Research On Asymptotic Statistical Algorithms Of Key Events In Financial Time Series

It is an effective way to achieve high returns and low risk in financial transactions by key trading events for buying or selling with low frequency and reliable prediction.The past value of the financial sequence has a direct or indirect effect on the future value,which leads to the context dependency of the key transaction events that the trading pivot makes sense only in the corresponding context subsequence.In this paper,different features are extracted from the price RSI sequence and the volume RSI sequence to form the auxiliary sequence R.Based on the auxiliary sequence R,the context subsequence of RB and RT and the key trading points therein are defined,and the corresponding segmentation algorithm is designed to automatically segment the RB and RT context subsequence and the corresponding trading key points.The volatility and noise in financial markets lead to random key trading points,where the randomness is to be studied to make reliable predictions.In light of measure theory,this paper proposes an event mapping model to formally define the randomness,expresses the transaction points in the sequence in a low and normalized metric space Q by mapping functions and transforms the prediction of key transaction points into the prediction of key events in metric space Q.According to measure theory,the probability of the occurrence of a critical event can be approximated from the probability of random events nested in Q when the mapping function is monotonic.Due to the sparsity of key transaction events,this paper designs an LSTM-based neural network model to learn this event mapping function.During the training process,different data blocks divided from the obtained context subsequence form a new training set with the corresponding transaction key points to solve the imbalance caused by the sparsity.Based on the monotonic convergence theorem of measure theory,multiple random sample points are used to estimate key events during the prediction process,which can be considered as a method between oversampling and data enhancement,and the trading decision is finally made according to the statistical convergence properties predicted.This paper refers to this method as the asymptotic statistical learning algorithm(ASL)and proposes an improved version based on the upper and lower bounds of asymptotic statistical learning algorithm(UL-ASL)which introduces the concept of upper and lower boundary approximation that respectively approaching the point from the upper and lower boundaries of the price and determining the key trading point simultaneously.This makes the decision-making process more accurate and robust.The experiment shows that algorithms ASL and UL-ASL,compared with twelve existing trading algorithms,can predict rare key trading events from a large number of random events based on the six real data sets including selected underlying stocks,the S&P 500 index and cryptocurrencies.The algorithms ASL and UL-ASL outperformed the other algorithms in different market conditions by examining the three indicators of return(RoR),Sharp Rate(SR)and Maximum Retraction Rate(MDD)respectively.