Deep Learning-Based Corporate Performance Prediction Model Using Financial Panel Data

This research explores the application of deep learning techniques in predicting the performance of consumer goods companies listed on the Nigerian Stock Exchange, leveraging panel data spanning from 2000 to 2021. Utilizing datasets sourced from Mendeley and African Financials platform, we preprocess the data by imputing missing values and transforming it from wide to long form. We then devise a 4-class performance target based on profitability ratios extracted from the dataset. Four deep learning models—Gated Recurrent Units (GRU), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM)—are designed, fine-tuned through hyperparameter optimization, and trained on the preprocessed data. Evaluation of model performance reveals LSTM as the overall best-performing model, exhibiting a mean absolute error of 0.000405. Remarkably, all deep learning models demonstrate perfect classification results, as evidenced by confusion matrices devoid of misclassifications and ROC curves with an area under the curve (AUC) of 1 for all models. These findings underscore the efficacy of deep learning methodologies in accurately predicting corporate performance, thus offering valuable insights for stakeholders in finance, investment, and strategic decision-making.