Stock Price Prediction By HMM Base On Bayesian Inference

Stock price prediction has been an important and difficult topic researching for many years.Comparing with foreign markets,stock price is more complicate and has high variance in China market.This thesis is to suggest a new compound model BHMM(which is HMM based on Bayesian inference)to make prediction.Comparing with traditional HMM,BHMM provides stable price prediction and makes HMM pattern prediction more accurate.BHMM makes improvement in three aspects.Firstly,BHMM splits raw data into pieces with different lengths by using Bayesian inference.The traditional way of HMM modeling uses fixed-length window to slide over train data which is prone to learn more noise while BHMM just providing data with high split probability for HMM to fit.Hence BHMM is more likely to learn patterns with less noise.Secondly,BHMM makes trend prediction by mixed windows.BHMM uses many windows to slide over train data,and calculates the mean percentage variance for each most proper HMM.When predicting,BHMM combines results from different windows for the final result which could achieve higher coverage and accuracy.Finally,BHMM combines Bayesian point estimation and loss function to improve the price prediction by HMM.It is used to make prediction by the next value of best-fitted HMM with problem of high error rate,so BHMM makes assumption that current data conforms Normal distribution as prior distribution which parameters determined by itself and HMM's probability density function as likelihood function so as to estimate the posterior distribution as predictions which could be adjusted by loss function.Comparing with ARIMA,LSTM and traditional HMM,BHMM is more stable with less MAPE and higher accuracy.At last,an experiment is took to test its performance through ten China market index including SSE and it achieves 53.5% of accuracy with ROI of 16.4% on average which proves the stability and feasibility of BHMM further.