AI analyzing 10-K filings to extract predictive constraints

Part 2 in a series on using LLMs for financial research — see Part 1 on earnings releases

Compared to earnings releases, 10ks are really long, often 100+ pages. This makes it even more risky to use an LLM to summarize them for stock research, but it's such a timesaver people do it anyway.

If you simply ask ChatGPT or Gemini, it will produce an excellent seeming summary of a 10-K. A closer read shows that LLMs miss critical info a seasoned analyst would spot; and include a lot that looks important but is actually not relevant for predicting long-term stock outcomes.

As we did with earnings releases, we analyzed all the 10-K filings from every company in the S&P 500, and correlated them against high quality long-term forecasts of company fundamentals. We took the top 50 we deemed most interesting, and looked for patterns.

Key Takeaways

LLMs summaries of 10-Ks gravitate toward business descriptions, strategy statements, and ESG disclosures, which rarely matter for forecasts of revenues, margins, and payouts
LLMs don't reliably include seemingly-boring info that human experts would flag is important, in seven categories covered below
You can save hours of 10-K reading by using LLMs, but you need to use them wisely and provide them the right context, and we explain how below.

7 Types of Key 10-K Findings

Category 1: Regulatory Constraints With Dates and Dollars

Vague regulatory mentions ("subject to regulation") are worthless. But constraints with dates ("rate moratorium through September 2029") can fit right into long-term cash-flow models.

Found in 12 of 50 companies.

Example: Alliant Energy's $11B capital program hits 4-year rate freeze

When Alliant Energy disclosed their Iowa subsidiary operates under a retail electric base rate moratorium from October 2025 through September 2029—during the peak of an $11 billion capital program—this transformed a vague "rate agreement" into a modelable constraint. The specific end date enabled forecasting margin pressure through 2029 followed by recovery in 2030. Without these details, analysts might project either sustained margin expansion or permanent compression, missing the temporary nature of the constraint.

Category 2: Legal Exposure With Quantified Ranges

Earnings calls say "involved in litigation." 10-Ks say "$452M sought plus 40 similar pending claims with $221M insurance remaining."

Found in 15 of 50 companies.

Example: Universal Health Services's $180M verdicts with 40 more claims and shrinking coverage

Universal Health Services disclosed two massive jury verdicts totaling $180M with approximately 40 additional similar claims pending, while only $221M in insurance coverage remained for the relevant period. New policies beginning March 2025 exclude coverage for sexual abuse claims. This pattern—one large adverse verdict plus many similar pending claims—signals a systemic problem likely to recur. The specific quantification enabled modeling 50-150 basis points of direct margin impact plus elevated risk premiums.

Category 3: Customer/Supplier Concentration With Trends

Not just percentages but trends over time. "69% from Apple, up from 58%" tells a different story than "significant customer concentration."

Found in 18 of 50 companies.

Example: For Skyworks Solutions, Apple concentration grows to 69% despite diversification strategy

Skyworks Solutions disclosed Apple revenue concentration increased from 58% (FY2022) to 69% (FY2024)—contradicting their stated strategic goal of diversification. This multi-year trend, only visible in the 10-K's three-year comparison tables, revealed the strategy was failing. Similarly, Supermicro showed one customer represented 20% of sales but 44.8% of accounts receivable, signaling potential payment risk beyond the revenue concentration.

Category 4: Segment Economics and Business Mix

Corporate averages obscure reality. When low-margin segments grow faster than high-margin segments, average margins compress even if both improve.

Found in 22 of 50 companies.

Example: Qualcomm's 71% margin licensing shrinks as 30% margin chips grow

Qualcomm's 10-K revealed the QTL licensing segment has 71% operating margins versus 30% for QCT chips. As the chip business grows faster, blended margins compress despite operational improvements in both segments. TKO Group showed the IMG acquisition delivers 26% EBITDA margins versus 39% for core UFC/WWE. These segment-level details—especially multi-year trends—enabled forecasting margin trajectories as business mix evolves, typically driving 100-300 basis points difference in long-term margin forecasts.

Category 5: Working Capital Patterns and Cash Conversion

Revenue growth without cash generation is a red flag. 10-K cash flow detail reveals the truth before working capital constraints hit earnings.

Found in 14 of 50 companies.

Example: Supermicro's 110% revenue growth consumes $2.5B cash instead of generating it

Supermicro's 10-K revealed that despite 110% revenue growth, the company consumed $2.5 billion in operating cash flow (versus generating $664 million the prior year). The detailed cash flow statement showed $3.0 billion in inventory buildup and $1.3 billion in accounts receivable increases. This signaled either unsustainable growth requiring continuous external funding or quality of revenue issues. These patterns enabled forecasting significant growth moderation as working capital constraints emerge, while also influencing payout ratio forecasts by 10-30 percentage points.

Category 6: Reserves, Insurance, and Leading Indicators

Insurance markets and reserve trends predict problems before they hit the P&L. Rising self-insurance or coverage exclusions signal what's coming.

Found in 11 of 50 companies.

Example: Universal Health Services's self-insurance reserves triple to $79M as coverage excluded

Universal Health Services' self-insurance reserves increased from $25M to $79M, while new policies excluded sexual abuse coverage. RTX showed a $674M increase in asset retirement obligations from revised EPA rules. These are leading indicators of future margin pressure—companies and their insurers know about problems before they hit the income statement. Reserve trends typically signal 50-150 basis point margin changes and reveal operational quality issues.

