An improved DenseNet model for prediction of stock market using stock technical indicators.

• A lightweight approach for stock trend forecast using DenseNet-41 and autoencoder. • A custom DenseNet-41 model with STI's for reliable features to improve the prediction accuracy. • Our framework can predict the buy/sell/hold signal based on a novel deep learning model. • Presented a robust model for stock prediction due to efficient STIs selection and deep features. • Achieved state-of-the-art results on a challenging dataset namely the Yahoo Finance database. Reliable estimation of future cost prices of stocks are a significant and exciting task in both educational and economic research. Current progressions in machine learning (ML), and explicitly, in the area of deep learning (DL) have permitted the research community to utilize the financial information from the social and finance websites to estimate the future prices of products in a viable manner. Besides, the accurate estimation of stock future trends is a complex task due to the unstable nature of the finance data. We have tried to overcome the limitations of existing studies by presenting a more reliable DL-based approach. More clearly, the proposed work is based on estimating the final prices of products by utilizing the 10 years of stock data from the Yahoo Finance website along with the computed Stock Technical Indicators (STIs). The calculated STIs are initially passed as input to the autoencoder to reduce the feature space by eliminating the highly correlated data and providing a more nominative set of STIs. The processed STIs together with the Yahoo finance data are passed as input to the DenseNet-41. The final feature vector computed by the DenseNet-41 is then passed to the SoftMax layer of the network to determine the closing cost prices of products for small, medium, and longtime horizons. Based on the predicted prices of stocks the proposed approach provides three types of signals designated as buy, sell, or hold to assist the investors in their decision-making. Performance evaluation demonstrates that our work outperforms the latest methods by acquiring a minimum MAPE score of 0.32. [ABSTRACT FROM AUTHOR]