Stock price forecasting in Indonesia stock exchange using deep learning: a comparative study

In 2022, the Indonesia Stock Exchange (IDX) listed 825 companies, making it challenging to identify low-risk companies. Stock price forecasting and price movement prediction are vital issues in financial works. Deep learning has previously been implemented for stock market analysis, with promising results. Because of the differences in architecture and stock issuers in each study report, a consensus on the best stock price forecasting model has yet to be reached. We present a methodology for comparing the performance of convolutional neural networks (CNN), gated recurrent units (GRU), long short-term memory (LSTM), and graph convolutional networks (GCN) layers. The four layers type combination yields 11 architectures with two layers stacked maximum, and the architectures are performance compared in stock price predicting. The dataset consists of open, highest, lowest, closed price, and volume transactions and has 2,588,451 rows from 727 companies in IDX. The best performance architecture was chosen by a vote based on the coefficient of determination (R2), mean squared error (MSE), root mean square error (RMSE), mean absolute percent error (MAPE), and f1-score. TFGRU is the best architecture, producing the finest results on 315 companies with an average score of RMSE is 553.327, MAPE is 0.858, and f1-score is 0.456.

DOI: https://doi.org/10.11591/ijece.v14i1.pp861-869

How to run

python.exe .\forecasting.py --code BBCA --type 1 --lookback '5,10,15,20,25' --scaler standard --batch_size 128 --epoch 150 --callbacks 1 --model TFGRU 

I'd like you to please open the file forecasting.py for more details. Existing models are CNN, CNN-GRU, CNN-LSTM, GCN-GRU, GCN-LSTM, GRU, GRU-CNN, GRU-LSTM, LSTM, LSTM-CNN, LSTM-GRU.

Custom architecture

import tensorflow as tf

class TFGRU:
    def __init__(self, n_classes=1):
        # Get model hyperparameters
        self.n_classes = n_classes

    def build_model(self,
                    input_shape,
                    dropout=0.25,
                    unit=128
                    ):
        self.dropout = dropout

        inputs = tf.keras.Input(shape=input_shape)
        x = inputs
        x = tf.keras.layers.GRU(unit, activation="tanh")(x)
        x = tf.keras.layers.Dropout(self.dropout)(x)

        # output layer
        outputs = tf.keras.layers.Dense(self.n_classes)(x)

        return tf.keras.Model(inputs, outputs)

Change the layers with a custom layer with accept keras input (x), put custom architecture in folder .\models, and import custom architecture in file .\models_init_.py