Zehong Yang

Short-term stock price prediction based on echo state networks

Neural network has been popular in time series prediction in financial areas because of their advantages in handling nonlinear systems. This paper presents a study of using a novel recurrent neural network-echo state network (ESN) to predict the next closing price in stock markets. The Hurst exponent is applied to adaptively determine initial transient and choose sub-series with greatest predictability during training. The experiment results on nearly all stocks of S&P 500 demonstrate that ESN outperforms other conventional neural networks in most cases. Experiments also indicate that if we include principle component analysis (PCA) to filter noise in data pretreatment and choose appropriate parameters, we can effectively prevent coarse prediction performance. But in most cases PCA improves the prediction accuracy only a little.

Automatic stock decision support system based on box theory and SVM algorithm

The stock market is considered as a high complex and dynamic system with noisy, non-stationary and chaotic data series. So it is widely acknowledged that stock price series modeling and forecasting is a challenging work. A significant amount of work has been done in this field, and in them, soft computing techniques have showed good performance. Generally most of these works can be divided into two categories. One is to predict the future trend or price; another is to construct decision support system which can give certain buy/sell signals. In this paper, we propose a new intelligent trading system based on oscillation box prediction by combining stock box theory and support vector machine algorithm. The box theory believes a successful stock buying/selling generally occurs when the price effectively breaks out the original oscillation box into another new box. In the system, two SVM estimators are first utilized to make forecasts of the upper bound and lower bound of the price oscillation box. Then a trading strategy based on the two bound forecasts is constructed to make trading decisions. In the experiment, we test the system on different stock movement patterns, i.e. bull, bear and fluctuant market, and investigate the training of the system and the choice of the time span of the price box. The experiments on 442 S&P500 components show a promising performance is achieved and the system dramatically outperforms buy-and-hold strategy.

Recognition of gray character using gabor filters

In this paper, a novel Gabor filter-based feature extraction method for low resolution gray character classification is proposed By applying Gabor filters to a character image, dominant orientation matrix is obtained and used to form feature vector for recognition. We compared our Gabor feature with other Gabor features. Experiments show that the proposed feature extraction method achieves high recognition accuracy and is also not sensitive to noise and other distortions. This method has been used in a Vehicle License Plate Character System.

A novel hybrid feature selection algorithm: using ReliefF estimation for GA-Wrapper search

A new feature selection method named ReliefF-GA-Wrapper is proposed to combine the advantages of filter and wrapper. In the ReliefF-GA-Wrapper method, the original features are evaluated by the ReliefF method, and the resulting estimation is embedded into the genetic algorithm applied to search optimal feature subset with the train accuracy of induction learning algorithm for the evaluation function. Experiments are carried on handwritten Chinese characters dataset, which is a large-scale dataset, and several other typical datasets with features more than 20. The results show ReliefF-GA-Wrapper has better performance then ReliefF and GA-Wrapper, indicating that the proposed ReliefF-GA-Wrapper algorithm is competitive and scales well to large datasets.

Self-adaptive design of hidden Markov models

Hidden Markov models (HMMs) are stochastic models widely used in speech and image processing in recent years. The number of states in a classical HMMs is usually predefined and fixed during training, and may be quite different from the real number of hidden states of the signal source. Moreover, in pattern recognition applications, different signal sources probably have different state numbers, thereby cannot be well modeled by HMMs with a fixed state number. This paper proposes a self-adaptive design method of HMMs to overcome this limitation. According to this design, an HMM automatically matches its state number to the real state number of the signal source being modeled. To realize a practicable training algorithm for the new HMM, this paper first introduces an entropic definition of the a priori probability of the model and accordingly a maximum a posteriori (MAP) training strategy, and then designs an MAP training algorithm in the case of fixed state number based on the deterministic annealing (DA) technique. Based on this MAP training, a complete training method named shrink algorithm is finally proposed for the new HMM. Experimental results indicate that self-adaptive HMMs can model stochastic signals more accurately and have better performance in pattern recognition than classical models.

Feature selection in recognition of handwritten Chinese characters

Recognition of handwritten Chinese characters is a large-scale pattern recognition task, which is difficult and time consuming to build the corresponding classifiers. In this paper, two feature selection methods are proposed to reduce the complexity and speed up the handwritten Chinese recognition: one is the ReliefF-Wrapper method which evaluates the original features with the ReliefF method, and then uses the wrapper method to decide the number of features to be selected; and the other is GA-Wrapper that uses genetic algorithm to search the optimal subset of features with high training accuracy. Experiments were performed on 800 most frequently used Chinese characters, with 80,000 handwritten samples. Results show that the ReliefF-Wrapper method has good interpretation and high speed and GA-Wrapper gains higher accuracy. Limitations of the both methods and future work are also discussed.

The application of echo state network in stock data mining

Stock data, which is among the most complicated time series, is difficult to analyze and mine. Neural network has been a popular method for data mining in financial area since last decade. In this paper, we explore the use of Echo State Networks (ESNs) to perform time-series mining on stock markets. The Hurst exponent is applied to adaptively determine initial transient and choose sub-series with greatest predictability before training. With the capability of short-term memory provided by ESN, a stock prediction system is built to forecast the close price of the next trading day based on history prices and technical indicators. The experiment results on S&P 500 data set suggest that ESN outperforms other conventional neural networks in most cases and is a suitable and effective way for stock price mining.

A new approach for off-line handwritten Chinese character recognition using self-adaptive HMM

Research on off-line handwritten Chinese character recognition is an important, yet difficult field of artificial intelligence and pattern recognition. In this paper, a kind of pseudo 2D HMM is established to solve the problem. This HMM has a special topology and is a powerful tool for modeling the Chinese handwriting. Furthermore, a self-adaptive design method is applied to this HMM and proved to be able to bring better performance to the HMM classifier. Finally, recognition experiments show that the proposed HMM has a high correct rate of 95.9%.

Adaptive parameter tuning for relevance feedback of information retrieval

Relevance feedback is an effective way to improve the performance of an information retrieval system. In practice, the parameters for feedback were usually determined manually without the consideration of the quality of the query. We propose a new concept (adaptiveness) to measure the quality of the query. We built two models to predict the adaptiveness of the query. The parameters for feedback were then determined by the quality of the query. Our experiments on TREC data showed that the performance was improved significantly when compared with blind relevance feedback.

Design and implementation of educational platform in RoboCup simulation games

RoboCup is an international game and academic activity which focuses on improving the education and research of distributed AI, intelligent robotics, machine learning and other related fields. To stimulate the students interests of AI research and introduce the RoboCup games to more students, we transplant the RoboCup to an educational platform for students to study and research. This paper shows the structure and design of the whole platform, the specific implementation of basic modules on the educational platform. The platform is divided into three parts: low-level module, libraries, and program scheme. Students should work in the programming scheme part, build the basic skill and top-level strategy layer. This paper describes the main theory of the basic skills such as dribble, kick, and shoot, including the distributed planning in realizing these skills. The top-level strategy part analyzes the realization of several basic strategies and machine learnings impact on the game. We also provide the method and framework for implementing the basic skills and top-level strategy, called the qsinghuAeolus program. Finally, we show the educational value of the platform in RoboCup simulation games.