Using Custom Reddit Data to Trade Stocks

Very interesting work Jovad. How did you decide which submissions statistics to use as inputs to sentiment score? Is this based on work done elsewhere? And why the choice of those ticker symbols? 

Oddlot theory is proving to be resoundly incorrect this year. Recent market moves in the NASDAQ have been attributed to big Softbank option bets, however, I have seen a few analyses from vol traders that suggest 1) dollar premium spent on small-lot option trades have swamped the recent trades allegedly from the 'NASDAQ Whale' by as much as 10x; 2) small lot trades have mostly been short dated (<30 days DTE) tenors vs 6-12months from Softbank 3) a dealer/MM 'gamma squeeze' situation more likely to be caused by these shorter dated flows since shorter (longer) dated options have more gamma (vega).

There is probably still money to be made following the 'dumb-money'...!
 

Hi P Chen, 

Thanks for sharing the podcast! After realizing a few flaws in the algorithm, I made a few changes. A backtest of the "upgraded" algorithm is attached, and here is the link to the commented python code for analyzing Reddit. 

We use the SentimentIntensityAnalyzer from nltk.sentiment.vader to determine the sentiment of a body of text. Originally, we only analyzed sub comments but it now also analyzes submission title and submission body text. Once all scores are collected, they are averaged and returned as averageScore. 

This work is not based on work elsewhere. However, there are plenty of youtube tutorials on Reddit API and a few on analyzing sentiment.

I picked those stocks by briefly scrolling through wallstreetbets and picking four tickers. I have included a backtest with 11 more tickers that I saw mentioned in the subreddit. If this algorithm were to be run real-time, it would be intuitive to create a separate program to pull a list of tickers mentioned on the subreddit within a 30-day rolling window (ie. will not pull ticker 'F' if it was mentioned 120 days ago; it must be a recent discussion).

Yes, you are right. Odd lot theory has been incorrect this year, and I believe it will continue to be inaccurate. As the friction for small investors to participate in the market declines, a collection of small investors can have a greater market impact and company news reports may create increase short-term volatility for a stock price. Here is an interesting short article on a mutual fund's manager thoughts on Robinhood. 

Some pitfalls with this algorithm:

1. There is no filter for predatory wallstreetbets submissions (ie. link)
2. Only searches for mentions of ticker names but not company names (quick fix)
3. No credibility model

An idea to prime this algorithm for live trading is to use a neural network. It will create a prediction for the next set of sentiment scores and use those predictions to manage risk. The CSV file would also need to be updated and fed into the algorithm daily.  If anything I said is confusing please let me know. 

(One change I made to the backtest is increased the trailing stop loss from 0.05 to 0.08 to allow more room for recovering losses)

Cheers,

Jovad Uribe

Hey Jovad, I have a similar type of strategy I am looking at doing albeit without sentiment analysis and rather the adding of the top 10 tickers in a csv sitting on a public google drive.  The CSV changes daily via a python grab. Would there be an interest in assisting with this to contribute the results to the community?

Hi Jim,

For assistance on integrating a reference data source, refer to our documentation.

Best,
Derek Melchin

thanks Derek,  I am going to be putting something together around this.  I appreciate the link regarding the ref data source.

Hi Jim, 

I would be happy to assist. You can send me pm via the QuantConnect slack. 

Hi Jovad. The emails I'm sending to you are bouncing:

DNS Error: 8865401 DNS type 'mx' lookup of meridiancapitaladvisors.com responded with code NXDOMAIN Domain name not found: meridiancapitaladvisors.com

Posting here is the only way I know of to reach you. Could you try me from a different email? Thanks much.

-Rob

Thanks, Jovad. Very interesting. Hope you could successfully apply this to real-time trading!