High-frequency Financial Time Series Prediction And Anomaly Detection Based On LSTM And Autoencoder

Recently,with the crazy price fluctuations of Bitcoin,extensive attention and discussion have been attached to cryptocurrencies around the world.Cryptocurrency enriches the form of high-frequency financial time series(HF-FTS)and promotes the development of highfrequency quantitative trading which is an interdisciplinary field of computer science and finance.Researches on HF-FTS will benefit the participants in profiting in the market and the monitors in supervising the market.Correspondingly,forecasting and anomaly detection of HFFTS have become important parts of the researches.Financial time series are usually accompanied by chaos,high noises,nonlinearity,and nonstationary.HF-FTS magnifies these characteristics at the micro-level.Thus,it is challenging to research HF-FTS accompanied by a phenomenon of the “lost”.This paper studies the forecasting and anomaly detection of HF-FTS.The contributions are as follows:(1)This paper describes the used Ethereum high-frequency time series multi-feature dataset and discusses the feature scaling.Pre-diff Simi-standardization is proposed in response to the “lost” in the regression prediction which is after common scaling methods.Training the neural network with Pre-diff simi-standardized HF-FTS features can alleviate the "lost" of the prediction curve.Targeted comparative experiments showed that the Pre-diff Simistandardization method has certain advantages when the potential "lost" occurs.(2)This paper builds an HF-FTS multi-classification prediction model which is based on Bidirectional Long Short-Term Memory network,Attention mechanism,and Dynamic ProximValidation Learning Rate mechanism(DPVLR).Based on Long Short-Term Memory network,the Bidirectional mechanism strengthens the extraction of microscopic features of highfrequency sequences.The Attention mechanism compensates for the attenuation of feature information.The DPVLR is proposed to enhance the ability of the neural network to fit the data segments closed to the future in time and thus enhance the robustness of the model.This paper conducted a simulated real offer experiment based on the built model.The results of comparative experiments proved that the model this paper built performs well on the task of HF-FTS multi-class forecasting.The results of ablation experiments showed the effectiveness of the model mechanism used or proposed in this paper.(3)This paper builds two HF-FTS anomaly detection models which are based on autoencoders.The anomaly detection model based on Deep Autoencoder is built in order to detect the occurrence of the potential "lost" phenomenon caused by abnormal HF-FTS data.The anomaly detection model based on Variational Autoencoder is built in order to optimize the strategy of HF-FTS multi-classification prediction.This paper achieved experimental verification combined with the results of multi-classification prediction.It was verified that the models this paper built could identify the “lost” anomalies and optimize the prediction results.