This paper examines the effectiveness of recent large language model-based news sentiment estimation for stock price forecasting with the combination of latest transformer-based prediction models. To achieve a better accuracy in sentiment classification, experiments are designed to compare six different models (GPT 4, Llama 3, Gemma 2, Mistral 7b, FinBERT, VADER) in financial news sentiment classification, and it was found that recent large language models can outperform FinBERT and VADER, which are the most commonly used models in financial sentiment analysis. Based on the experiment results, Llama 3, with relatively stable performance, is chosen to classify the news sentiments of the selected companies. Informer, Transformer, TCN, LSTM, SVR, Random Forest and Naive Forecast are used to predict the stock prices with different sliding window sizes. Experiments with different scenarios are designed to evaluate the prediction ability of news sentiment. Results show that adding news sentiment data can indeed improve the stock price prediction. Informer, one of the state-of-the-art transformer models for long-term prediction tasks, yields the best performances in most cases. Ablation study of Informer suggests that the generative style decoder plays an important role in performance improvement.