Category 7: Targets vs. Reality and Governance Signals

Multi-year comparison of management targets vs. actual results reveals reliability. One year is luck; three years is capability.

Found in 19 of 50 companies.

Truist Financial: Targets 18-20% return, delivers 12.3%

Truist Financial targeted 18-20% return on tangible common equity but delivered 12.3% in 2024. Skyworks aimed for diversification but saw customer concentration increase. These multi-year gaps between stated targets and actual results—only visible by comparing the current 10-K's MD&A against prior year targets—reveal management capability and strategic execution. This assessment typically results in 5-15 percentile point reductions in reliability scores and wider forecast ranges.

What to Ignore

Eight information types consistently non-predictive (which LLMs over-weight):

Boilerplate Risk Factors — "We face competition" appears in every 10-K. Exception: New risk factors or substantially expanded discussion of previous boilerplate warrant attention.
Business Descriptions and Product Overviews — Unless you're unfamiliar with the industry or the company is adding/exiting business lines. The information is available elsewhere and rarely changes year-over-year.
Generic ESG Disclosures — Sustainability goals without specific capital commitments or regulatory requirements. Exception: ESG creating real costs (carbon taxes) or benefits (tax credits).
Management Bios and Standard Board Composition — What managers do matters more than their résumés. Exception: Recent turnover, compensation misalignment, related party transactions.
Historical Summaries Beyond 3 Years — Focus on 2-3 year trends unless conducting full-cycle analysis for cyclical businesses (peak-to-peak or trough-to-trough).
Forward-Looking Statement Safe Harbors — Pure legal protection with no information value.
Generic Competitive Landscape Discussions — "We compete on quality, price, and service" tells you nothing. Exception: Market share specifics, pricing dynamics, or win/loss trends.
Unchanged Accounting Policy Descriptions — Technical details rarely change year-over-year unless there's been a policy change or new standard adoption.

How to Prompt AI for 10-K Analysis

In the previous post on earnings releases, we recommended a single high quality prompt with the right documents as context.

Here, we are less sure this is optimal. 10-Ks are so long that doing multiple passes through, looking for different types of information, may outperform a single prompt.

If you do run a series of prompts on a 10-K, here's what we recommend you ask for.

As for passing context, unlike with earnings releases, 10-Ks often have enough historical data that adding additional filings, such as the previous year's 10-K, as context may not be as valuable. In our research, we sometimes provided more, and sometimes not, with unclear results.

Prompts

10-Ks are 200+ pages of dense technical information. General summaries lose critical details. Instead, use focused extraction prompts for each category.

Regulatory Constraints

Extract all regulatory constraints, rate agreements, export controls,
licensing requirements, consent decrees that have dates and/or dollar amounts. For each: (1) specific constraint,
(2) dollar impact if quantified, (3) start/end dates, (4) what triggers
changes. Flag constraints >2 years or >$500M.

Legal Exposure

Extract legal proceedings you deem as material that have numbers. Ignore 'ordinary course' language.
For each material case: plaintiff claims, dollar amounts sought, number
of similar pending cases, insurance coverage amounts, reserve amounts.
Flag multiple similar claims or insurance gaps.

Concentration Trends

Extract customer/supplier/geographic concentration tables for current
and prior 2 years. Calculate year-over-year changes. Flag concentration
>50% or increasing when strategy claims diversification.

Segment Economics

Extract segment revenue, operating income, margins for each segment over
3 years. Calculate margin by segment and growth rate by segment. Flag
low-margin segments growing faster than high-margin segments.

Cash Flow Quality

Extract operating cash flow, changes in working capital components
(AR, inventory, AP) over 3 years. Calculate cash flow/net income ratio
and working capital as % of revenue. Flag cash consumed despite
earnings growth.

Reserve Trends

Extract all reserve rollforwards (legal, warranty, environmental,
self-insurance, loan loss). For each: beginning balance, additions,
releases, ending balance. Extract insurance policy changes. Flag reserves
building faster than business growth or new coverage exclusions.

Accounting Changes

Extract significant accounting policies section. Identify changes in
useful lives, capitalization policies, reserve methodologies. Calculate
dollar impact on earnings. Flag any changes that increased earnings.

Targets vs. Actuals

Extract forward-looking targets stated in MD&A. Compare stated targets
vs. actual results. Calculate gap for each metric. Flag targets missed
by >20% or consistently over multiple years.

Conclusion: What's Unique to 10-Ks

Since you're using LLMs to read a lot of filings, you may wonder what you should emphasize in your research pipelines from 10-Ks vs other documents like 10-Qs or earnings releases.

The 10-K's unique value lies in three elements: binding constraints with specific dates (rate moratoriums, regulatory deadlines, contractual obligations), multi-year trends in footnotes (reserve buildups, concentration changes, cash conversion patterns), and management credibility assessment through comparing targets versus actuals. This information exists nowhere else—earnings calls discuss what management wants to emphasize, quarterly reports provide updates, but only the 10-K must disclose the complete picture with quantified constraints and three-year comparison tables.

The extraction approach prevents information loss that inevitably occurs with summarization. When you ask an LLM to "summarize this 10-K," it will produce an readable overview focused on what seems important—business strategy, market position, recent performance. But it will miss that one customer represents 44.8% of accounts receivable, that insurance policies now exclude certain coverage, that useful life changes artificially inflated earnings by $128 million, or that management has missed their ROTCE target by 500-700 basis points for three consecutive years.

In our experience, these boring, technical, highly specific details are precisely what determines long-term outcomes.