Crypto Trading Bot using Twitter Sentiment

Danny Hines

Danny Hines • 15 min read

Posted Jan 20, 2022

Measuring crypto sentiment

TL;DR: Built a Heroku app that connects to the Twitter and FTX APIs to listen for tweets from top influencers, measure the sentiment, and make trades based on the sentiment, follower count, and price history. This bot is currently not trading, but I'm saving data and running backtests. If you're interested in working with me to improve it, hmu.


I'm not a big crypto guy. Blockchain tech is undoubtedly cool, but I resist investing in crypto because I have no idea what the price should be. That is, it doesn't produce anything. There is no CEO, no revenue, no cash flows, and there ever-present threat of the U.S. outlawing it. But that doesn't mean that you can't profit off of it. Instead of hodling random coins with good marketing, I trust myself more to trade with software. Even if I'm unsuccessful, trading crypto still sounds like a fun way to lose money.

The Elon effect

Because there's no company to analyze or a market to compare it to, crypto price movements are dominated by hype and FOMO. This is especially true for smaller-marketcap coins that have the potential take over the world. And when the next big thing happens in crypto, Twitter is where most people find out.

The father of value investing Ben Graham once said, "In the short run the market is a voting machine, but in the long run it is a weighing machine.” It's too early to tell the weights, but I know the markets have voted for Elon Musk.

Elon's tweet sends doge to the moon

Elon tweeted that you will be able to buy Tesla merch with Dogecoin and the price shot up 30%. That's what I'm talking about.

I don't care as much about the big bois BTC and ETH, because bitcoin is mentioned all the time and its market cap is so large that any given tweet will cause a ripple in the tsunami of bitcoin-related tweets. So I'm going to look for influencers tweeting about altcoins, and see if I can buy before crypto twitter does.

The Strategy

Here's what we're going to do:

  • Listen for tweets from prominent crypto influencers
  • When they mention a coin, we measure the sentiment of the tweet and the price information
  • If the sentiment is positive enough, place a market buy order
  • Place a trailing stop sell order, so if they price moves X% below its high, it will sell

This will essentially act like a twenty-something year old who loves crypto twitter, except a lot faster and we'll actually sell. The goal is to buy at the market price as soon as the news/endorsement is made, and place a stop order slightly below our buy price (let's say 2% below market price). Instead of the typical limit sell order which you would place above the current price to lock in a profit, a stop sell order is placed below the market price, so in the worst case scenario there's a fixed amount you can lose. In our case, we're hoping for a “pop” that is larger than the trailing stop % to lock in gainz.

This buy/trailing stop combo is essentially a bet that the price will go up by more than 2% before it goes down by 2%.

Assuming the price goes up (because it always goes up), the stop order will automatically update our stop price to be a certain percentage (i.e. 2%) below the highest price. That way, if we buy at 100, it goes to 110 and drops back down, we'll still lock in a ~7.8% gain because we'd sell around 107.8 (2% below 110).

You're probably asking, couldn't you just put a limit sell order at 10% higher than the price you bought at to make 10%? Sure, but what happens if it goes down? Or even more annoying, what if goes up by 50%? We're looking for moonshots here, and the trailing stop is our best chance of capturing big swings upward.

To see how I trade with Node.js, check out my code snippet for creating this buy/trailing stop combo using the FTX API with Typescript.

The Twitter API

Twitter has a solid API that lets you search for previous tweets as well as stream them live. By opening a stream, we can listen for certain keywords and/or from certain users. Without paying $100/month, we can fetch 2M tweets/month and 25 query strings, each with a max 512 characters. The queries look like advanced google searches, using quotes to match exact text and finding tweets from a certain user with from:username. I would run out of query space if I wanted to search for every keyword, so my query is a big list of usernames, so I get all the influencers' tweets and check whether a keyword is mentioned on my server.

Matching tweets to tokens

Identifying which tweets referenced which coins was a challenge. The problem felt simple at first, but increasingly became more complicated as I noticed the edge cases over time.

To know what's available to trade, I hardcoded a list of symbols and keywords that are closely related. My original bot used FTX.US which could only trade 17 symbols using US dollars. This list uses Binance.US, which has 65 symbols but required more effort to set up.

export const SYMBOLS = { "1INCH": ["1inch"], AAVE: ["aave", "aaveaave"], ADA: ["ada", "cardano", "iohk", "Charles Hoskinson"], ALGO: ["algo", "algorand"], AMP: ["flexa", "amptoken"], ANKR: ["ankr"], ATOM: ["cosmos"], AVAX: ["avax", "avalanche", "avalancheavax"], AXS: ["axs", "axie infinity", "axieinfinity"], BAND: ["band protocol", "bandprotocol"], BAT: ["basic attention token", "brave browser", "brave rewards", "attentiontoken"], ... }

The symbols are stored in a dictionary using the symbol as the key (eg. DOGE) and the value is an array of keywords which are guaranteed to refer to the symbol.

