Analysis: WallStreetBets Discussion & Investor Behavior

WallStreetBets Discussion & Investor Behavior

WallStreetBets may contain some of the loudest voices among retail investors, but how well does discussion on the subreddit reflect the actual trading behavior of the retail investing community as a whole?

To answer this question, we tapped into Nasdaq's new dataset on retail investor trading activity, alongside our own data on WSB discussion.

THE DATA

We'll be using our data on WallStreetBets discussion, and specifically be looking at the number of times tickers are mentioned every day in WSB's daily discussion threads.

To get data on retail behavior, we will be using Nasdaq's retail activity. This gives us an 'activity' metric which quantifies what percent of retail trading volume is in a certain ticker on a particular day.

Combining these datasets allows us to visualize the largest single-day ticker mentions on WSB alongside the proportion of trading volume in those tickers on that particular day:

We can also visualize the largest single-day retail activity for tickers alongside WSB mentions of those tickers on those particular days:

Looking at how many total ticker mentions there have been on WSB over time (shown below), it's pretty clear that we will want to do something to normalize the data. There was a lot less comment volume on the subreddit in the past, so we will want to make sure that 2019 mentions aren't underweighted relative to 2021 mentions.

In order to do this, I will look at the day-over-day percent change in mentions of tickers on WallStreetBets, instead of the raw mention count. I will also use the day-over-day percent change in retail activity, instead of the raw value.

For the purposes of this analysis, I would also like to focus on tickers which have demonstrated some baseline amount of retail interest. I ended up narrowing the study down to the 120 tickers which had received a 'retail activity' of .005 on at least ten days.

ANALYSIS

The first thing I'd like to do is look at the correlation between the percent change in mentions of a ticker on WallStreetBets, and its percent change in retail activity.

Looking at the results for all of the top tickers combined, we get a correlation of around .23. While this tells us that there has been some positive correlation, I think we can get more interesting information by looking at the results by ticker.

Below is a table of tickers with the highest and lowest correlation between WSB mentions and retail activity:

HIGHEST

$SNRE (correlation of .99)
$CLOV (.83)
$WORK (.80)
$GME (.75)
$INO (.64)

LOWEST

$HOOD (-.01)
$SPY (.02)
$COIN (.03)
$AMD (.04)
$GILD (.04)

One drawback to this approach is that correlation is heavily influenced by outliers. To demonstrate the issue, below is a scatter plot of the data for SRNE. There's a near perfect correlation due to the extreme outliers in retail activity and WSB mentions on May 15th, 2020.

To avoid this issue, we can filter out all days with above a 10,000% increase in retail activity. This doesn't have much affect on the overall correlation (it is now .21), but it does give us different results when looking at the tickers with the highest and lowest individual correlation. Those results are shown below:

HIGHEST

Outliers removed

$CLOV (correlation of .84)
$VXRT (.81)
$WORK (.80)
$GME (.75)
$AMC (.73)

LOWEST

Outliers removed

$IVV (-.02)
$VOO (.04)
$CEI (.05)
$FUBO (.05)
$SPXL (.06)

MODELING

Because we have demonstrated that there is a relatively strong positive correlation between retail activity and WallStreetBets discussion for many stocks, a natural next step is to train models on the data to get a sense of the predictive power of WSB discussion.

This could be quite useful if done successfully, as the data on retail activity is only available on a daily basis. Because we have live WSB discussion data, we could make predictions on retail activity before the data is released.

For the sake of brevity, we'll fit a simple linear regression model to data for all of the tickers, using retail activity as the dependent variable and WSB discussion as the sole independent variable.

Fitting models on the 120 tickers, we find 102 to be highly statistically significant (P<.001).

In order to attempt to estimate the predictive power, we will examine the R-squared value of these regressions, which tells us what proportion of the variation in retail activity can be explained by WSB discussion.

Here are the stocks whose models have the largest R-squared value:

$CLOV (R-squared of .70)
$VXRT (.67)
$WORK (.65)
$GME (.56)
$AMC (.54)
$LK (.51)
$NKLA (.48)
$QS (.48)
$INO (.43)
$SNOW (.39)

We can also look at the coefficients to see how much the model predicts a 1% increase in WSB discussion will increase retail activity.

Here are the stocks whose models have the largest slope coefficient:

$CLOV (.86)
$SAVA (.76)
$VXRT (.63)
$GME (.60)
$BBIG (.55)
$ATER (.41)
$AMC (.38)
$WORK (.34)
$NKLA (.31)
$BBBY (.31)

CONCLUSION

Using these models, we can make live predictions on retail activity throughout the day. We have added those predictions to our WallStreetBets dashboard.

There are a couple ideas for further work which we didn't end up pursuing for this analysis piece:

1. Combine our live data on WallStreetBets discussion with live price movement data in order to improve our predictions. This would allow us to incorporate information about why certain stocks are getting talked about on a particular day.

2. Investigate what impact retail behavior has on stocks' future performance. If there is strong evidence that stocks with large increases in retail activity overperform or underperform in the short-term, it might make sense to build an alrogithmic trading strategy around the live predictions which we are generating.