Analyzing Cryptocurrency Sentiment on Twitter with LangChain and ChatGPT | CryptoGPT

CryptoGPT’s sentiment pipeline turns an author’s Twitter activity into daily sentiment scores by combining LangChain with ChatGPT and forcing structured JSON output. The core move is to group a single account’s tweets by date, enrich each tweet with engagement signals (like view counts), and then ask ChatGPT to return an aggregate bullish/bearish score on a 0–100 scale for each date—without any extra commentary.

The workflow starts after tweets have already been collected and cleaned. For sentiment analysis, the system takes one Twitter handle at a time and groups that author’s tweets by their posted date. For a chosen date range, it passes ChatGPT a list of tweets in a custom text format: each tweet entry includes the tweet’s view count and the cleaned tweet text. The design also caps the number of tweets sent to ChatGPT—if an author has more than 100 tweets, the pipeline samples 100 to keep prompts within practical limits.

On the LangChain side, the implementation uses a ChatOpenAI model configured with temperature set to 0 to reduce randomness and improve repeatability. A PromptTemplate is built with two key inputs: the Twitter handle and the formatted tweet list. The chain is created with LMChain, then executed by feeding in the handle and the generated tweet text. The output is then parsed as JSON, relying on the prompt to keep responses machine-readable.

The prompt itself is tailored to steer ChatGPT toward consistent scoring. It frames the model as an experienced crypto trader who pays attention to historical predictions made by that specific Twitter account, then weighs those patterns against the new tweets provided for each date. The scoring instruction is explicit: return only a JSON object where each record contains a date and a sentiment value from 0 to 100, with the mapping defined as bearish at the low end and bullish at the high end. It also includes a “no explanations” constraint—“return just the roll Json response… do not explain”—to prevent the model from adding prose that would break downstream parsing.

A helper function prepares the tweet list for the prompt by iterating over tweets grouped by date and concatenating entries in the format “views - tweet text.” The pipeline then runs the sentiment analyzer for each handle stored in session state, storing results as a dictionary keyed by author and date.

In testing, the system produces a JSON-like response with multiple dates and sentiment scores for accounts such as Michael Saylor (shown as “michael sa” in the example) and Elon Musk (“L musk”). The results demonstrate that ChatGPT can follow the structured-output requirement and generate daily aggregate sentiment values from engagement-weighted tweet text. The next step, teased for a follow-up, is visualization—turning the returned sentiment dictionary into a DataFrame and chart using Streamlit.