We are doing a project for crypto market trend prediction where we are trying to predict whether a crypto should buy, hold or sell based on the price and sentiments. First we are preparing a historical price dataset including ochlv and technical indicators with timestamp and a twitter tweets dataset which contains both time stamp and the tweets. Tweets which are preprocessed (removing bots and promotional tweets, other nlp preprocessing steps) are passed to both vader and roberta model that predicts sentiment scores corresponding to them. And we are merging both the datasets (historical and sentiments) using the timestamp. The problem is that after passing all the tweets to the preprocessing steps, we lost a lot of tweets and some days as no tweets, then we passing these remaining tweets to the model and aggregate the model predicted scores on day based, we have a lot of days as NaN scores on both vader and roberta. How to solve these problem, wheras crypto market is highly volatile?