WallStreetBets Exploration Dashboard
How do Retail Investors React to the Market?
In this section we analyse Cumulative Abnormal Returns (CAR) and cumulative returns leading up to and following posts on WSB
We calculate and plot the average / median cumulative log returns starting at 14 days prior (-14) to a posts and ending 14 days after a post - the zero mark indicates the day that a post is made
CAR / Cumulative Returns Across all WSB Posts
Average CAR for Top 20 Tickers on WSB
What does the WSB Discussion Landscape Look Like?
We create a ticker-to-ticker interest overlap graph which measures the fraction of authors that share post about both tickers
The graph displays interesting properties, where we see targeted followings forming around certain tickers / sectors - such as the tech sector, the "short-squeeze" group (GME, BB), and more targeted / smaller groups such as the cruise line cluster
Discussion Topics Over Time
Relationship between WSB Signals and Future Asset Returns
Relationship between WSB signal and Asset Returns between One and Twenty-Five Weeks after Signal Observation:
We consider the linear relationship between signal at time t about an asset and its log-returns. The signals include:
CAR: the cumulative abnormal return in an asset on the day a WSB post is made
sentiment: the sentiment expressed in a posts about future asset performance (+1, -1)
prev_day_return: the log-return in an asset on the day that a post is made
url: an indicator for whether a post contains a credible news-source URL
proactive: an indicator for whether sentiment is different from current asset return
We propose several options on how to filter the dataset, including (0 implies no filter is selected):
Flair Filter: if set to True our sample includes only posts marked as DD posts; if False the sample includes all DD posts
Graph Measure: an option to filter sample by the graph properties of the comment structure of posts
Sentiment Filter: either filter sample to posts with positive (+1) or negative (-1) sentiment
Meme Stock Filter: option to exclude GME and AMC (True)
Quality Measure: an option to include an additional control variable (such as CAR) in the regression
Portfolio performance:
We consider the performance of a portfolio trained and validated on WSB + market data. Left panel considers week-long time horizons for training, validation and testing; the right panel considers a monthly time horizon.
We select a training period between 5-19 weeks (left) and 2-11 months (right)
We perform feature selection on a set of market + WSB features using training and validation data
The validation and test time periods are both two time periods
After model training we construct a long / short portfolio from our signals and observ ethe log-return in two subsequent test periods