Due to hypercompetitive internet service provider market, customers can easily move from one companies or operator to another if they did not obtain a good service. This customers’ movement is a major issue for companies as the reason most often found why customer churn. Churn management is an important program for companies to maintain valuable customers thus predicting customer churn is crucial. In literature, churn analysis has been widely studied with a various churn analysis techniques. These techniques mostly utilize customer complaint data, customer tenure, customer usage, etc. This paper employs customer opinions from Twitter with some specific keyword. It investigates whether measurements of collective sentiment (mood) states about a product, extracted from Twitter feeds, are correlated to the value of the churn rate of the observed product. This study examines the text content of daily Twitter feeds by Convolution Neural Network that measures positive, negative or neutral sentiment. Cross-validate the resulting mood time series with Granger causality analysis was conducted. A Recurrent Neural Network was then used to test the hypothesis that mood states from Twitter are predictive of changes in churn rate values. The results indicate that the accuracy of churn rate predictions can be improved by the inclusion of specific mood dimensions, that is negative sentiment, but not others. The Mean Average Percentage Error (MAPE) is about 1.47%.