A combination of the deep learning and metaheuristic algorithms for profit prediction using financial accounting information systems

Proper forecasting of profits is the key to efficient financial decision-making. This paper presents a hybrid architecture of long short-term memory (LSTM) networks and Harris Hawks Optimization (HHO) to achieve a better predictive performance with adaptive tuning of hyperparameters. It was applied to a Chinese stock market dataset containing balance sheet items, income statements, and cash flow variables of several fiscal years. The comparative experimental results with the standard LSTM, gated recurrent unit (GRU), artificial neural network (ANN), convolutional neural network (CNN), support vector regression (SVR), and autoregressive integrated moving average (ARIMA) indicate that LSTM-HHO has better precision with RMSE of 1.85 + 0.12, MAE of 1.45 + 0.10, MAPE of 3.78 + 0.21, and R 2 of 0.93 + 0.01, and these results show 23-45 percent better accuracy compared to the standard LSTM and 35-50 percent better accuracy compared to Convergence analysis indicates that HHO can select the best hyperparameters more rapidly than particle swarm optimization (PSO), differential evolution (DE), and genetic algorithm (GA). Robustness tests ensure consistent behavior in the context of changing the size of datasets, missing data, and noise, whereas sensitivity analysis suggests low RMSE with respect to changing hidden units, dropout rates, learning rates, and layer structure. These findings represent that the suggested LSTM-HHO learns the nonlinear correlations and temporal factors existing in the financial data with reliable and robust profits predictions. The suggested architecture can do financial and accounting management, while it makes strategic planning and reduces risk.