Recently, Korea’s stock market regulator approved Interactive Brokers to offer US retail investors direct trading on Korean stocks for the first time.

When I read about the approval I had a pretty simple reaction.

A new persistent buyer cohort the size of US retail walking into an equity market that has been operationally walled off for decades should move some prices.

The trade is positioning before the buyers arrive.

The thing is, I do not actually cover Korean equities....I know Samsung and SK Hynix from working through the global AI chip supply chain. The other ~2,500 names in the Korean universe are essentially unknown to me.

So I created with Claude Code a ~75 page dossier going deep on what names to pay attention to...

(Link to the dossier if you don't want to read how I built this: Link)

This is exactly the kind of project AI is now genuinely good at. Three things mattered for this one:

AI lets you boil the ocean. Screening every name in a 2,555-stock universe used to be something even an entire sell-side team can't do. I did it in a weekend.
AI lets you get up to speed in a market you do not cover. The Korean AI value chain, K-defense, the Value-Up Program (Korea’s new corporate-reform initiative aimed at improving shareholder returns), the chaebol structure (the big family-controlled industrial groups like Samsung, Hyundai, LG, SK). Hours of guided research instead of months of relationship-building.
AI lets you read primary sources in a language you do not speak. The Korean financial press and DART filings (Korea’s equivalent of the SEC’s EDGAR system) are largely untranslated. Claude Code reads them in Korean and summarizes in English on demand.

Here’s the plan:

Pull the full 2,555-name Korean equity universe tradeable through IBKR.
Score every name on a 4-block flow-sensitivity framework I designed.
Validate the top of the screen against Korean primary sources.
Surface the 5 highest-conviction deep dives and 15 mid-tier names.

The "Prompt"

We are using Claude Code for this one! You can also use Codex if you're on the OpenAI ecosystem.

I structure research like this top-down. Steel man framework first, then working hypothesis, then the test instrument that operationalizes both.

The framework prompt is the test instrument.

The Framework

IBKR opening direct Korean trading to US retail is the kind of new-buyer event that historically pushes prices up in markets where foreign ownership has been thin.

Korean stocks have been institutionally accessible to global allocators for years. US retail capital has had no practical route in beyond a small set of US-listed Korean ADRs.

The Korean universe maps unusually well onto the narrative themes US retail consistently buys: AI chips (Korea has the global duopoly in HBM, the high-bandwidth memory inside every Nvidia GPU), robotics (highest industrial-robot density in the world), K-defense, K-content (K-pop, K-drama), and EV batteries.

The opportunity sits in the mid-cap layer where each marginal dollar of incoming demand moves price the most.

The Working Hypothesis

Names with low foreign ownership (under 15%), in retail-attractive sectors, with mid-cap market cap and meaningful but uncrowded daily volume, will outperform the broader Korean equity market over the next 6 to 12 months as US retail flows arrive.

Our Screening Framework

Block A: Flows Sensitivity (40%). Free float, foreign ownership headroom, turnover ratio, ADR competition.
Block B: AI / Thematic Exposure (35%). Direct AI revenue, value-chain position, Korean media and retail-investor signal.
Block C: Korea Structural Factors (15%). Index inclusion, Value-Up program participation.
Block D: Quality Filters (10%, can disqualify). Sub-$1M daily volume disqualifies. Holdco-discount and foreign-ownership-cap names get penalized.

Top names get put on our shortlist

How Claude Code Ran It

A Python data layer pulled the universe, scraped Naver Finance (Korea’s main retail finance portal) for foreign ownership, and pulled the Korean exchange CSVs for index constituents.

The deep dives and mid-tier profiles were written by twenty parallel Claude Code subagents, each pulling and summarizing Korean-language primary sources for a single name.

The framework prompt is what AI will not replace! The Python plumbing and multi-agent orchestration are commodity.

