When applying artificial intelligence technology to quantitative trading, high noise and unpredictability of market environment are the first practical problems to be considered. Therefore, how to select the learning features of the market based on rapidly changing financial data is particularly important. In this paper, the real time financial data are first processed by K-line theory, which uses candlesticks as a generalization of price movements over a period of time, so this process can play the role of de-noising. Then, the candlesticks are decomposed into different subparts by mean of a specified spatio-temporal relationship, based on which cluster analysis of the subparts to get the learning features. Further, the learning features that are clustered by the above K-lines are put into the model, and the online adaptive control of the parameters in the unknown environment is realized by the deep reinforcement learning method, so as to realize the high frequency transaction strategy. In order to verify the performance of the model, the data on different financial derivatives transactions such as stocks, financial futures and commodity futures are used. The proposal approach is compared with other methods which are based on price, fuzzified price and K-lines for features learning. In order to verify the accuracy of the proposal approach, prediction-based methods such as recurrent neural network and fuzzy neural network are used for comparison. Experimental results show that the proposed method has higher robustness and prediction accuracy. INDEX TERMS Deep reinforcement learning, K-line decomposing, clustering, trading system.

No abstract

Most of time series deriving from complex systems in real life is non-stationary, where the data distribution would be influenced by various internal/external factors such that the contexts are persistently changing. Therefore, the concept drift detection of time series has practical significance. In this paper, a novel method called online entropy-based time domain feature extraction (ETFE) for concept drift detection is proposed. Firstly, the empirical mode decomposition based on extrema symmetric extension is used to decompose time series, where features in various time scales can be adaptively extracted. Meanwhile, the end point effect caused by traditional empirical mode decomposition can be avoided. Secondly, by using the entropy calculation, the time-domain features are coarse-grained to quantify the structure and complexity of the time series, among which six kinds of entropy are used for discussion. Finally, a statistical process control method based on generalized likelihood ratio is used to monitor the change of the entropy, which can effectively track the mean and amplitude of the time series. Therefore, the early alarm of concept drift can be given. Synthetic data sets and neonatal electroencephalogram (EEG) recordings with seizures annotations data sets are used to validate the effectiveness and accuracy of the proposed method.

No abstract

No abstract