Application Of Hidden Markov Model To Futures Market

Compared with other models,the hidden Markov model has the advantages of distinguishing market conditions and explaining price trends.Two classical prediction methods are included.The first method is to estimate model parameters using BW algorithm,and select the state with the largest probability value for prediction through the implicit state transition matrix.The second is to look for the day most similar to the likelihood value of the current hidden state in the historical data according to the Viterbi algorithm as a similar day,and forecast through the price changes of similar days.In this thesis,the hidden Markov model is introduced into the Chinese futures market,and the quantitative timing strategy is constructed according to the two prediction methods.Ten futures varieties are selected to study the observation sequence and parameter setting of the model.Strategy backtest was carried out for different observation series under one and multi-dimensional conditions,with different data quantities and different hidden state numbers respectively.According to the prediction victory rate of the backtest results and the prediction accuracy and stability of the maximum strategy backtest evaluation model,the dominant observation series and parameter selection rule under different methods were obtained.PTA futures were selected to further optimize the model by applying the obtained rules,BIC criterion was used to determine the number of hidden states,and the multidimensional observation series was treated with principal component analysis to reduce the correlation between data,improve the model prediction effect and enhance the strategy income ability.The results show that under the two forecasting methods,the forecasting effect of the discrete closing price rise and fall and the logarithmic rise and fall series is significantly better than the original closing price series after discretization;the prediction effect is better when the implicit state transition matrix is used to predict with a small amount of data,while the prediction effect is better when the historical data is used to find similar days;the model using BIC criterion to determine the number of hidden states has a good prediction effect,and the strategy return is much higher than the benchmark return;continuous data after dimensionality reduction of principal component analysis should meet the mixed normal distribution.After removing the sequence that does not meet the normal distribution,the prediction accuracy of the model is better,the strategy returns are higher,and the maximum retracement is smaller.