The stock market is an important domain in which the investors are focused to, therefore accurate prediction of stock market trends remains a hot research area among business-people and researchers. Because of the non-stationary features of the stock market, the stock price prediction is considered a challenging task and is affected by several factors. Anticipating stock market trends is a difficult endeavor that requires a lot of attention, because correctly predicting stock prices can lead to significant rewards if the right judgments are made. Due to non-stationary, noisy, and chaotic data, stock market prediction is a huge difficulty, and as a result, investors find it difficult to invest their money in order to make a profit. In order to predict stock market movements, a number of strategies have been established. Earlier studies based on statistical models and machine learning techniques have been focused on short term stock price prediction. In this aspect, this study designs a novel hyperparameter tuned hybrid convolutional neural network with long short term memory (HPT-HCLSTM) for stock price prediction. The proposed HPT-HCLSTM technique encompasses three different processes namely pre-processing, prediction, and parameter optimization. The HPT-HCLSTM technique employs the HCLSTM technique for the prediction of stock prices. In addition, teaching and learning based optimization (TLBO) algorithm is applied for the hyperparameter optimization of the HCLSTM technique and thereby results in minimal error values. In order to demonstrate the enhanced prediction performance of the HPT-HCLSTM technique, a wide range of simulations were carried out and the results highlighted the better performance of the HPT-HCLSTM technique under several aspects. The HPT-HCLSTM technique is found to be a proper tool for forecasting stock prices. the HPT-HCLSTM technique has showcased better performance with the increased R2 value of 0.9154.

The stock market is a platform where the stocks are traded, circulated, and transferred. It has a record of four hundred years and it is utilized as channel to companies for increasing funds [

Continuous growth in the AI fields results in extensive utilization of DL algorithms in practical scenarios and many research fields [

Several areas have tested the precision of DL methods for predictive performance, namely gene analysis image classification and. Experimental outcomes are also attained for time-series prediction and data analyses with a DL method; e.g., DL method is utilized for predicting offline store traffics [

This paper presents an effective hyperparameter tuned hybrid convolutional neural network with long short term memory (HPT-HCLSTM) for stock price prediction. The proposed HPT-HCLSTM technique involves the design of effective DL model using HCLSTM for the prediction of stock prices. Moreover, teaching and learning based optimization (TLBO) algorithm is applied for the hyperparameter optimization of the HCLSTM technique and thereby results in minimal error values. Higher order neural networks pay more attention than traditional neural networks because they have greater computational capabilities, as well as better learning and storage capacity. This work represents a novel attempt to effectively optimise the performance of a higher order neural network (specifically, the Pi-Sigma neural network) for classification purposes. The neural network was efficiently trained using a newly developed population-based teaching learning-based optimization algorithm. The design of TLBO model for hyperparameter tuning of the HCLSTM technique shows the novelty of the work. For examining the improved performance of the HPT-HCLSTM technique, a comprehensive simulation analysis is made and the results are inspected under varying aspects.

Mehtab et al. [

In Vijh et al. [

Zhang et al. [

Abe et al. [

In Yu et al. [

In this study, a new HPT-HCLSTM technique is derived to forecast the stock prices accurately. The design of HPT-HCLSTM technique involves three stages of operations namely pre-processing, HCLSTM based prediction, and TLBO based hyperparameter optimization.

Primarily, the data collects in the business fields endure pre-processed to the change of input financial data as to helpful data by 3 subprocesses such as data transformation, class labeling, and min-max normalization. It mostly uses for enhancing the quality of business data. Afterward, the data transformation procedure contains the translation of categorical as to numerical value. Then, the data sample has been chosen for appropriating class labels from the class labeling procedure. Finally, the min-max normalization technique attains finalization for regularizing the data to unchanging level.

Once the input stock market data is pre-processed, the next stage lies in the design of HCLSTM based prediction process.

The LSTM resolves the issue of gradient disappearance of RNN. The LSTM has the ability for striking out or improve data to cell state. This ability has been provided as the framework named as gate. An LSTM takes 3 gates such as input, forget, and output gates that are utilized for providing read, write and reset functions correspondingly.

Amongst them, _{t−1} implies the cell state in the earlier module, _{t−1} refers the outcome of preceding component, _{t} stands for the present input, utilized for generating novel memory, and the resultant data contains the cell state _{t} transferred later, novel output _{t}.

