconsistent with the actual operating conditions of all listed companies. When operating in a quarter, a shorting strategy is adopted.

The decision tree algorithm of the present invention is a binary classification. In order to reduce the gray space between healthy and unsound financial health, the matching of the top 30 MSCI constituent stocks (excluding financial related industries) and the full delivery stock is used to observe the MSCI The constituent stocks of the Taiwan Index have the characteristics of equity, large market capitalization, sound financial constitution, and high exclusivity in various industries, and the full-delivery shares are based on the Taiwan Stock Exchange 2005). 200532524 Co., Ltd. Business _ Article 9 Provisions In the event of a listed company's prescribed circumstances, its listed securities are traded through the Taiwan Securities Exchange, a nuclear greener, and a greener private-seller scheme. ° Article 49 stipulates as follows. If a listed company has any of the following circumstances, the company shall request the competent side to approve the change of the original transaction wire for the listed securities of the company: Sixteen Laidian announced and declared the most official bribe report showing that the net value has been lower than one-half of the paid-up capital. 2. The regular shareholders' meeting was not held within six months after the end of business.

3. Those who have issued a qualified opinion by an accountant. 4. Violation of the relevant provisions of the listed company's major information verification and disclosure regulations, the notification of the supplementary bank to handle the public procedures, failed to complete the deadline and the case is significant. 5. Claim to the court for reform in accordance with Article 282 of the Company Law. 6. Company law considers it necessary for other reasons. The MSCI index refers to the constituent stocks and the full-delivery stocks of the season respectively. The msci constituent stocks = 0, the physique 2 kings, and the king's father cut stocks = imperfect physique. These two groups are used as training samples for decision trees. Pass 6: The top 30 constituents of the Taiwan Stock Index, and refer to the training samples of the lions. The total amount of MSCI constituent stocks in the project is 1 1301-Taisu 1450 2 1303 South Asia 1602-Taidian 3 1326- Taihua 1805. Kaiju 4 2002-Sinosteel 2342-Silicone 5 2303-UMC 3053-Dingying 6 2311 · Sun Moonlight 2506 · Taishi 7 2317-Hon Hai 2525-Baoxiang 8 2324-Compal 2528-Huang ^ Pu: 9 2330-TSMC 2MO-Lin III 10 2353-Acer 4801-Maastre 11 2357-Asus 5011-Jiuyang 12 2382-Quanta 5505-Hewang 13 2412 · CLP Power 5702-Unification 14 2454-MediaTek 5901 15 8913-Huaxia Leasing 200532524 Step 3: Sensitivity analysis of neural network-like: The input variable of the decision tree is the financial ratio and the output The variable is (sound health = 0, unsound health = 1). The attributes of its financial ratio can be divided into financial structure stability, short-term debt service ability, long-term health ability, operating ability, turnover ability, productivity index, cash flow index, There are 105 indicators in eight categories of market price indicators. In order to measure the importance between input variables and output variables in order to screen the important details of Wei Jin's determination, we first use sensitivity analysis as a method of riding financial ratios. Step 4: Use cash basis and other basis to enter the decision tree. Beaver (1966) In the study of “Financial Ratio Prediction of Operation Failure”, it was found that “cash flow / liability_” is the best indicator of Lai Linniu, please see _Figure 7 : The cash flow indicators are not 'independent of the cash flow of this county &come; the decision process of mosquitoes entering the decision-making process produces decision criteria' and the financial ratios of the other seven outlines Financial ratios of 1.5 and greater than 2, and then enter the training process. And refer to the decision tree before training training as follows:

Step 5: Considering the return on equity (ROE) Buffett believes that the company's stock price is determined by its intrinsic value, and the stock price will swing up and down the value formed by the intrinsic value 94 200532524 as shown in the following table, and the return on shareholder equity ( RO (E) is the best definition for measuring intrinsic value. In this research, ROE is used as the first screening criteria for screening stocks, and then the criteria formed by the decision tree are used to perform screening.

