Overview

A one-day development log of introducing an Expert Agent Team architecture into a KIS OpenAPI-based AI trading system. Covers the four-expert AI + Chief Analyst discussion simulation, a pure-Python technical indicator calculator, and three KOSPI200 data source swaps that ended in hard lessons.

graph TD
    A[KOSPI200 constituents] --> B[Volume/Change TOP50 intersection]
    B --> C[5 to 25 candidate stocks]
    C --> D[Daily candle data collection]
    D --> E[Technical indicator calculation]
    E --> F[4 Expert parallel analysis]
    F --> G[Chief Analyst discussion]
    G --> H[Trade signal generation]
    style F fill:#e8d44d,color:#333
    style G fill:#e74c3c,color:#fff

Expert Agent Team Architecture

The previous MarketScanner analyzed stocks with a single Claude call. This was replaced by a discussion structure: four specialists analyze from their own perspectives, and a Chief Analyst synthesizes their views.

The Four Specialists

The key is calling all four in parallel via asyncio.gather:

async def run_expert_panel(data_package: dict) -> list[dict]:
    experts = [
        ("Technical Analyst", "MA alignment/divergence, RSI, MACD ..."),
        ("Momentum Trader", "Volume surge, Stochastic K/D ..."),
        ("Risk Assessor", "ATR-based volatility, RSI overbought ..."),
        ("Portfolio Strategist", "Cash allocation, sector concentration ..."),
    ]
    tasks = [_call_expert(persona, focus, data_package)
             for persona, focus in experts]
    return await asyncio.gather(*tasks, return_exceptions=True)

Chief Analyst Discussion Simulation

Once four opinions are in, the Chief Analyst reviews the bullish/bearish ratio and makes a final call. The prompt is designed to evaluate the reasoning behind minority views, not just count votes:

# even 3 bullish vs 1 bearish can result in HOLD if the bearish reasoning is strong
prompt = f"""
Expert opinion summary:
{analyses_text}

When the vote is not unanimous, pay special attention to
the concerns raised by the minority opinion.
"""

Pure Python Technical Indicator Calculator

To eliminate external library dependencies (TA-Lib, pandas-ta), RSI, MACD, Stochastic, Bollinger Bands, and ATR were implemented directly.

def calculate_rsi(closes: list[float], period: int = 14) -> float | None:
    gains, losses = [], []
    for i in range(1, len(closes)):
        diff = closes[i] - closes[i - 1]
        gains.append(max(diff, 0))
        losses.append(max(-diff, 0))

    avg_gain = sum(gains[:period]) / period
    avg_loss = sum(losses[:period]) / period

    # Wilder's smoothing — exponential smoothing, not SMA
    for i in range(period, len(gains)):
        avg_gain = (avg_gain * (period - 1) + gains[i]) / period
        avg_loss = (avg_loss * (period - 1) + losses[i]) / period

    rs = avg_gain / avg_loss if avg_loss != 0 else float('inf')
    return round(100 - (100 / (1 + rs)), 2)

Wilder’s Smoothing is used because it’s more sensitive to recent values than a plain SMA, improving the timeliness of trading signals.

The KOSPI200 Data Source Saga

Three data source swaps in a single day. Here’s each failure and how it was resolved.

graph LR
    A["KIS API<br/>inquire_index_components"] -->|failed| B["KIS API<br/>market_cap"]
    B -->|30-item limit| C[pykrx]
    C -->|session cookie LOGOUT| D["NAVER Finance<br/>scraping"]
    style A fill:#ff6b6b,color:#fff
    style B fill:#ff9f43,color:#fff
    style C fill:#ff6b6b,color:#fff
    style D fill:#2ecc71,color:#fff

Attempt 1: KIS API inquire_index_components

❌ Not registered in domestic_stock.json → API call impossible

KIS OpenAPI’s inquire_index_components exists in the documentation but was never registered in the actual SDK. A ghost API.

Attempt 2: KIS API market_cap (fid_input_iscd=2001)

⚠️ Call succeeds but returns a maximum of 30 items

Even with a KOSPI200 filter (2001), only the top 30 market-cap stocks are returned. Not enough for screening all 200 constituents.

Attempt 3: pykrx

A popular Python library for pulling KRX official data. But:

❌ KRX endpoint returns LOGOUT without a session cookie

pykrx’s internal HTTP session sometimes fails to manage KRX server authentication cookies properly, causing the server to return only the text LOGOUT.

Final Solution: NAVER Finance Scraping

The most stable source turned out to be NAVER Finance:

def _fetch_kospi200_via_naver() -> dict[str, str]:
    session = requests.Session()
    session.headers["User-Agent"] = "Mozilla/5.0"
    session.get("https://finance.naver.com/")  # acquire session cookie

    codes: dict[str, str] = {}
    for page in range(1, 25):  # iterate 24 pages
        resp = session.get(
            "https://finance.naver.com/sise/entryJongmok.naver",
            params={"indCode": "KPI200", "page": str(page)},
        )
        pairs = re.findall(
            r"item/main\.naver\?code=(\d{6})[^>]*>([^<]+)",
            resp.text,
        )
        if not pairs:
            break
        for code, name in pairs:
            codes[code] = name.strip()
    return codes  # returns exactly 199 constituents

Key points:

• session.get("https://finance.naver.com/") must run first to acquire the session cookie
• indCode=KPI200 in entryJongmok.naver is the KOSPI200 filter
• Iterating 24 pages retrieves all 199 constituents
• Results are upserted into SQLite for a same-day cache with automatic next-day refresh

Market Scanner Pipeline

The final pipeline runs in four stages:

10 commits in a day, 2,689 lines added — the entire architecture migrated from a single Claude call to an Expert Team discussion system.

Quick Links

• sharebook-kr/pykrx — KRX stock data scraping library (not adopted due to session issues)
• NAVER Finance KOSPI200 — the final data source

Insights

The biggest lesson here is that financial data API reliability can only be verified by actually running it, not by reading the docs. KIS API had endpoints documented but missing from the SDK; pykrx had session management bugs that made it unsuitable for production.

The Expert Agent Team pattern is applicable to any AI system that needs to make decisions — not just stock analysis. The key is the Chief Analyst’s prompt design: evaluating the reasoning behind minority opinions, not just counting votes. Three bullish vs. one bearish can still result in HOLD if the bearish view is backed by ATR-based volatility data.

Pure Python technical indicator implementation fully eliminates the TA-Lib installation headache (C library dependency) while maintaining algorithmic accuracy like Wilder’s Smoothing. A valuable approach for projects with deployment environment constraints.