「Claudeが決算を読む」— 無料の財務データAPIとLLMで作る日本株AIアナリストを本にまとめたA developer compiled a book on building a Japanese stock AI analyst using free…

A Japanese developer has published a book documenting how to build an AI-powered stock analyst that pairs Anthropic's Claude large language model with freely available financial data APIs to automate the reading of corporate earnings reports. The project targets individual investors who want to apply large language models to fundamental analysis without paying for professional data terminals or research subscriptions, and it reflects a broader trend of hobbyist developers packaging practical LLM workflows into shareable guides.

At its core, the approach separates two tasks that were historically bundled together in equity research: gathering structured financial data and interpreting it. The data-gathering layer relies on public APIs, while Claude handles the interpretation, summarization, and narrative synthesis. According to the summary, the book walks through concrete steps so that a reader can reproduce the system rather than simply understanding it conceptually. That practical, reproducible framing appears to be the main value proposition for retail investors who lack the tooling or budget of institutional analysts.

The mention of free financial data APIs is significant because Japan has several accessible sources. J-Quants, an API service operated in connection with the Japan Exchange Group, provides historical stock prices, listing information, and financial statement data for Japanese equities, and offers a free tier suitable for individuals. EDINET, the Financial Services Agency's electronic disclosure system, exposes securities reports and quarterly filings, and provides an API for retrieving XBRL-tagged financial documents. These sources allow a developer to pull balance sheets, income statements, and cash flow figures programmatically, which is the raw material the LLM then analyzes. The article does not specify exactly which APIs the book uses, but these are the commonly referenced options in the Japanese retail-investing developer community.

On the LLM side, Claude is well suited to this kind of document analysis because of its comparatively large context window, which can accommodate lengthy earnings materials, and its ability to follow structured instructions for extracting figures and generating commentary. A typical pipeline likely involves fetching a company's financial data, formatting it into a prompt, and asking Claude to identify trends, compute or comment on ratios, flag risks, and produce a plain-language summary. Developers often add structured output constraints so the model returns consistent fields, and some chain multiple calls together to handle year-over-year comparisons or sector context. The book presumably covers prompt design and API integration as the connective tissue between the data source and the model.

This work sits alongside a growing set of tools and industry moves. Anthropic has been positioning Claude for agentic and analytical tasks, and its API is increasingly used in finance-adjacent automation. Comparable efforts exist across the ecosystem, including retrieval-augmented generation setups, frameworks such as LangChain or LlamaIndex for orchestrating data and model calls, and libraries like yfinance for market data outside Japan. Bloomberg's development of a finance-specific language model and the emergence of AI features in retail brokerage apps illustrate that automated financial interpretation is being pursued at both institutional and consumer levels. A self-published, step-by-step book occupies a distinct niche by making the technique approachable for non-specialists.

Several caveats deserve emphasis. Large language models can produce confident but incorrect statements, a failure mode known as hallucination, which is especially consequential when numbers and financial conclusions are involved. Users would be wise to treat Claude's output as a drafting or screening aid rather than an authoritative valuation, and to verify figures against the original filings. Free API tiers also impose rate limits and coverage constraints that may affect how comprehensively a system can operate. None of this constitutes financial advice, and the reliability of any generated analysis depends heavily on prompt quality, data freshness, and the reader's own judgment.

Even with those limitations, the project is a useful illustration of how accessible components can be assembled into a functional analytical workflow. By combining open financial data with a general-purpose language model, the author demonstrates a pattern that individual investors and developers can adapt to their own needs. Whether the automated analysis proves accurate enough for real decision-making will vary by use case, but the documented methodology lowers the barrier to experimentation and offers a concrete starting point for those curious about LLM-assisted fundamental analysis.