The forgetting gate from LSTM has been really a valve. If the input gate has been continuously open, much data flood as to the memory. At present, a forgetting process requires that more for removing the data from the memory [_{t−1} (earlier output) and X_{t} (existing input) and output the number amongst 0 with 1 to all digits from the cell state C_{t−1} (preceding state). 1 demonstrates the entirely kept, and 0 indicates the completely removed. The computation equation as:

Amongst them, W_{f} signifies the weight matrix, b_{f} implies the bias term, and the outcome F with this network is number from the range 0 and 1, representing the probabilities of preceding cell state being forgotten, and 1 refers as “Entirely reserved” 0 represented as “entirely discarded”.

An input gate in LSTM needs to enhance the state-of-the-art memory in the present input then circulate NN “forgets” part of earlier state. The input gate contains 2 parts. The primary part, a sigmoid layer called as input threshold layer” chooses that values are required to renew. The second part, a tanh layer, makes a novel candidate vector

Amongst them, W_{n} implies the weight matrix, b_{n} stands for the bias items, W_{M} indicates the weight matrix to update the state of unit, b_{m} denotes the bias item to upgrading the state of unit, and C_{t} demonstrates the state of upgraded memory units. In _{t} and _{t} gets the dot product with C_{t−1} for deciding that for retaining the original state of time-step memory units.

The resultant gate from LSTM has been outcome of present moment which requires that created afterward computing a novel state that has been utilized for controlling greatly the state of memory units from this layer has been removed. The resultant gate defines the output at that moment based on newest state, the outcome at last moment, and the present input. Its computation equation has been followed:

Specifically, primarily utilize the sigmoid activation function for obtaining O_{t} with value from the interval 0 and 1, and afterward, increase the memory cell state C_{t} with the tanh activation function and afterward multiply it with O_{t} that has been the output of this layer. d_{t} could not only be compared with the input x_{t} in the timestep t and the activation value d_{t−1} of hidden layer from the preceding timestep, apart from compared with the memory unit state C_{t} in the present time step.

The LSTM network is increased time features and procedure data with sequential characteristics [

The data has been pre-processed initially, and the managed data has been utilized for training the CNN as well as LSTM networks correspondingly. Afterward, the feature data removed with CNN and the feature data removed with LSTM are correspondingly managed as to the similar dimensional with mapping layers, and the outcomes of CNN as well as LSTM have been linked from parallel with concatenating, and lastly classified by softmax.

For properly tuning the hyperparameters involved in the HCLSTM model, the TLBO algorithm is applied to it. TLBO algorithm is based on the classical technique of education from classrooms contains 2 important parts. A primary part has been compared with choosing an optimum solution (teacher) and sharing skills amongst the teacher and another solution from the teacher phase. The second one defines the learner procedure for finding optimum solution with utilizing knowledge amongst specific candidates and by chance chosen one. These 2 parts role important plays from the TLBO technique. Combine of intensification as well as diversification have been guaranteed from the teacher and learner stages correspondingly.

The optimize technique of TLBO begins with the group of arbitrary populations known as students. Afterward the initial estimation of solutions, an optimum solution defines the teacher. The skill has been shared amongst the teacher and another solution from the teacher stage. The recently upgraded solutions have been estimated, and the greatest optimum solution is elected and exchanged with the old ones [

During the TLBO technique, a primary candidate solution is regarded as class with nS student. The group of randomizing students (S) has been processed with subsequent formula:

Initially, the students have been estimated, and its equivalent penalization objective function vector (PFit) has been produced. Afterward, an optimum student (the student with optimum-penalization objective function values) was elected as teacher (T). A step size upgrades the student nearby its teacher. The step size has been attained dependent upon the teacher skill and the average skill of every student (AveS). The teacher stage has been expressed as follows:

The term of srepsize_{i} refers the step size of ith students, newS signifies the vector of novel student, rand_{i,j} defines the arbitrary number elected in the interval of 0 and 1 and the teacher factor (TF_{i}) has been regarded as for changing the result of teacher skill on the class average that is also 1 or 2.

