Scraping NYT crossword stats ⬛⬜⬜⬜⬜⬛

When I first subscribed to the New York Times crossword in January 2017, I could solve some Monday puzzles and maybe the occasional Tuesday. Over time, I slowly got better at noticing the common patterns and picked up enough crosswordese to be able to finish it fairly consistently. As of November 2024, I've now maintained a daily streak stretching back 7+ years to August 2017 (knock on wood...).

The NYT crossword app exposes some solve time statistics like all-time average and fastest time, both broken out day. But there is no good way to see how much you have improved over time. You can open each puzzle in the archive individually and see your times, but there aren't any overview plots or officially-sanctioned developer APIs to fetch your data. Luckily, all puzzle stats are fetched via client-side Javascript, making it easy enough to scrape the data.

Regularly-updating plot:

Some observations:

• The crossword clues generally get harder over the course of the week, peaking in difficulty on Saturday, and the correlation with solve times shows up pretty clearly in the solve times.
• The clues of Sunday puzzles are roughly at a Thursday or Friday difficulty level, but the grid is extra-large, so it's usually the slowest day.
• Thursdays and Sundays usually have some sort of theme and/or trick, which usually added extra difficulty, especially early on when I wasn't as familiar with the usual patterns that constructors follow.

Caveats:

• I didn't count any puzzles that I didn't finish or that I used the "check" or "reveal" assists on, so there's some survivorship bias. Again, this only affects the early data, as I've since stopped using those features.
• I generally solve puzzles on my phone, but every now and then I'll solve them on my computer, which shaves some time off. This is just another source of noise.

Scraping the data

This repo contains a Rust crate that scrapes data from the NYT's servers. To use it, install Rust and then run the following to generate a CSV file containing the data:

# See help screen
$ cargo run --release -- --help

# Example usage starting search from the crossword on January 1, 2016 onward
# Subsequent program runs will use existing file as a cache 
$ cargo run --release -- -t <your NYT token> -s 2016-01-01 data.csv

# Example usage with increased quota to set rate-limit to 10 requests/second
$ cargo run --release -- -t <your NYT token> -s 2016-01-01 -q 10 -o data.csv

The NYT subscription token must be extracted via your browser (see below).

The program will fetch results concurrently, but by default, requests are limited to 5 per second to reduce the load on NYT's servers. While you can choose to override that limit to speed up the search, be nice and use something reasonable. There shouldn't be any need to run this script very often so it's better to just err on the side of being slow. Regardless of the setting you use, default or not, I'm not responsible for anything that happens to your account.

Design goals

1. Reduce requests made to NYT's servers. If provided, data from previous runs is loaded and used to avoid re-requesting information pulled from previous runs.
2. Reduce load on NYT's servers. Requests made to the server are rate-limited.
3. Maximum concurrency. Requests are made as concurrently as possible given the other two constraints thanks to async/await. It's totally overkill with the default amount of rate-limiting 🤷🏽‍♂

Extracting your subscription token

These instructions are for Google Chrome, but you should be able to do the equivalent with other browsers' developer tools.

1. Open the Developer Tools dialog
2. Open the Network tab
3. Navigate to https://www.nytimes.com/crosswords
4. Look for a request for some kind of json file e.g. progress.json, mini-stats.json, or stats-and-streaks.json.
5. In the headers pane, find the list of cookies, and fine NYT-S in that string. That is your token. If you can't find the NYT-S cookie in the request, try a different json file.

Under the hood

The script scrapes data using some undocumented but public REST APIs used by the official NYT crossword webpage. The APIs can be found by using the included set of developer tools in Chrome or Firefox to peek at HTTP traffic while browsing the crossword webpage.

Some details if you want to bypass the script and replicate the functionality yourself:

1. Each puzzle is assigned a numerical id. Before we can fetch the stats for a given puzzle, we need to know that id. To find it, send a GET request as below, specifying {start_date} and {end_date} in YYYY-MM-DD format (ISO 8601). The server response is limited to 100 puzzles and can be limited further by adding a limit parameter.

   curl 'https://www.nytimes.com/svc/crosswords/v3/36569100/puzzles.json?publish_type=daily&date_start={start_date}&date_end={end_date}' -H 'accept: application/json'

2. To fetch solve stats for a given puzzle, send a GET request as below, replacing {id} with the puzzle id. This API requires a NYT crossword subscription. {subscription_header} can be found by snooping on outgoing HTTP requests via Chrome/Firefox developer tools while opening a NYT crossword in your browser. Alternatively, you can supposedly extract your session cookie from your browser and send that instead (see linked reddit post below), but I haven't tried it myself.

   curl 'https://www.nytimes.com/svc/crosswords/v6/game/{id}.json' -H 'accept: application/json' --cookie 'NYT-S={subscription_header}'

3. Check out the calcs and firsts field of this response to get information like solve duration, when the puzzle was solved, and whether any assists were used.

4. Rinse and repeat, collecting data for the dates of interest.

Plotting the data

Use your favorite tools to analyze and plot the raw data stored in the CSV file.

My plots are generated via the Python script in the plot folder. To use it, run the following:

# Install prerequisites
$ pip3 install -r plot/requirements.txt

# Generate plot
$ plot/plot.py <path to csv file> <path to save plot image>

The output path can be an SVG or PNG file.

Auto-updating

The plot above is auto-generated by a regularly-scheduled job running on the Google Cloud Platform.

cloud_run.py implements a Flask server that glues together the stats fetching and plotting scripts, and the whole thing is containerized and run via Google Cloud Run.

References

• Relevant Reddit post: for figuring out how to find the right APIs to hit
• Rex Parker does the NY Times crossword: grumpy old man

Disclaimer

This is not an officially supported Google product