An asymptotic statistical learning algorithm for prediction of key trading events

In financial trading, trading at key time points (i.e., the best times for buying or selling in specific contexts) is one of the most effective trading methods. Since key points are random and sparse, their randomness needs to be investigated to predict key points reliably. Based on measure theory, we propose an event mapping model to formally define this randomness and express it in a low dimensional and normalized space Q through mapping functions. Interestingly, when mapping functions are monotonic, the probability of a key event can be approximated by those of nested random events in Q. Based on this finding, we designed a long short-term memory based neural network to learn the mapping functions and derived an asymptotic statistical learning (ASL) algorithm. ASL automatically analyzes the convergence of random events in space Q and makes point estimates on key events in trading. Compared with 12 existing algorithms on six real datasets from NYSE, S&P 500, NASDAQ, cryptocurrency markets, etc., the algorithm ASL reliably predicts sparse key events from many random events, which provides a method to deal with the imbalance classification problem in predicting key events in trading. ASL significantly outperforms other algorithms under different market situations in both bull and bear markets.