Open-source quant infrastructure for Korean equities — built so that the rejections get published too.
PhD in mechanical engineering (mathematical optimization) · ML & perception engineer @ Doosan Robotics by day.
The through-line is optimization: strategy search is an optimization problem, and every trial is paid for.

한국어 · English

Three things the data taught me, each with the code that found it:

• KRX price limits are not symmetric. The +30% cap is floored to the tick of the target price, the −30% floor to the tick of the base price, so the two moves differ by up to 0.26 points depending on where you start. → krx-quant-core
• 31% of the Korean stocks that ever traded are delisted. Drop them and a 2008 backtest loses 37% of its universe. The pipeline keeps them. → quant-airflow
• Pure noise passes "5 of 6 folds positive" about half the time. So a strategy has to beat its own randomized version, not a fold count. → swing-it

Three stacks, one shape — collect → store → research. Korean equities get the most of it: client libraries feed one Airflow pipeline into TimescaleDB, a shared core fixes the market rules, and the research and trading repos read from there. The search itself is treated as optimization with a bill attached: grid sweeps, Optuna TPE and a CMA-ES evolutionary engine write every config to a run ledger, Deflated Sharpe and PBO charge for the number of trials, and the out-of-sample window opens once.

%%{init: {"flowchart": {"useMaxWidth": false, "nodeSpacing": 12, "rankSpacing": 18, "padding": 6}, "themeVariables": {"fontSize": "11px"}}}%%
flowchart TB
    subgraph KR ["🇰🇷 Korean equities"]
        direction LR
        SRC["kiwoom-client<br/>krx-news-client<br/>krx-fundamentals-client"] --> AF["quant-airflow"] --> DB[("TimescaleDB")] --> RS["swing-it · daytrade-it · scalp-it"]
        KC["krx-quant-core"] --> RS
    end

    subgraph US ["🇺🇸 US equities"]
        direction LR
        SH["Sharadar"] --> DD[("DuckDB")] --> PR["portfolio-research · macro-sector-agent"]
    end

    subgraph CX ["🪙 Crypto"]
        direction LR
        EX["Exchange APIs"] --> CR["binance-quant-engine"] --> CI["crypto-quant-insights"]
    end

    KR ~~~ US ~~~ CX

    classDef source fill:#2563EB,stroke:#1E40AF,color:#FFFFFF
    classDef move   fill:#B45309,stroke:#78350F,color:#FFFFFF
    classDef out    fill:#059669,stroke:#065F46,color:#FFFFFF

    class SRC,SH,EX source
    class AF,DB,DD move
    class RS,KC,PR,CR,CI out

    style KR fill:#0F172A08,stroke:#64748B
    style US fill:#0F172A08,stroke:#64748B
    style CX fill:#0F172A08,stroke:#64748B

_{Blue — data sources · amber — collection & storage · green — research & engines. fin-checkup stands alone; scalp-it and daytrade-it are private.}

	Project	What it is
	kiwoom-client	Kiwoom Securities REST API wrapper — every domestic-equity endpoint, real-time WebSocket, sync + async, auto token refresh. Ships an MCP server so AI agents call the same API as tools, real orders opt-in only · pip install kiwoom-client
	krx-news-client	Korean market news and disclosures — DART filings and Toss Securities — as one schema, article bodies and per-company history included · pip install krx-news-client
	krx-fundamentals-client	Korean corporate fundamentals — DART statements and dividends, Naver valuation (PER, PBR, yield), screening — as Pydantic models. No standing server
	quant-airflow	The one pipeline behind both equity stacks: Korean prices, supply/demand, earnings, consensus, news and disclosures into TimescaleDB, with delisted stocks backfilled so downstream backtests aren't survivorship-biased. The US Sharadar rebuild into DuckDB is paused since the subscription lapsed
	krx-quant-core	Where the Korean-equity systems agree on market mechanics — tick sizes, asymmetric price limits, a dated transaction-tax schedule, order guards, DART disclosure classes, a kill switch — plus the validation statistics and a research layer (grid sweeps, Optuna TPE, BIPOP CMA-ES) whose run ledger records every trial and hard-locks the out-of-sample window. No strategies; all three consumers pin the same version
	fin-checkup	Risk-disclosure alerts (rights offerings, CB issues, audit opinions, delistings) routed to Telegram, plus a traffic-light financial checkup over DART & SEC EDGAR. Facts and measurements only — never a recommendation
	swing-it	KOSPI/KOSDAQ swing research on two axes: an observation screen of where money went by sector, and a judgment track where alpha hypotheses face walk-forward, random null controls, purged CV and Deflated Sharpe enforced as CI guardrails. The rejections are the product
	portfolio-research	US equity factor engine — strategy parameters walk-forward optimized on point-in-time, survivorship-bias-free data, with Deflated Sharpe & PBO charging for the search; tactical ETF allocation goes through the same gate · writeup
	macro-sector-agent	Top-down US sector-cycle research that asks which industry has been forgotten, not what to buy. An LLM judge argues from evidence at one narrow waist only; the machine never picks — it only excludes
	binance-quant-engine	Strategy-agnostic Binance USDT-M futures backtest & execution engine — zero lookahead, backtest↔live parity, exits placed as exchange-side algo orders at entry, optional MCP server
	crypto-quant-insights	Chart-backed research notes from validating a live futures system — crash stress tests, win rate vs. reward-to-risk, in- vs. out-of-sample divergence, Kelly leverage vs. what's actually run, backtest vs. live fills. The strategy stays private; the findings don't
	automated-stock-trading-systems	Backtester for Bensdorp's seven non-correlated trading systems (educational reimplementation) — runs offline on a synthetic universe, no data vendor, no API key

Strategies and parameters stay closed. Only structure and discipline are written down. Happy to walk through any of these on request.

Project	What it is
scalp-it	Korean intraday framework — live tick/orderbook collection, a pre-registered validation program (hashed pre-registrations, hard-locked out-of-sample window, measured once) and an order execution loop that now runs dry-run only
daytrade-it	Korean-equity news day-trading system — a daemon polls Toss news, Claude extracts facts only (subject, novelty, direction) and code does the scoring, behind a DART hard-severity gate that refuses entries when its cache is missing or stale. Runs in paper mode now; the replay engine, ledger and write-ups stay
quantbox	Binance USDT-M futures breakout/momentum system, traded live and currently paused. binance-quant-engine is the public extract with the strategies removed
crypto-pair-trading	First iteration of the crypto pair-trading framework — predecessor of quantbox
resume-private	Private résumé source (LaTeX)