This is the continuation of part V of Machine Learning in Capital Markets.
The determined values can be converted into a uniform interval by scaling. The interval can range from -100 to +100 units. It is only important that all values are scaled so that the relative size is not changed. In contrast to the RSI method, individual outliers are not eliminated, only the size of the individual values is adjusted. The adjusted values can then be used as input for individual classifiers. Furthermore, with the univariate scaling, individual characteristics can be compared with the overall unit. In the already mentioned example with the loan approval, the relationship between “good” and “bad” customers could be illustrated by scaling. To get the result, these must be put in relation. In doing so, the expression or the probability that the individual characteristics belong to the good or bad customers must be taken into account. In order to obtain the scale value for the “good” customers, the number of characteristics of the good customers must be divided by the number of characteristics of the good and bad customers.
The primary task of feature selection is to create a certain balance between the inventory of incremental features and an increase in efficiency when processing the training data by removing redundant features. Due to the high probability of randomness and the non-linear nature of the financial market-specific data, it is extremely difficult to carry out an efficient way of selecting characteristics. Within the Thom field trial, the performance of correlation and regression-based filter methods and wrapper-based selection methods were tested. A wrapper is a program for extracting the content of information. For example, wrappers are used to extract the relevant information such as price or speed of a notebook from supply pages. These are then used to compare offers. A learning problem is a regression if y is a number such as tomorrow’s weather. In this case, y is the positive result of a hypothesis regarding the prediction of new implemented sample data. Realistically speaking, in the case of a regression problem, one is only looking for an expected or average value, since the probability of finding the exact value for y is in the range 0. In contrast to the regression-based filter methods, the wrapper-based selection methods such as the “floating” or “forward” search method were able to determine and select new usable subsets of characteristics on the basis of a k-nearest-neighbor classification. However, the result of the classification could not be improved with these additional features, but could only be kept within a constant range. The ratio of the increased computing time is therefore not positively related to the training and classification time. Based on these facts, the use of filters and wrappers has been dispensed with in the next analysis processes.
Classification of objects in classes
The basic assumption in the classification of objects is the distinction of the studied objects in rising and falling courses. Furthermore, the basic assumption applies that rising and falling prices exist with the same frequency. These assumptions are risky in that there is a possibility that actual price movements will not be taken into account. All price drops are assigned to class zero, while all rising price movements belong to class one. This assignment remains valid regardless of the amount of the course change. This should take into account course changes of 0.0001%. In such cases, the real change is so low that it can be neglected and is considered “unchanged” in practice. In order to intercept these minimal changes, a third class is used. The introduction of a third class minimizes the number of assigned objects and increases the complexity of the classification process. A “vertical” delimitation of the course of the course is made possible by the fact that all course changes which take place below 10% of a defined standard deviation are assigned to the third class. It is irrelevant whether said minimum course deviation is defined positively or negatively. This delimitation is defined as an evaluation-free zone that all course changes which take place below 10% of a defined standard deviation are assigned to the third class. It is irrelevant whether said minimum course deviation is defined positively or negatively. This delimitation is defined as an evaluation-free zone that all course changes which take place below 10% of a defined standard deviation are assigned to the third class. It is irrelevant whether said minimum course deviation is defined positively or negatively. This delimitation is defined as an evaluation-free zone.
After implementing this method, it can be determined that this does not increase the classification quality. The quality of the classification is increased as soon as the classification time is defined at the period high of a period under consideration. This is justified by the fact that the period high is a period with an actual upward movement. The period high is naturally reached before the end of the period under review. The period high must be defined in terms of time and value. The period duration is also defined for this. 15-minute periods are studied in practice. If the maximum price within the defined period is over 0.15% of the closing value for the period, this is defined as the maximum period.
Stock market data are naturally subject to uncertainties and randomness. In connection with the binary selection of the classification, it is conceivable that even very small price changes are interpreted as a recommendation to sell or buy shares. An incorrect classification in class zero would not initiate an action, contrary to an incorrect classification in class one. It follows that primarily the number of correct class one classifications should be maximized relative to the incorrect class one classifications. It is therefore of secondary importance to minimize the overall misclassification. When dividing into classes zero and one, both priorities are treated identically, which requires that a new prioritization takes place. The prioritization is achieved by adjusting the limit values. The limit for classification in class one is increased, which means that the number of objects in class one is minimized. Vice versa maximizes the number of objects in class zero. Objects that actually describe a positive change of course are assigned to class zero, because the course change is too low despite the positive trend. Despite this, the informative value of class one is increased overall assigned to class zero, since the course change is too low despite a positive trend. Despite this, the informative value of class one is increased overall assigned to class zero, since the course change is too low despite a positive trend. Despite this, the informative value of class one is increased overall.
The aim of the classification is to look for recurring feature patterns in order to be able to react better to new situations. It is assumed that there are a number of objects that can each be assigned to a class. In the classification, an attempt is made to create a classifier that assigns new or unknown objects to the correct class. In relation to a stock price forecast, an attempt is made to better characterize the price trend by using characteristic patterns.
If a classifier is adjusted several times to improve the classification quality, the less likely it is that the result will be reproducible when classifying unknown objects. By validating the classification algorithm on the unseen data, knowledge about its generalization ability can be gained. Furthermore, the choice of methodology for data evaluation is very important. Neural networks offer the advantage of being able to recognize nonlinear relationships well. However, there is a higher risk of overfitting or underfitting in neural networks. If the number of neurons is too large in relation to the individual layers, it is possible that the overfitting effect will occur and the generalization ability on unseen data will be inhibited. For this reason, there is a conflict of goals between adaptation and generalization ability of the neural network. In contrast to the neural networks, the creation of support vector is less dependent on the choice of parameters.
To be continued in next part.