Discover more from Daniel Stone
Money Stuff is Linear-ish
What was once a pre-market morning newsletter became an afternoon newsletter, 12 seconds at a time
I recently spent a bit too much time trying to find out when a man in Brooklyn presses a button.
His name is Matt Levine, and he writes a popular daily finance newsletter for Bloomberg called Money Stuff. He’s known for his incisive and humorous commentary on Wall Street, which makes complex concepts accessible to non-experts while remaining relevant to a large audience of industry insiders.
I’ve now subscribed to Money Stuff for six years and noticed it comes out later than it once did. Out of curiosity, I decided to graph the times Money Stuff has come out.
And the result was surprisingly linear-ish.
It’s just so tidy! If you do the math, you find that Money Stuff has come out about 47 minutes later every year since 2015.That works out to about four minutes a month or 12 seconds a business day. Clearly, the rate has varied a bit and it’s not perfectly linear (hence the “ish”) but the sheer contiguity is striking.
What’s strange is that in life schedules tend to move in blocks—not monthly four-minute increments. You move and your commute gets seventeen minutes longer, you start having a standing meeting Tuesday mornings—or even your editor says “let’s put the newsletter out at 11 instead of 9.”
So, you’d expect this chart to look a bit more like a bunch of floating steps.This is actually true from 2019 to 2021 when the times bunch around noon. You can pretty safely guess that Money Stuff was scheduled to be released at that time, since during this period newsletters were not released before noon.
Probably the almost uniformly-gradual shifts in release times have just happened, but I would like to imagine that it’s all part of a sophisticated daypart repositioning strategy. Rather than move Money Stuff from the morning to the afternoon at once, a clever marketing VP thought it would be wise to move it back by a few seconds each day so as to not upset discerning readers.
Now, there’s some other stuff you can do with the data. For instance, you can look at the days of the week on which Money Stuff has come out.
Lastly you can see that Money Stuff has gotten a bit longer in recent years.
The word counts are better viewed quarterly totals, which show us that there has been increasingly more Money Stuff.
I suppose you might wonder about the point of the above exercise. Other than being sort of interesting, reading Money Stuff is part of the routines of tens of thousands of people. The charts above describe how certain fundamental elements of that experience have changed: whether you read it with oatmeal or a tuna sandwich, and whether you can finish reading by the time you’re done eating.
See infra Methodology.
I should also note that Levine took parental leave in 2020, which is the story behind the gap in this and subsequent charts.
Using OLS, I estimate a rate of 2.1674*10⁻³ hours/calendar day, which works out to 3.95 minutes a month or 11.87 seconds per business day. R²=0.93, 95% confidence interval: [11.7 seconds per business day, 12.0 seconds per business day], p=2.22*10⁻¹⁶.
The release times are also extremely consistent across weekdays (the graph is boring), so it’s also not the case that some weekly event on say Tuesday mornings is pushing the average time later over time.
Note: I don’t account for the non-Money Stuff articles Levine writes, though I do exclude the period of Levine’s 2020 family leave and trading holidays.
I used the Money Stuff emails that I have to determine the word counts, and I’m missing about fifty, so the totals are slightly lower across the board than they probably are.
At first, I put all my Money Stuff emails (most of the run from 6/2016-present) into a folder, extracted the timestamps, and put them in a spreadsheet. This gave me a chart suggesting how much the data followed the trend described above.
The only slight problem with this data is that a) I have deleted a few emails over time b) the timestamps are lagged from the official filing dates and c) the earliest emails are from mid-2016, while Money Stuff was first published in 2015.
So I decided to create a list of Money Stuff articles going back to 2015 from scratch.
It is possible to access a list of Levine’s articles from his author page in a sort of weird JSON/HTML hybrid format. From this I extracted the metadata of about 1,300 articles: headlines, subheads, timestamps, article links, and section titles. Using the links, I hoped to scrape data from the 1,300 article pages.
But, I found that Bloomberg.com has a good scraping defense system. (I tried using a simple request with a user-agent, a request using the headers/cookies I copied from my browser, Selenium, and a more cloaked Selenium. Every time, the first request was blocked.)
Even browsing the Bloomberg website by hand (I got a trial subscription), you can only make about 5 requests a minute without triggering an “are you a robot” page. This is to Bloomberg’s credit, I suppose. You don’t get a shiny building that’s like a spaceship by letting strangers copy sizable chunks of your IP.
The main upshot of not being able to automate scraping was that I could not use article text to verify whether a Levine-authored article was indeed a Money Stuff newsletter (many aren’t) or calculate word counts. Further, along the way, I found that the timestamps I had initially collected reflected the latest edits to articles and not publication times. Without the contents of individual article pages, I could not automate checking whether an article timestamp in the list I had was the original timestamp.
So instead I had to be slightly creative. Here’s what I did:
I nixed: a) a series of clearly duplicate articles in the list (resulting from publishing an article in two sections) and b) every result from a Google site search of “this post originally appeared in Money Stuff.”
For the remaining articles, I tagged articles as Money Stuff when:
“Money stuff” was in the URL, headline, or section name. (Bloomberg’s labeling practices have changed over time.)
The subhead began with “Also” which is true of 72% of Money Stuffs, but only 4% of non-Money Stuffs.
For the period after February 2015, when Money Stuff was launched, I skimmed the 277 articles that were not Money Stuff articles and identified about 30 Money Stuffs I had missed
Since dates are in Bloomberg article URLs, for every instance in which the date in an article’s URL differed from the timestamp I had, I checked the article and amended the date/timestamp to reflect the initial time of publication.
ran a regression on the times of the Money Stuffs and
double checked any articles (about 50) for which the residuals seemed large ( >1.5 hours), and in 9 of 10 cases they were newsletters that had been edited earlier in the day. So I amended the times to reflect publication instead of edits.
I think this is kosher (i.e. not fudging things) because even though I did not capture instances in which an edit was made 15 minutes after publication, the thing I was really trying to capture here was initial delivery. There is admittedly a small chance that the error goes the other way in some cases (i.e. some emails were extra-early but edited later), but in spot checks of 50 newsletters where the residual was basically zero I didn’t find any. Out of an abundance of caution, I do not include a chart about posts before/during trading hours due to not being able to catch all the original timestamps.
In the end, I’m sure there are some Money Stuffs I missed and erroneous timestamps, but I think I caught most of them.
For the word counts, I went back to the email data, extracted the text (and removed the headers and footers).