Intrinsic value and stock price relationship Step 6: Gray relation ranking The advantage of gray correlation is that researchers can find their correlation between random factor sequences with a small amount of information and specific data processing. Step 6 is all the listed companies. After step 4 and step 5, the remaining stocks are screened. In this step, through the selected technical analysis indicators, find out the purchase priority and determine the final investment list. The technical indicators selected in this study are the buying momentum indicator (AR), the willingness to buy and sell indicator (BR), the shock indicator (0SC), the psychological indicator (PSY), the relative strength indicator (RSI), the quantity strength indicator (VR), and William. Indicator ⑽S% R), the detailed description of which is shown in Table 1 'and the criteria set by the gray correlation order are shown in Table 2: Item description [^ Number formula AR (buying momentum indicator) 硏 judge the funds and popularity of both long and short 26 [(Today Highest price-today's opening price) 26 days total cumulative] + [(today's opening price-today's lowest price) 26 days total total] BR (buying willingness indicator) 硏 judge the long and short sides of funds and popularity 26 [( Today's Highest Price—Yesterday's Closed Price) 26 Days Total Accumulation] + [(Yesterday's Closed Price—Today's Lowest Price) Cumulative Total in 20 Days] OSC (Oscillation Index) Uses the rate of stock price changes during the period to measure how long the market is Strength 10 (Ct + Ct-n) _0 Ct: Closing price of the day Ct_n: Closing price of n days ago 200532524 PSY (Psychological indicator) Measure the bullish or bearish mentality of stock market investors 13 (Total number of rising days within 13 days +13) X100 RSI (phase (Strength and Weakness Index) Measure whether individual stocks are overbought or oversold24 (Average total increase within 24 days 値 + Average average increase and decrease within 24 days 値) VR (Volume Strength Index) 硏 Judgment of trading volume buying and selling gas indicators 26 【 (26 days the stock price rose on the day, the sum of + 0.5x 26 days the stock price unchanged on the day of the day, the sum) + (% day the stock price fell on the day of the transaction, the sum of the stock price on the 26th day of the same day, the sum of the day, the sum of the day) x WMS% R (William's indicator) Applying the swing principle to determine whether it is overbought or oversold 9 [(Highest price within 9 days-9th day closing price) + (Highest price within 9th day and lowest price within 9 days)] xlOO K% R (William's indicator) Applying the swing principle to determine whether it is overbought or oversold 9 [(Intraday deals-Average turnover during the period x 0.5.) + (Highest turnover during the period-Lowest during the period- Transaction 値) x2] Table 1 Technical Indicators Explained Item Giant Strategy Broad Market Gain AR BR OSC PSY RSI VR WMS K% R 1 Μ Unlimited purchase 0.25 0.3 90 7.69 13 50 80 0.3 2 Buy 10% or more 1.69 4.62 107.1 83.848 85.9 320 26 3.18 3 Unlimited buy 1.85 5.1 109 92.31 94 350 20 3.5 Table 2 sets the standard

Since the operating time of the present invention is fixed at May 30, September 30, and November 30, if the poverty investment strategy is to buy, in order to avoid the relatively high cost of purchasing individual stocks on that day, a screening criterion is specifically formulated, that is, Calculate the reasonable purchase price on that day. If the stock price on that day is greater than the reasonable purchase price, you will not buy it at that time. In practice, you can buy the stock when the stock price is lower than the reasonable purchase price, and the reasonable purchase price is calculated. As follows: Reasonable purchase price = i After-tax shareholder's equity_ (single quarter)) x recent quarterly earnings per share X purchase price / earnings ratio V 100)

x PER R〇E) X NAV 、

Too '-/ In summary, the stock selection process is drawn below. The first step of the buy and sell strategy 200532524 is to use the after-tax shareholder return (accumulated) to screen. Although the purchase strategy is deleted, Since the accumulated capital is used for screening, in the second quarter (9th day), τ third quarter ⑴ tz) ♦ In operation, in principle, the counties of shareholders' after-tax rate will be doubled and doubled. It is different because of _ economy. In principle, the number of homes after the first step of screening is about 5 0 100 豕, and then the screening is carried out using the criteria generated by the decision tree. On the other hand, in the short-selling section 'due to securities lending Six days before the trading cessation date (two days after the ex-rights and ex-dividend dates), you must make up for it. Therefore, when the shorting strategy is adopted, the stocks with ex-rights and ex-dividends in the current year are excluded as the last shorted portfolio.

