Module 2 Lab: Empty Folder to Verified Figure
In this lab you run the same loop the module page lays out: empty folder, download script, clean panel, publication-style figure, reproducibility README. The difference is that you choose the measure. The demo built R&D intensity, and your lab measure is a different one, so when you check your figure at the end you are checking your own work rather than a copy of the demo. The starter pack (lab2_starter.zip) contains an opening-prompt template with blanks for your measure, tags, and aggregation; a verify checklist; and a card explaining DECISIONS.md and LOG.md. The zip has no data in it. FSDS is public, and downloading it yourself is part of the exercise.
Choose your measure
Start with an empty folder: mkdir ~/lab2 && cd ~/lab2 && claude. Because the folder is empty, Claude sees this project and nothing else. Then pick one of the three measures below. They are tiered by ambition. Choose the tier that matches how much cleaning judgment you want to practice in this module, which need not be the hardest one.
Base: leverage by industry and year. Compute book leverage (long-term debt divided by total assets) as the median by two-digit SIC code (sic2) and year. This is the closest variant of the demo: the same join and the same aggregation, with a different pair of XBRL tags. Expect the debt tags to cover only about a third of 10-K filers. In the pinned quarters, LongTermDebtNoncurrent appears in 26–28% of 10-Ks and LongTermDebt in 6–9%, while Assets appears in 98–99%. A missing debt tag can mean the firm has no long-term debt, reports it under another tag, or shows it only in a footnote, which FSDS does not carry. Decide in DECISIONS.md whether a missing debt tag counts as zero or as missing, and report how that choice moves the median.
Standard: R&D intensity, different industry grain. Build the demo’s R&D-intensity measure (R&D expense over total assets), but aggregate by one-digit SIC (sic1) instead of the demo’s sic2. You already know the measure. What is new is the grouping decision and how it changes the pattern you see.
Stretch: revenue concentration. For each year, compute the share of total reported revenue that comes from the ten largest filers by revenue. This tier takes more cleaning judgment than the other two. You have to decide how to handle firms that report revenue under more than one tag, and how to break ties at the tenth-largest boundary. Pick it if you want more practice with ambiguous cleaning decisions rather than with the mechanics of the join. Two things to know before you start. The main revenue tag changed with the 2018 revenue standard: RevenueFromContractWithCustomerExcludingAssessedTax appears in no 10-Ks in the 2013 and 2016 files and in 43% of them in the 2024 file, while Revenues falls from about 41% to 30%. And keep only rows with a blank segments field, or you will add each firm’s segment revenue on top of its total.
Whichever tier you pick, use the known-good quarters the starter pack pins for you: Q1 of 2013, 2016, 2019, 2022, and 2024, which hold most 10-Ks for fiscal years 2012, 2015, 2018, 2021, and 2023. Each zip is about 95–125 MB, so the download is roughly half a gigabyte. (FSDS groups filings by the quarter they arrived, and most firms file their 10-K in February or March; the Q4 files hold only a few hundred 10-Ks each.) Five quarters give you a long enough time series for a meaningful plausibility check without downloading and cleaning fourteen quarters in an 80-minute lab. Tag coverage in FSDS also varies by quarter. A value reported under one tag in an earlier filing sometimes appears under a slightly different tag in a later one. That is another reason to pin exact quarters instead of defaulting to all of them. These years skip 2019 and 2020, so the module page’s 2020 example will not work here. Choose a plausibility check your five years can show, such as the number of filers per year or one well-known firm’s value checked against its 10-K.
Requirements, regardless of tier
The deliverable has the same shape as the demo’s:
- named scripts only;
DATA_PROFILE.md,DECISIONS.mdandLOG.mdrequested in your opening prompt before any analysis runs, not added later once something looks off;- a plot of the raw firm-level series before you aggregate it;
- an attrition table showing every point where your sample size changed and why;
- one publication-style figure;
- a README that lets a stranger, or you in three weeks, rerun the whole pipeline from scratch.
Adapt the opening-prompt template in the starter pack to your tier. Fill in the measure, the tags, and the aggregation. Leave the paper-trail section exactly as written, and do not trim it if you are short on time. It is the part that produces DATA_PROFILE.md, DECISIONS.md and LOG.md.
Then work in this order. Send the prompt. Read DATA_PROFILE.md as soon as it is written, before any measure is computed, and check the row counts and tag coverage. Ask for a plot of the raw firm-level ratio by year. Write your two named checks into the README before the figure exists. Let the figure render and refine it in at most three rounds. Run the checks. Ask Claude to finish the README last, then rerun the pipeline from the README in a fresh copy of the folder.
Stretch for finishers: if your pipeline is done with time to spare, do one of two things. Either rerun the same research question with a second measure and compare what changes, or write a paragraph (not code) on what the WRDS version of this pipeline would need to change: which table replaces sub.txt and num.txt, which key replaces adsh, and which filters Compustat would require that FSDS does not. This previews Module 4. It is not homework, and a few sentences is enough.
Debrief and exit ticket
Before you move on, ask yourself the debrief questions. Was your first-draft figure right on the first attempt? Usually it is not, and that is the expected outcome. Where did Claude recover from an error on its own, and where did you have to step in and redirect it?
Then answer the standard three exit-ticket questions in your README or your notes (one concept that clicked, one thing that broke, one thing you’d want to verify before trusting a similar result), plus this module’s sub-agent question: if you saw Claude start a sub-agent at any point during your lab, what do you think it was for? Module 4 answers that properly. When the course runs with a group, the debrief is a short discussion and the exit ticket is handed in, along with the WRDS checkpoint above.