Matching text to symbols

I set up the matching function to look out for variations of the symbol, which are the keys in the dictionary. Because certain tokens like $ZEN and $LINK are actual words, I can't reliably check for “link” or “Link”. So I decided to check for an all-caps version of the symbol, as well as a $ and # preceding the symbol no matter its capitalization. The token $ABC would match “ABC”, “$abc”, “$ABC”, “#abc”, etc.

Anything else that refers to the token should be listed as a keyword, i.e. the symbol for ChainLink is $LINK. I set up the keywords so that if the keyword is lowercase, it will match the text no matter if it's capitalized or not. This is helpful to catch totally unique words like “axieinfinity”, which is capitalized in different ways. If there is capitalization in a keyword, it will need to match the capitalization exactly. Examples are “The Graph”, “Dash” or “Helium”.

For example, given the symbol/keyword for the token ChainLink:

LINK: ["chainlink"]

We would be able to match LINK, #LINK, $link, #link, ChainLink, chainlink, Chainlink

Originally I included the capitalized symbol name, i.e. “Btc”, but ChainLink caused a unique problem because “Link:” is incredibly common in tweets that link to an article. This strategy avoids “link” and “Link” but still catches “$link”, and I can add any other word or phrase in the keywords.

Measuring Sentiment

For measuring the sentiment and therefore how positive the tweet is toward the token I used the Javascript version of an open-source project called VADER. Vader is “a lexicon and rule-based sentiment analysis tool that is specifically attuned to sentiments expressed in social media”. **It uses a list of ~7,000 words and their relative positive/negative sentiment on a scale from -4 to 4, and the measures a string of text by counting the valence of different sections of the text along with negations, capitalization, and punctuation which affect the overall sentiment. The output of the sentiment function is an object that gives you the percentage of the phrase that was positive, negative, and neutral, and the **compound** score which is the overall sentiment (from -1 to 1).

const vader = require('vader-sentiment'); const input = 'Danny is very smart, handsome, and funny'; const intensity = vader.SentimentIntensityAnalyzer.polarity_scores(input); console.log(intensity); // {neg: 0.0, neu: 0.299, pos: 0.701, compound: 0.8545}

In general, a score ≥ 0.05 is considered positive, and ≤ -0.05 is negative. It's not perfect, but the compound score is usually a solid estimate of the sentiment.

I added about 200 words to the lexicon with crypto-related words and phrases like “bullish”, “just bought”, “tanking”, “to the moon”, etc. Here are tests using mocha/chai for my fork of the VaderSentiment project:

import { SentimentIntensityAnalyzer } from '../src/SentimentAnalyzer/index.js'; import { expect } from 'chai'; const getScore = (txt) => SentimentIntensityAnalyzer.polarity_scores(txt).compound; describe('Sentiment', function () { it('Exclamation point', function () { let strings = [ 'Danny is smart, handsome, and funny.', 'Danny is smart, handsome, and funny!', 'Danny is smart, handsome, and funny!!!', ]; expect(getScore(strings[1]))[0])); expect(getScore(strings[2]))[1])); }); it('Question mark(s) - ending in ? lowers sentiment', function () { let strings = [ 'Danny is smart, handsome, and funny.', 'Danny is smart, handsome, and funny?', 'Danny is smart, handsome, and funny?', 'Danny is smart, handsome, and funny???', ]; expect(getScore(strings[1]))[0])); expect(getScore(strings[2]))[1])); expect(getScore(strings[3]))[2])); }); it('Emojis', function () { expect(getScore('💘 and 💋 and 😁')); expect(getScore('🚀 🚀 🚀'))'🚀')); }); it('Negation', function () { let strings = ['I am buying Bitcoin', 'I am not buying Bitcoin']; expect(getScore(strings[1]))[0])); expect(getScore(strings[0])); expect(getScore(strings[1])); expect(getScore('problem'))'no problem')); }); it('Phrases', function () { let strings = [ "I like Eth, but I'm afraid of the gas fees", 'I like Eth, I think it will work out', ]; expect(getScore(strings[0]))[1])); expect(getScore('exchange is now accepting LTC')) getScore('exchange is accepting LTC') ); expect(getScore("if you're into crypto ❤️ this")); expect(getScore('This coin is not a buy right now')); expect(getScore('This coin is the shit')); expect(getScore('that would be worth a lot')); }); it('Non-influential text', function () { expect(getScore('Here are my top 10 favorite coins')); expect( getScore("If you bought $CRYPTO 1 year ago, this is how much you'd make") ); }); });

When to buy

This is obviously the hard part. At first I looked at buying when the sentiment was positive. That works, but if Elon Musk is tweeting, he only needs to mention the word “Doge” and it can go up. I decided to split the influencers into 5 categories, each with a different sentiment threshold for considering it a “buy”. These are exchange, influencer, news, trader, and superinfluencer. I've messed around with the thresholds a lot, here's an example of how it looks:

export const BUY_THRESHOLD = { superinfluencer: -0.05, exchange: 0.6, influencer: 0.35, trader: 0.2, news: 0.3, };

I originally had a very low threshold for the exchanges, for example, because I wanted to catch whenever a company like Bitcoin announced a new coin on their platform. After some backtesting and messing with the lexicon, it wasn't working well so I made phrases like “now trading” worth high valence scores like 3.2 on the -4 to 4 scale, which allowed me to bring the threshold up and filter out noise.