Analysis of the empirical results of standardized investment procedures. The market's long and short positions are judged. The empirical period of this research is from November 30, 2000 to April 8, 2005. A total of 13 77 12 200532524 operations were conducted. Table 4 shows the correlation between the R * ratio and the broad market index:

10,000 9.0008.000 7.0006.000 5.000 4.000 3.0002.000 1,0000

10 8905 8909 8911 9005 9009 9011 9105 9109 9111 9205 9209 9211 9305 9309 9311 9312 9401 9402 8402 4 6 4 2 0 1 20001130-20010530 7.19 10.17 6185.14 5256.93 Buy 5265.93 5057.07 Change 41.45 Change -15.01 Change -3.8 2000Q3 2001Q1 20001130 20010530 20010530 20010930 2 20010530-20010930 10.17 6.79 5256.93 5057.07 Change 504.86 363694. 3.8 Change 27.97 2001Q1 2002Q2 20010530 20010930 20010930 20011130 3 20010930-20011130 6.79 5.16 5048.86 3636.94 Buy 3636.94 4441.12 Change -24.01 Change -27.97 Change 22.11 2001Q2 2001Q3 20010930 20011130 20011130 20020530 5.89 3636.94 4441.12 Buy 4411.12 5675.65 Change 14.15 Change 22.11 Change 27.80 2001Q3 2002Q1 20011130 20020530 20020530 20020930 5 20020530-20020930 5.89 3.99 4441.12 5675.65 Empty 5675.65 4191.81 27.80 of Price of Price of Price -32.26 -26.14 6 20020930-20021130 2002Q1 2002Q2 20020530 20020930 buy 2,002,093,020,021,130 3.99 3.04 5675.65 4191.81 4191.81 4646.69

7H 200532524 Change -23 · 81 Change -26.14 Table 4 R wood and big change The investment of the present invention «Traditional shooting method is _ decision sickness classification, looking for a good physical and silky internal杂 杂 m_ 林 读 means that it is still a priority for Yi to find out the purchase. This study of the Wei Jing 13 times, from 51 years to 2 April, a total of 51 months is about 4.25 years, the annual return rate is about 89. _, the return on investment performance is better than the broad market. _Figure 8 shows. The contribution of miscellaneous research results is described as follows: The shareholders' equity report of the present invention _ as the first age of riding stocks, matching _ the top 30 formed by the matching of the manufacturing industry and the full delivery of the constituent stocks, The individual stocks after these heavy screenings are all companies with good health, which can be profitable, and earn cash dividends and stock dividends. -General investors and fund New Zealanders are obsessed with business affairs. Mo Mi deeply analyzes from the quarterly financial report, but the aspects of consideration are often similar. _ Suddenly do not know how to make a choice, but In the analysis of riding position, the financial ratio (losing Wei) is relatively important compared to whether it is good or not (output variable). The stock is slightly more slick than the line. It is not easy to record the details, so the author specifically addresses this situation from many overall factors (such as the growth rate of Liyue, the growth rate of leading indicators and the stock price index, or the growth rate of the GDP and the stock price index. Relationship, etc.) After a long observation, this ratio is easier to judge and interpret than other factors (较, GDP, leading indicators). This ratio is defined by the author's own definition. 'In 13 operations, there is a clear signal of short selling. , And the return on investment of the two operations is quite high. 〃 200532524 [Schematic description] The first diagram is the flow chart of the research method of the present invention. The second diagram is the basic structure of the inverted delivery network of the present invention. The third diagram is the decision tree classification table of the present invention. The fourth diagram is Attributes of the invention The training set for Patrons is the test. The fifth chart is the investment system architecture table of the present invention. The sixth chart is the top 30 of the MSCI Taiwan Index constituent stocks of the present invention. The seventh chart is the eighth cash flow index table of the present invention. Figure is the contribution table of the research results of the present invention

Claims (1)

