LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A point-in-time Russell 3000 tweet panel, built around each summer’s index rebalance and totaling on the order of 100 million posts, is large enough to answer plain questions that do not require a trading rule. Message length rose over the years. Exact-text duplicates settled near an average of roughly 15 percent, inside a band the lecture describes as about 15 to 20 percent. About 55 to 60 percent of posts in a given year named a single cashtag, the dollar sign plus a ticker. Those counts are descriptions of the archive. They are not a return.
Why it matters
The project, underway since about 2022, starts from a simple interest: social text and stock returns. The archive is the part the lecture actually opens. Constituents are taken from the Russell 3000 as it stands at the early-summer rebalance, around the end of June into early July, and held to the next rebalance. The 2013–2014 window is the first block, on the order of 5 million posts, with a chart label the lecture reads as about 5.7 million. Counts rise year by year, with a dip in 2018–2019 and a jump in 2020–2021 to almost 20 million posts for that year’s constituents. The sum is the roughly 100 million.
The idea in plain words
Length is measured in words, not characters. In the first window the distribution centered around 12 to 15 words. In the 2020–2021 window it centered around 25, and some posts reached the character cap then in force, a cap the lecture does not quote from memory. Comparing only the endpoints would be a guess about the middle. The year-by-year chart is the actual claim: length was fairly stable, then in 2017–2018 the average jumped to roughly 20 words and kept rising toward about 26. More posts and longer posts are separate facts. The lecture’s question, left open, is whether later years therefore carry more information into a later return study.
Duplicates here are not reposts of the same status. They are distinct messages whose text matches exactly, such as two unrelated “buy” lines on the same ticker. The first year’s duplicate share was the low one. The next year, 2014–2015, rose to almost 22 percent. Later years eased and sat near a mean of roughly 15 percent, and the lecture summarizes a typical year as about 15 to 20 percent exact copies, sometimes repeated far more than twice. There are not enough years to prove a pull back to that mean. It only looks stable. Unique counts are the total times one minus the duplicate share. The open question is whether those copies change a downstream model that tries to relate text to cross-sectional returns.
How to use it
Freeze membership at the rebalance you can defend, and do not mix a 2014 ticker list into a 2021 pull. Recompute length, duplicate share, and single-cashtag share inside each window before you pool. A pooled sentiment score will inherit the duplicate rate and the multi-ticker rate whether you wanted them or not. About 55 to 60 percent of a year’s posts name one ticker. With several tickers, a simple lexicon can assign one mood to every name in the post. Mentions of other accounts, found with an @ pattern, came to 843,000 in the first window, and the lecture treats that count as a lower bound.
What usually goes wrong
Reading the length jump as a proven rise in information content repeats a conjecture. The lecture shows the jump to roughly 20 words in 2017–2018 and the later level near 26. It asks whether that means more content. It does not answer.
Questions people ask
How big is the panel?
On the order of 100 million tweets, summed across Russell 3000 windows from the 2013–2014 rebalance cycle through later years. The first window is about 5 to 5.7 million posts. The 2020–2021 window is almost 20 million.
What happened to message length?
It is in words. Early posts centered near 12 to 15 words. The average was fairly stable until 2017–2018, then moved to roughly 20 and later toward about 26. The 2020–2021 distribution centered near 25.
What share of posts are duplicates or single-ticker?
Exact-text duplicates were almost 22 percent in 2014–2015 and later sat near a roughly 15 percent average, with the lecture’s summary band about 15 to 20 percent. Single-cashtag posts were about 55 to 60 percent of a year’s posts, counting duplicates.