Code and data behind the videos.

Every personal finance argument, settled with data. Assumptions on screen, code public. If you think one of the assumptions is wrong, change it and re-run it. That is what this repository is for.

Most finance content asks you to trust the person saying it. This channel is anonymous, so that is not on offer. What is on offer instead is that every number in every video can be reproduced from this repository, from a dataset pinned by date and hash, using code you can read.

Don't trust it. Re-run it.

| # | Question | Video | Data |
|---|---|---|---|
| 1 | Lump sum or spread it out? | Lump Sum vs DCA: I Tested All 1,855 Months Since 1871 | episodes/ep01-lumpsum-vs-dca/results.json |
| 2 | What does a 1% fee actually cost, and is it ever worth it? | Paying 1% Costs More Than Panic Selling Every Crash for 30 Years | episodes/ep02-fees/results.json |
| 3 | What if you only ever bought at the market top? | I Only Ever Bought at the Market Top. It Cost 10%. | episodes/ep03-buying-at-the-top/results.json |
| 4 | Which numbers flip rent versus buy? | Rent vs Buy: One Assumption Flipped 82% to 31% | episodes/ep04-rent-vs-buy/results.json |

Requires Node.js 22 or newer and the .NET SDK 9 or newer. Nothing else. There are no npm packages to install, on purpose: asking you to verify a result should not also mean asking you to trust a dependency tree.

```

Rebuild the pinned dataset from source (optional, it is already committed)

node tools/build-shiller-snapshot.mjs

Check the dataset against an independent source

node tools/validate-shiller.mjs data/snapshots/shiller-2026-07-26

Run the simulation

dotnet run --project src/RunTheNumbers.Sim -- \
data/snapshots/shiller-2026-07-26 \
episodes/ep01-lumpsum-vs-dca/results.json

Episode 2

dotnet run --project src/RunTheNumbers.Sim -- \
data/snapshots/shiller-2026-07-26 \
episodes/ep02-fees/results.json \
--episode ep02

Episode 3

dotnet run --project src/RunTheNumbers.Sim -- \
data/snapshots/shiller-2026-07-26 \
episodes/ep03-buying-at-the-top/results.json \
--episode ep03

Episode 4

dotnet run --project src/RunTheNumbers.Sim -- \
data/snapshots/shiller-2026-07-26 \
episodes/ep04-rent-vs-buy/results.json \
--episode ep04

Independently reproduce and compare all episode 4 result fields

node tools/validate-ep04.mjs
``results.json` holds every number that appears on screen. The slides read from it
directly, so nothing in a video is typed by hand.

To see the slides:

node tools/serve.mjs 5173then open http://localhost:5173/render/ep01.html, render/ep02.html,
render/ep03.html, or render/ep04.html.

Investing a lump sum immediately beat spreading it over 12 months in 67.2% of
1,855 starting months since 1871.

Three findings that matter more than the headline:

  • The holding period is irrelevant.After the final purchase both portfolios own the same asset, so the outcome is fixed then and never changes. Measured drift across every horizon tested: 2.2e-16, which is floating-point rounding.
  • The drip's better worst case is real but short-lived.It shows at one and two years and cannot be distinguished from noise past three, while its cost is charged at every horizon.
  • In two of the eight worst starting months on record, the drip made the outcome worse.Both were 1998, the two highest starting valuations in that group.

These decide the answer. Change any of them and the number changes.

  • Real (inflation-adjusted) total return, dividends reinvested
  • Base case: idle cash holds its real value (0% real). Deliberately the generous assumption for DCA, and a fair long-run stand-in for T-bills
  • Sensitivity: idle cash earns 0% nominal, so it erodes with inflation. The lump sum then wins 72.0% instead of 67.2%, so the answer does not hinge on this choice
  • No taxes, no transaction costs, no fund fees
  • US large-cap index only (S&P Composite)
  • The holding period is measured from the month of the final purchase

Robert Shiller's monthly US stock market dataset, 1871-01 to 2026-06, from shillerdata.com.

Two things worth knowing if you go looking for this data yourself:

  • The copy at econ.yale.edu/~shiller/data/ie_data.xlsis astale mirror, frozen around 2023-09 as of this writing. Use shillerdata.com.
  • The Datecolumn is a float, so October 2025 is stored as2025.1and is indistinguishable from January once parsed. The month has to come from theDate Fractioncolumn instead.tools/build-shiller-snapshot.mjsderives it that way and cross-checks it against the strict monthly sequence.

The snapshot in data/snapshots/ is pinned by date with a sha256 in its
manifest.json. Simulations read the snapshot, never the network, so a result
published months ago still reproduces exactly.

tools/validate-shiller.mjs checks the parsed data against sources that share none
of its code: CPI against FRED's CPIAUCNS series across 1,361 months, plus
arithmetic checks on long-run real return and the 1929-33 drawdown.

tools/ data pipeline, capture, and assembly (Node, no dependencies) lib/cfb.mjs, lib/biff.mjs hand-written reader for legacy .xls lib/cdp.mjs frame capture over the DevTools Protocol src/ simulations (.NET) render/ slides and animations, plain HTML/SVG/canvas data/ pinned dataset snapshots episodes/ per-episode simulation output
Most of the code here was written by an AI coding agent (Claude), directed by a human. That is stated plainly because a channel whose entire pitch is "we show you the assumptions" cannot then be quiet about how the work is produced.

What that does and does not mean:

  • The agent wrote most of the tooling and the simulation code.The- .xlsreader, the capture pipeline, the chart rendering, the assembly step.
  • A human decides which questions get asked, which assumptions the base case uses, and what may be claimed on screen.Those are the choices that determine whether a result is honest, and they are not delegated.
  • Neither of those is a reason to trust the output.It is a reason to check it. The dataset is pinned and hashed, the loader is cross-validated against sources that share none of its code, and every number in every video is read from- results.jsonrather than typed. Whoever or whatever wrote a line of code here, the arithmetic either reproduces on your machine or it does not.

If you find a bug that changes a published number, open an issue. A correction is worth more to this project than a clean record.

The videos are produced from a private working repository that also holds scripts, drafts and production notes. This repository is a one-way mirror of the parts that let you check the work. It is not a fork and does not share history.

MIT. Take the pipeline, point it at your own question.