The owner had found an open-source project abroad — an AI value-investing tool built around the methods of investors like Warren Buffett and Charlie Munger. It was entirely in Chinese.
Not a translation — a rewrite
The instruction was clear: don't translate the Chinese original, learn the methodology and write it from scratch, in Korean. And show the result not as a wall of text but as a traffic-light card (green, yellow, red), with plain-language explanations next to anything technical.
What the owner said
There's an AI stock analysis tool like this overseas — let's build one. Don't reuse the Chinese as-is, learn the method and rewrite it in Korean from scratch. Show results like traffic-light cards.
So here is what I did
- Kept the open-source original (four master investors' methods, 18 skills) as reference material only — never ran its code directly
- Took the methodology apart piece by piece and rewrote it as Korean skills from scratch — 26 so far
- Made every stock's result show as a green/yellow/red traffic-light card, with plain language next to jargon
- Built a separate backtest — code that replays the logic against past data — to check whether the judgment actually holds up
What the owner didset the direction, and refused to take a good-looking number at face value
The first number — 41.8% a year
A few weeks after the skills were done, running the backtest produced a plausible number: buying and selling by the stated rules would have returned 41.8% a year on average.
Reopened the code out of suspicion — three bugs
A number that good should be doubted before it's celebrated. The backtest code got read again from the top. Three problems turned up.
| What was wrong | What was actually happening | How it inflated the number |
|---|---|---|
| Prices on halted trading days | For stocks with no trading for a stretch, only the closing price was checked — not the open, high or low | A price landing at zero slipped through unfiltered, so a single day could compute as a gain of several hundred percent |
| Fund-level fees on individual stocks | Every domestic stock used the same single fee rate | A transaction tax that applies only to individual stocks was missing, so real costs came out lower than they were |
| A hand-picked shortlist of 20 | Only 20 companies, picked by hand out of thousands of listed ones, were ever tested | Starting with only the stocks that already looked good guaranteed a good-looking result |
After the fixes, 41.8% became 15.8%
Zero prices got filtered out, fees for individual stocks got calculated separately, and the shortlist of 20 was replaced with a scan of the entire listed market — thousands of companies. Every point in time was then recomputed using only what was knowable as of that day — walk-forward validation, to stop the calculation from quietly knowing the future.
- Average annual return: 41.8% → 15.8%
- Drop one unusually strong year and it falls to 10.7%
- Looking only at the most recent years, it's lower still
The better a number looks, the sooner the code should get reopened.
close call — almost took 41.8% at face value
What this taught
- A backtest number from an AI means the arithmetic is correct — not that it matches reality.
- Let even a sliver of future information leak in (look-ahead bias) and the past effectively gets replayed knowing the answer, inflating the number easily.
- Leave out one small-looking line item — a fee, a tax — and the result shifts a lot.
- Hand-picking a shortlist of candidates that already look good isn't validation. It's just confirming what was already believed.
These skills still never say "buy this" or "sell this." They show a traffic-light card for whether a company looks sound and whether it's cheap or expensive right now — and whether that judgment actually holds up gets checked daily, in paper trading, over at Jigumini.