Predicting stock prices remains a critical research focus in finance, complicated by numerous external factors. Recent approaches utilize high-level data, such as tweets and news articles, for financial sentiment analysis. However, these methods often struggle to accurately capture public sentiment, leading to prediction inaccuracies. We introduce a high-quality dataset and a novel data collection method using a programmable search engine and FinBERT sentiment analysis, resulting in weekly sentiment metrics and stock prices for 20 stocks. The dataset was used to train and enhance a Flair sentiment model, integrated into a neural network for stock price prediction. Our results show that the neural networks with multiple time series output parameters outperform single-output models. Moreover, fine-tuned Flair models achieved higher accuracy than FinBERT-based models in predicting stock prices. This research highlights the potential of novel databases and refined sentiment models, offering improved insights into sentiment analysis in finance and introducing an innovative data collection method applicable across sectors using natural language processing.