Automated Sales Data Extraction and Forecasting Using Deep Learning

Sales data is vital for informed business decision-making, inventory control, and demand forecasting. However, in many organizations, this data exists in unstructured formats such as scanned PDFs and invoices, posing significant challenges for manual data extraction due to its time-consuming and error-prone nature. This project proposes an AI-driven, end-to-end system that automates the extraction and forecasting of sales data. The system employs TableNet for table structure recognition and Tesseract OCR for text extraction from scanned documents. The extracted data is structured into CSV format and then used for forecasting future sales trends using deep learning models—Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN). A custom dataset comprising over 500 annotated invoice images and corresponding sales data was used for training and evaluation. The proposed system achieves an average table detection IoU of 88.5, OCR accuracy of 92.3, and a forecasting RMSE of 18.6 on monthly sales data. Key contributions include: (1) development of a robust, automated pipeline integrating TableNet, OCR, and deep learning forecasting models; (2) application of modified TableNet for real-world, noisy invoice layouts; (3) quantitative evaluation demonstrating high accuracy in both data extraction and trend forecasting; and (4) a scalable solution that significantly reduces manual effort while improving business analytics capabilities.