The value of TF_{i} is not provided as input to this technique, and this technique arbitrarily chooses their value. The demonstrated schematic generation of novel solutions from the teacher stage of TLBO projected exposes that the feasible area of novel solutions is probable amongst 2 vectors of present solution (S) and randomized step size [

During this step, recently created students have been estimated and exchanged with its equivalent old ones from the easy greedy approach. In this manner, the recently created student with an optimum-penalized objective function has been chosen to equivalent old one. So, a novel class with nS students has been designed.

During the learner phase, primarily, all the students arbitrarily elect another one (Srs) other than himself. Afterward, the student shares his skill with arbitrarily chosen one. The student nears the another chosen student when the other elected one has further skill than him (PFit_{i} < PFit_{rs}). The learner phase has been expressed as:

Step five (replacement manner): The replacement approach was carried out again.

Step six (end condition): When this technique’s end condition has been fulfilled, this technique was ended. Else, go to step 2.

For experimental validation, the Shanghai Composite Index (000, 001) stock is chosen. The day-to-day trading data of 7083 trading days in the duration of July 1, 1991 to June 30, 2020 are attained. Every instance comprises 8 attributes namely opening price, highest price, lowest price, closing price, volume, turnover, ups and downs, and change. Besides, the results are inspected interms of, mean absolute error (MAE), root mean square error (RMSE), and R-square (R2).

Folds | MAE | RMSE | R2 |
---|---|---|---|

Fold-1 | 20.425 | 30.530 | 0.9123 |

Fold-2 | 20.517 | 30.360 | 0.9046 |

Fold-3 | 20.203 | 30.124 | 0.8963 |

Fold-4 | 20.235 | 30.185 | 0.8950 |

Fold-5 | 20.586 | 30.596 | 0.9223 |

Fold-6 | 20.176 | 30.455 | 0.9404 |

Fold-7 | 20.264 | 30.535 | 0.9059 |

Fold-8 | 20.449 | 30.421 | 0.9157 |

Fold-9 | 20.467 | 30.486 | 0.9266 |

Fold-10 | 20.552 | 30.201 | 0.9348 |

The R2 analysis of the HPT-HCLSTM technique under different folds is provided in

In order to showcase the better performance of the HPT-HCLSTM technique, a comprehensive comparative analysis is made in

Method | MAE | RMSE | R2 |
---|---|---|---|

MLP | 31.392 | 39.156 | 0.8659 |

CNN | 25.561 | 36.774 | 0.8695 |

RNN | 26.718 | 35.697 | 0.8711 |

LSTM | 24.257 | 34.227 | 0.873 |

BiLSTM | 23.305 | 33.475 | 0.874 |

CNN-LSTM | 23.091 | 32.536 | 0.8752 |

CNN-BiLSTM | 22.611 | 31.961 | 0.876 |

BiLSTM-AM | 22.233 | 31.851 | 0.8761 |

CNN-BiLSTM-AM | 21.848 | 31.59 | 0.8764 |

HPT-HCLSTM | 20.342 | 30.336 | 0.9154 |

The MAE analysis of the HPT-HCLSTM technique with existing ones takes place in

Finally, the comparative R2 analysis of the HPT-HCLSTM technique with existing methods is provided in

In this study, a new HPT-HCLSTM technique is derived to forecast the stock prices accurately. The design of HPT-HCLSTM technique involves three stages of operations namely pre-processing, HCLSTM based prediction, and TLBO based hyperparameter optimization. In addition, the TLBO algorithm effectually adjusts the hyperparameters involved in the HCLSTM model and it results in improved prediction outcomes. For examining the improved performance of the HPT-HCLSTM technique, a comprehensive simulation analysis is made and the results are inspected under varying aspects. The resultant outcome ensured the better performance of the HPT-HCLSTM technique over the other techniques in terms of several performance measures. In future, the feature selection techniques can be incorporated to choose the proper influencing factors for enhanced prediction outcomes. Furthermore, the RMSE of the BiLSTM-AM and CNN-BiLSTM-AM techniques was lowered to 31.851 and 31.59, respectively, by using the BiLSTM-AM and CNN-BiLSTM-AM approaches. Finally, with a minimal RMSE of 30.336, the proposed HPT-HCLSTM method produced better results. Features selection strategies may be used in the future to select the most appropriate influencing elements, resulting in better predictions.