Btw ... you can use this same logic for any market :)

The Result

Full output here: Link

We also created a full excel: Korea_Universe_Screen_v1.xlsx

Here is the funnel:

The funnel runs in four cuts:

Universe (2,555 names). Every Korean stock tradeable through IBKR
Priority cohort (291 names). Every name in the five retail-attractive sectors.
Investable shortlist (20 names). Apply market cap $300 million to $30 billion, average daily trading volume of at least $3 million, foreign ownership 1 to 35 percent, no US ADR, no foreign-ownership cap, no holding company.
Top deep dives (5 names) and mid-tier profiles (15 names). Sector-diverse highest-conviction picks get institutional-depth profiles.

The most surprising finding from the screen is structural. The Korean equity market has a bimodal foreign ownership distribution that I have not seen in any other developed Asian market.

Mega-caps in the AI chip cluster sit at 35-55% foreign ownership. The mid-cap layer of the same value chain sits at 5-15%.

The flow asymmetry concentrates in that gap, and the IBKR-tradeable universe gives US retail direct access to it.

A second finding worth flagging. LG Energy Solution (Korean ticker 373220) screens at 5.1 percent foreign ownership for a $79 billion mega-cap, which looks like a structural anomaly until you realize LG Chem (the parent company) owns 81.8 percent of the stock.

Float-adjusted foreign ownership (the percentage out of the shares actually available to trade) is closer to 28 percent, similar to Samsung SDI. The structural headroom thesis still holds, but it is tied to LG Chem’s announced plan to step its parent stake down to 70 percent over five years, releasing incremental tradeable shares.

That is the kind of question the screen is good at surfacing.

The Top 5 Deep Dives

Hanmi Semiconductor (042700). Global market leader at 71 percent share in HBM thermal-compression bonders (the specialized machines that stack and bond those memory chips for Nvidia GPUs). Foreign ownership 7.3 percent for a $25 billion company, $217 million in average daily trading volume. Q1 2026 earnings on May 20.
Doosan Robotics (454910). Korea’s leading collaborative-robot (cobot) maker. Foreign ownership 3.4 percent, the lowest in the Korean cohort above $5 billion market cap. Doosan group backing, an Nvidia Agentic OS partnership, humanoid roadmap milestones in 2027 and 2028.
Doosan Tesna (131970). Mid-cap semiconductor test specialist. Market cap $1.8 billion, foreign ownership 8.4 percent. Just won a major AI inference chip test contract in late April with a potential 300 billion won revenue impact (roughly $220 million). Filed a Value-Up plan in April.
Hanwha Systems (272210). Defense electronics. Foreign ownership 8.6 percent is the cleanest defense play in the Korean cohort. The other major Korean defense names are already crowded with foreign holders (Hanwha Aerospace at 45 percent, Hyundai Rotem at 35 percent, Korea Aerospace Industries at 27 percent). Market cap $16 billion.
LG Energy Solution (373220). EV battery cell mega-cap. Float-adjusted foreign ownership of 28 percent has structural headroom from the LG Chem parent stake-down.

The 15 mid-tier profiles are in the dossier and span all five retail-attractive sectors. Highlights include Rainbow Robotics (humanoid pure-play, Samsung-controlled), HYBE (the K-pop label group behind BTS), HPSP (a near-monopoly in specialty annealing equipment for chip manufacturing), Cymechs (a hidden HBM beneficiary flagged by Korean research), and POSCO Future M (battery cathode materials).

The Methodology Lesson

The cost of producing institutional-depth research on a multi-thousand-name universe in a market you do not cover has gone from a six-week analyst project to a weekend. The cost of writing a good framework has not.

The framework is the alpha. The breadth is what AI now lets you do!

Personal

On the personal side, I have been running an experiment on myself. I started tracking my meals by taking pictures of them, and I am combining that with biometric data from my watch (daily steps, sleep, heart rate variability).

What AI is genuinely great at is multivariate confluence: finding patterns across a bunch of different data sources at once that no single human can hold in their head.

Next I am adding AI-driven blood work analysis to the stack.

Hit reply if you are running anything similar. Always interested in what other people are stitching together.

Talk soon,

-Dave

When you're ready, here's how we can work together:

Join Wall Street Prompt - On-demand AI learning
Live Corporate Training - Live AI for finance workshops

Get one high impact AI prompt every week