Because it's very confusing how I match keywords and decide to trade on single or multiple matches, I made a flow chart that attempts to explain how I approach it without giving you all the code.

Flowchart: how my bot decides what to do

Basically if the tweet mentions a single coin or one coin is mentioned more than others, it will see if it's a "good buy" by calling the shouldBuy() function. This function will resolve to true of the sentiment is above the threshold for that type of influencer given their follower count. It also looks at whether or not the influencer is associated with the token (i.e. the CEO of Binance Changpeng Zhao should have a higher buy threshold for buying $BNB). Additionally, there is a check on the 14-day RSI value of the coin's price, but this has given mixed results so I only consider this if the RSI is especially high.


To measure the performance of the strategy, I pulled the last few months of relevant tweets and added to the dataset over the past couple weeks. Although I want to have US dollars to trade with, I realized that the price of Bitcoin is the major driver of price fluctuations in altcoins. So when backtesting, I usually look at the price movement relative to Bitcoin, i.e. trading DOGE/BTC instead of DOGE/USD. This gives me a better perspective of how the price of the coin moved relative to the market as a whole. I use the regular Binance api (instead of to backtest because of the larger volume and they have every type of coin against BTC.

Some (unbelievable) results

I originally tested using a 1% trailing stop percent (that is, if the price is at 100, it will set the stop price at 99). Because of the inherent volatility, I noticed it would hit the stop price and sell at a loss frequently before a pop. I decided to try alternate trailing stop percentages ranging from 0.5% to 3%, to see how the trades would perform. Here's a few months worth of data for a bunch of influencers, it shows the net profit/loss % if you would have if you traded using my current strategy.

Backtesting data

This (probably confusing) table shows that every trailing stop percentage resulted in a positive P/L except using 2.5%. Despite this outlier, the totals give a sort of bell curve, with 2% being the most profitable. If I traded using a 2% trailing stop, I would've traded 1190 times off of 3946 tweets, with a net profit of 142.5%. This doesn't include fees, which would be quite high for the number of trades, but yeeeesh that's a huge profit.

This table also reveals the most and least profitable types of influencers. The most “correct” were the inlfuencers like TheCryptoLark. The least profitable were the exchanges like Binance and Gemini, which has made me rethink how I'm scoring tweets from exchanges.

Deploying it

At first I deployed this bot on AWS using a cron job to look for tweets every 1 minute. This meant I was up to 59 seconds behind every other twitter crypto bot, so I made the change to use Twitter streams to listen for tweets live. Because I despise running virtual machines on AWS (what's a subnet plz help), I decided to go with Heroku.

I deployed the bot, connected to a PostgreSQL database (also on Heroku), and use Coralogix for logging. I'm giving the bot about $100 to see how it goes.

Update 4/12/22

After running it off and on for a few months and losing about 10%, I decided to stop trading with this bot. Between all the noise on twitter, the fees, and losing money on the spread (difference between bid/ask), it's currently not a profitable strategy. The giant profit from backtesting was most likely a result of not considering the fees + spread, and because my test was during the biggest bull run in the history of crypto.

Sometimes a well-known trader will tweet the same thing a couple times in a row and the bot will buy several times, but the person tweeting is just live-tweeting while they're chasing their loses. Other times they're using sarcasm and the bot can't tell. Even when an exchange tweets that they're now trading a certain coin, it's not guaranteed to be a winning trade. I think it's because people are front-running the announcement somehow, but it's hard to tell because of the inherent volatility.

Another reason for stopping is a downturn in the market, both the crypto market and the global economy as a whole. As the crypto industry matures and the questions about it start to have answers, the news is less positive. The price of any liquid asset will be determined by its future, so when the future felt limitless, the prices reflected that. But building things is hard, and there are thousands of different coins, all claiming to have intrinsic value because it's the future.

Amazon announced it would be delivering packages via drone in 2013. I'm sure there was a small pop in the stock price when the news was announced, but in the past decade the news surrounding Amazon drone delivery has been predominantly negative. And I feel like that's what is happening in crypto.

I'll continue to collect the tweets and run backtesting to see if they would've been profitable trades, and maybe I can find a winning strategy later.

That said, I'm still going to trade on Elon Musk.

Subscribe to the newsletter

Get early access to articles on tech, finance and more.

Totally free, no ads, no spam.