[ blog · tutorial ]12 min read

How to Build an Investment Research Agent: Markets, Fundamentals, SEC & Economic Data in One API

Sarah Choy·نُشر في 16 يونيو 2026·قراءة 12 دقائق

الترجمة غير متوفرة بعد، يظهر النص بالإنجليزية.

An investment-research agent needs five different data layers — prices, fundamentals, filings, macro, and news — each normally a separate vendor, key, and schema. Here's how to wire all five behind one endpoint set, with working code and the cost math.

الخلاصة

• •A useful finance agent needs five data layers: real-time markets (prices), company fundamentals (statements), SEC filings, economic indicators, and news. Stitched from separate vendors that's 5 contracts, 5 keys, 5 schemas.
• •API Pick exposes all five as consistent JSON search endpoints — /search/markets, /search/financials, /search/sec, /search/economic, /search/news — plus /extract for full documents. One key, pre-shaped for LLM tool calling.
• •The agent pattern: run the relevant endpoints in parallel as tools, merge the JSON, and let the model reason over grounded data instead of hallucinating numbers.
• •Credit pricing is only-on-success: markets 120, financials 200, sec 120, economic 50, news 15 per call. A typical multi-tool research turn costs well under $0.01–$0.10 depending on depth.
• •Build-vs-buy: assembling Polygon + a fundamentals vendor + SEC EDGAR + FRED + a news API yourself is weeks of integration and 5 monthly bills; the single-endpoint route is a day.

The five-layer problem

Ask an LLM "is NVIDIA expensive right now?" and it will confidently invent a P/E ratio. The fix isn't a bigger model — it's grounding. A research agent that earns trust has to pull live, cited data across five layers, then reason over it:

• Markets — the current price, market cap, and how it's moved. Crypto, forex, ETFs, and the day's movers when relevant.
• Fundamentals — balance sheet, income statement, cash flow, dividends, and insider transactions. The "is the business actually healthy" layer.
• SEC filings — 10-K risk factors, 10-Q detail, 8-K events, earnings-call language. The qualitative dimension numbers miss.
• Economic indicators — rates, inflation, employment, GDP from FRED, BLS, World Bank, and the IMF. The macro backdrop every thesis sits inside.
• News — the timely catalyst: a downgrade, a product launch, a regulatory action.

Stitched from separate vendors, that's five contracts, five API keys, five rate-limit regimes, and five response schemas you have to normalize before a model can touch them. The integration is where finance-agent projects stall.

One endpoint set, five layers

API Pick exposes each layer as a consistent JSON search endpoint, so the agent talks to one key and one response shape:

• Markets Search — global & US equities, crypto, forex, ETFs, funds, commodities, and US market movers.
• Financials Search — balance sheets, income statements, cash flow, dividends, insider transactions.
• SEC Filings Search — 10-K/10-Q/8-K, earnings transcripts, equity statistics.
• Economic Data Search — FRED, BLS, World Bank, IMF, USAspending, Destatis.
• News Search — date-filtered news across major outlets.
• Extract — pull a full filing or article to clean markdown when a snippet isn't enough.

The agent architecture

Register each endpoint as a tool. When a question comes in, the agent decides which layers it needs, calls them in parallel, merges the JSON, and reasons over the grounded result. A ticker question hits markets + fundamentals + news; a "how is the sector positioned" question hits economic + news + a couple of comparables.

import asyncio, httpx, os

API = "https://api.apipick.com/v1"
HEADERS = {"x-api-key": os.environ["APIPICK_KEY"], "Content-Type": "application/json"}

async def search(client, path, query, **kw):
    r = await client.post(f"{API}/{path}", headers=HEADERS,
                          json={"query": query, **kw})
    r.raise_for_status()
    return r.json()["results"]

async def research(ticker: str):
    async with httpx.AsyncClient(timeout=30) as c:
        markets, fundamentals, filings, macro, news = await asyncio.gather(
            search(c, "search/markets",    f"{ticker} price and market cap"),
            search(c, "search/financials", f"{ticker} latest balance sheet and cash flow"),
            search(c, "search/sec",        f"{ticker} 10-K risk factors", end_date="2026-06-16"),
            search(c, "search/economic",   "US interest rates and inflation latest"),
            search(c, "search/news",       f"{ticker} latest news", end_date="2026-06-16"),
        )
    return {"markets": markets, "fundamentals": fundamentals,
            "filings": filings, "macro": macro, "news": news}

# Feed the merged JSON back to your LLM as grounding, with the source URLs,
# and ask it to synthesize — never to recall numbers.
context = asyncio.run(research("NVDA"))

Each result carries a source URL. Pass those through to the final answer so a human can audit every claim — and so the agent's output is citable, which is what makes it useful in a real workflow.

Build vs. buy

For one data type, going direct is reasonable. For an agent that needs all five and explores unpredictably, the single-endpoint route ships in a day and bills only when a call succeeds.

What this unlocks

The same five tools power more than ticker lookups: earnings-season briefing agents, sector screens grounded in fundamentals, macro-aware portfolio commentary, and diligence assistants that read the actual 10-K via Extract. The pattern is always the same — grounded tool calls, parallel retrieval, synthesis over real data with sources attached.

Start with a free key (100 credits, no card) and wire the five tools into your agent framework of choice. From there it's prompt-engineering, not plumbing.

بقلم

Sarah Choy

CEO, API Pick

سارة تشوي هي الرئيسة التنفيذية لشركة API Pick. تكتب عن بناء واجهات برمجية جاهزة للإنتاج لوكلاء الذكاء الاصطناعي وسير عمل نماذج اللغة.