Trend Prediction And Algorithmic Trading Of Chinese Stock Index Based On The Time-CNN

With the introduction of online trading platforms and the development of computer technology,algorithmic trading strategies based on trend forecasting are increasingly favoured by investors.However,due to the complexity and dynamics of financial data,it is difficult to manually design new features and achieve good prediction accuracy using traditional prediction methods.Therefore,because neural networks can automatically extract deep features and have stronger fitting capabilities,this paper uses neural networks to predict the trend of stock indexes and proposes a close-knit "neural network+ algorithmic trading" framework.The framework directly uses the prediction results as trading signals to generate timing strategies with autonomous intelligent decisionmaking.Taking stock indexes in the Chinese market as examples,this paper studies the framework from four aspects: data processing,neural network models,algorithmic trading strategies,and model evaluation measures.In terms of data processing,aiming at the problem of how to define the trend and choose the prediction step,this paper designs a back trade experiment to select the one that maximizes the strategy’s return among a variety of labels as the final prediction target.On the model side,the Time Convolutional Neural Network(Time-CNN)is used to predict stock index trends.The comparative experimental results with the other six neural network models show that the Time-CNN model has the strongest predictive ability.In terms of algorithmic trading strategies,our framework yields seven neural network timing strategies,five of which have higher returns than baseline strategies such as "Buy&Hold".The TimeCNN strategy has the highest return,and it also has a good anti-risk ability from the perspective of Sharpe ratio and maximum drawdown,which can provide investors with a certain reference.Model evaluation measures can evaluate the model’s performance and help researchers select the optimal model.However,the traditional measures(accuracy and F1 score)have shortcomings in reflecting the "revenue capability of the strategy corresponding to the model".This paper proposes two new model evaluation measures: Inertia Weighted Accuracy(IWACC)and Percentage-change Weighted Accuracy(PWACC).The new measures have a higher correlation with the returns of the corresponding strategies,which can better help investors to choose models with higher returns.The empirical results show that the new measures can increase the average return from 3.37% to 7.95%.