A 10-K filing runs 100+ pages. An earnings call transcript hits 8,000 words easy. Multiply that across a watchlist of 20 companies and you're buried before you start.

LLMs can read all of it in seconds. But "summarize this filing" gives you garbage. The difference between useful financial analysis and confident-sounding hallucination comes down to how you structure the prompt.

Here's what actually works, what fails silently, and the specific prompt patterns that produce output an analyst can trust.

The Accuracy Problem You Need to Understand First

Before you write a single prompt, know the failure modes. The FinTagging benchmark from Columbia and Georgia Tech found that LLMs extract financial facts at roughly 72% F1 score. Sounds decent. But when asked to link those numbers to the correct US-GAAP taxonomy concepts, accuracy dropped to 17%. On full taxonomy classification, top models scored 0%.

Translation: the model can find the number. It often can't tell you what the number means in accounting terms.

This matters because a prompt that says "extract revenue" might pull the right figure. A prompt that says "extract revenue recognized under ASC 606 for the cloud services segment" will get confused, give you a different line item, and sound completely sure about it.

The fix isn't avoiding LLMs for financial work. It's knowing where to use them and where to use structured data instead.

Use XBRL Data for Numbers, LLMs for Narrative

Every SEC filing contains XBRL-tagged structured data. Every revenue figure, every debt obligation, every share count is already machine-readable and tagged to its exact accounting concept. XBRL International pointed out that most AI tools ignore this data entirely, consuming only the HTML narrative portions. They compared it to "having access to a precisely catalogued library but choosing to read only the book covers."

The practical rule: never ask an LLM to extract a specific number from a filing's text when you can pull it from XBRL directly. Tools like EdgarTools (open-source, 2.3M+ downloads, MIT-licensed) expose this data through a Claude MCP server with 13 tools covering financial statements, insider trading, and company filings. No API key needed.

Use LLMs for what they're actually good at: reading the Management Discussion & Analysis section, identifying shifts in risk factor language, comparing the tone of guidance across quarters, and synthesizing qualitative information that doesn't exist in structured form.

Prompt Pattern 1: Earnings Call Sentiment Extraction

Earnings calls are pure narrative. No XBRL. No structured data. This is where LLMs earn their keep.

The mistake most people make: asking for a "summary." Summaries flatten the signal. What you want is structured extraction with explicit output schemas.

You are a financial analyst reviewing an earnings call transcript.

TASK: Extract management sentiment signals from this transcript.

For each topic discussed, provide:
- topic: The specific business area (e.g., "cloud revenue growth", "supply chain costs")
- management_tone: One of [confident, cautious, evasive, defensive, neutral]
- direct_quote: The exact sentence that best captures their stance
- analyst_pushback: Did any analyst challenge this point? Yes/No. If yes, summarize the exchange in one sentence.
- forward_guidance: Any specific numbers or directional language about future quarters

Return as a JSON array sorted by order of appearance in the transcript.

IMPORTANT: If management gives vague non-answers, label the tone as "evasive" -- do not soften it. Flag any topic where the CEO and CFO gave contradictory signals.

TRANSCRIPT:
{transcript_text}

This prompt works because it forces discrete categorization instead of prose. The management_tone enum prevents the model from hedging with paragraphs of qualifiers. The instruction to flag evasive answers counteracts the model's tendency to be charitable with corporate language.

Expected output (partial):

json [ { "topic": "cloud revenue growth", "management_tone": "confident", "direct_quote": "We saw 34% year-over-year growth in cloud ARR, and our pipeline entering Q3 is the strongest we've seen.", "analyst_pushback": "Yes. Goldman analyst asked whether growth was organic or driven by contract renewals with price escalators.", "forward_guidance": "Expects cloud mix to reach 45% of total revenue by Q4 FY2026" }, { "topic": "operating margins", "management_tone": "evasive", "direct_quote": "We continue to invest strategically in areas that will drive long-term value creation.", "analyst_pushback": "Yes. JPMorgan analyst pressed twice on timeline for margin expansion; CFO redirected to 'long-term framework' both times.", "forward_guidance": "None specific. Referenced 'operating leverage over time.'" } ]

That second entry is the kind of signal that gets lost in a summary. The CFO dodged the margin question twice. A human analyst would catch it live. This prompt catches it in the transcript.

Prompt Pattern 2: Risk Factor Comparison Across Filings

10-K risk factors are where companies quietly disclose what keeps them up at night. The sections are long, repetitive, and boring by design. But changes between years are where the real intelligence lives.

You are comparing risk factor disclosures between two annual filings for the same company.