200532524 X. Scope of patent application: 1. A method of investing and listening system architecture, which is based on the freshening process_stock offering into an investment portfolio, mainly includes the following steps: Step 1 Step 2 Step 3 Step 4 Step 5 Step 6 Section II 1. Step 1: The long and short stock price index is the leading indicator of the overall economy, and the broad market gain is a reflection of the good and bad operating conditions of all listed companies. This study uses The single quarterly quarterly report of all listed companies takes the after-tax shareholder return (R0E) greater than 8 as the screening threshold, and R * is used to measure the overall status of all listed companies in the quarter. ) χ100 Here, compare GRV⑹ with the stock price index at the time of the quarterly quarterly announcement; suppose that it is greater than 0, which means that the operating conditions of all listed companies in this period are better than the previous period, so the stock compensation 200532524 is only in that number. The period from the quarterly announcement time point to the previous quarter should show an upward trend; assuming that the mouth is less than 0 and the decline is greater than 20%, but the broader market shows a trend of consolidation or rise during the period (t— (t)). The performance of the broader market is inconsistent with the actual operating conditions of all listed companies. Therefore, in the next quarter, a shorting strategy will be adopted. 2. Step 2 · DECISION TREE data pre-processing: The decision tree algorithm of this study is binary classification. In order to reduce the grey space between financial health and imperfection, the top 30 MSCI constituent stocks are used (excluding financial (Related industries) and full-delivery stocks. Observe that the MSCI Taiwan Index constituent stocks all have equity, large market value, sound financial health, and high exclusivity in various industries. Full-delivery stocks are based on the Taiwan Stock Exchange. Article 49 of the company's business rules stipulates that if one of the prescribed circumstances occurs for a listed company, the listed securities have been approved by the Taiwan Stock Exchange to change the original trading method to a full settlement method. 3. Step 3 · Sensitivity analysis of neural network-like: The input variable of the decision tree is the financial ratio 'and the output variable is (healthy, unhealthy Wang 1 1). The attributes of its financial ratio can be divided into financial structure stability There are 105 indicators in eight categories: short-term debt repayment ability, long-term debt repayment ability, operating ability, turnover ability, productivity indicator, cash flow indicator, and market price indicator. There are 105 indicators in order to measure the importance between input variables and output variables. The variables entered into the training process of the decision tree, so sensitivity analysis was used as a method for screening financial ratios. Step 4: Use cash basis and other bases to enter the decision tree. Beaver (1966) found in the study of “financial ratio prediction of operational failure” that “the amount of cash w / total liabilities” is the best indicator for predicting operational failure. The flow indicators are independent and fixed into the scale of the human silk tree. Decisions are made, and he calls his seven major financial ratios-and advanced person sensitivity analysis _ selects, selects financial ratios greater than 丨 · 5 and greater than 2, and then Then enter the training process. V. Step 5 · Considering the return on shareholders' equity (rqe) Buffett believes that-Jiagong_ low is determined by its intrinsic value, and the stock price will be lower and higher than the value formed by the intrinsic value 200532524. It is a well-defined definition of the price. In this research, we will use _ as the first screening criteria for screening stocks, and then add the criteria formed by the decision tree to continue the screening. 6. Step 6: Sorting Grey Relations In the case of Wei in 4 quantities, after specific data processing, the sequence of mechanical and mechanical factors can find their correlation. Step 6 is all listed companies. After step 4 and step 5, screening = remaining stocks. In this step, through the selected technical analysis indicators, find out the purchase priority and t the final investment list. The selected technical indicators are the momentum indicator for buying and selling (UR), willingness to buy and sell =) Epicenter (0SC), psychological indicator (PSY), relative strength indicator (RS〇, volume strength indicator (VR), and William indicator (WMS%) R). Recognition = based on the recommendation of Li-tech asset portfolio, using decision tree classification, neural network, and gray view to analyze the financial indicators and technical analysis indicators of the enterprise. The present invention can be summarized as follows: ^ Road analysis' finding key financial principles related to sound physique and intrinsic value ^ 2. Use decision tree classification to clearly find the decision attributes of stocks. 3. Use gray-level stocks to show the priority order of buying and selling. 4. Build-a Standardized investment procedures for investors