PRIOR YEAR 10-K RISK FACTORS:
{prior_year_risks}

CURRENT YEAR 10-K RISK FACTORS:
{current_year_risks}

Produce a structured comparison with these categories:

1. NEW RISKS: Risk factors that appear in the current year but not the prior year. For each, quote the specific new language and assess whether it signals a known issue or a precautionary addition.

2. REMOVED RISKS: Risk factors present last year but absent now. For each, assess whether removal signals resolution or just consolidation into another section.

3. MATERIALLY CHANGED: Risk factors present in both years but with significantly different language. Quote the old and new versions. Flag any case where dollar thresholds, geographic scope, or regulatory references changed.

4. UNCHANGED BOILERPLATE: Risk factors that are essentially identical. List by topic only, no detail needed.

For categories 1-3, rate each item: HIGH / MEDIUM / LOW materiality based on potential financial impact.

The explicit categorization prevents the model from giving you a vague "the risk factors evolved" response. Asking it to distinguish between resolution and consolidation on removed risks catches a subtlety that simple diff analysis misses.

Expected output (partial):

NEW RISKS: - AI regulatory compliance (HIGH): "We may be subject to new and evolving regulations governing the development and deployment of artificial intelligence, including the Colorado AI Transparency Act effective June 2026." This is new language referencing a specific statute, suggesting active legal review rather than boilerplate.

MATERIALLY CHANGED: - Cybersecurity incidents (HIGH): Prior year referenced "potential unauthorized access." Current year adds: "including incidents involving third-party vendors and supply chain software." Scope expanded significantly, likely reflecting an actual vendor incident or near-miss.

Prompt Pattern 3: Multi-Filing Financial Narrative

This pattern works for building a longitudinal view across several quarters. Instead of analyzing one document, you're asking the model to find the story across documents.

You have excerpts from four consecutive quarterly earnings calls for {company_name}.

Q1: {q1_excerpt}
Q2: {q2_excerpt}
Q3: {q3_excerpt}
Q4: {q4_excerpt}

For each of the following, trace how management's position evolved across the four quarters:

TOPICS: {topics_list}

For each topic, provide:
- q1_position: One sentence summary of their stance
- q4_position: One sentence summary of their current stance
- trajectory: One of [improving, deteriorating, stable, reversed, unclear]
- credibility_check: Did any specific commitment from an earlier quarter get walked back? Quote both statements if yes.

The credibility_check field is the most important. Identify any case where management said X would happen and later admitted it did not.

The credibility check is what makes this prompt valuable. Executives are careful with language. Tracking where "we expect strong growth" in Q1 becomes "growth moderated due to macro headwinds" in Q3 reveals more than any single-quarter analysis.

When Prompts Aren't Enough: The Multi-Agent Approach

Single prompts hit a ceiling on complex analysis. Pathway's LiveAI system demonstrated this clearly: their multi-agent architecture scored 56% on the FinanceBench benchmark, compared to GPT-4 Turbo's 19% standalone. The architecture dispatches specialized agents (risk analyst, sentiment expert, fundamental analyst) with separate context windows running in parallel.

You don't need enterprise infrastructure to apply the principle. The concept translates to prompt chains: run your risk extraction prompt, your sentiment prompt, and your financial data query separately, then feed those three outputs into a synthesis prompt. Each prompt operates in a focused context instead of fighting for attention in one massive window.

The Honest Limits

Hallucination rates on complex financial documents hit 69-88% in studies tracked by Vectara and confirmed by MIT research. Models are, per MIT's findings, "most confident precisely when they should not be trusted."

This is not a reason to avoid LLMs for financial analysis. It's a reason to design prompts that produce verifiable output. Every prompt pattern above includes specific quotes, explicit categorization, or structured data that an analyst can cross-reference. The output format is the quality control.

The production consensus from teams running these systems at scale: LLMs are great at the edges (summarization, narrative detection, drafting memos) and unreliable at the core (precise numerical extraction, GAAP classification). Build your prompts accordingly.

Start Here

Pick one company on your watchlist. Pull their latest 10-K and the prior year's from EDGAR. Run the risk factor comparison prompt above. You'll find something the market hasn't priced in within 20 minutes.

That's the real value of prompt engineering for finance. Not replacing the analyst. Making the analyst faster at the part that matters: reading between the lines.

If your team is building financial analysis workflows and wants structured prompt engineering training, connect with Kief Studio on Discord or book a session at kief.studio/contact.

earnings calls

Prompting for Financial Analysis: Extract Insights From Earnings Calls and SEC Filings