High-Resolution Thematic Investing: Decomposing SEC Filings Into Words. Test Case: AI, the Pandemic, Remote Work and Tariffs. Talkers vs. Warners

Every listed US company files an annual report, the 10-K, each year. It has fixed parts. Item 1 describes the business. Item 1A, Risk Factors, lists what could hurt the company. Item 7 discusses the year's results. For any theme, the report puts each company in one of three groups.

Talkers mention the theme outside Item 1A: in the business description, in the discussion of results, or anywhere else except the risk factors. For a talker, the theme is part of what it sells, makes, buys or plans.

Warners mention the theme inside Item 1A. For a warner, the theme is a risk: a higher cost, a new competitor, a new law.

The silent do not mention the theme anywhere in the report.

A company that mentions a theme in both places is a talker and a warner at once, and for a theme that matters to a company this is the usual case. In 2025, 273 of the 292 S&P 500 companies that talked about AI also warned about it. NVIDIA wrote about AI more than any other company in nine of the ten years, and in three of them it also had the most AI in its risk factors. So for AI, a warning in the risk factors usually comes from a company that is deep inside the theme.

The overlap is different for each theme (Figure 2). For AI, the companies that both talk and warn grew from one in 2016 to more than half of the index in 2025. For the pandemic, almost every company talked and warned in 2021 and 2022. Remote work went the other way: in 2021 most companies that mentioned it talked about it, and by 2025 most of them only warned, while almost half of the index was silent again. Table 1 names the companies on each list in April 2025.

Figure 3 shows how a filing becomes a portfolio. A computer program reads every original annual report that the largest US companies filed from January 2015 to early October 2026: the SEC's large accelerated filers, about 2,000 companies a year with $700 million or more in shares held by outside investors, 24,132 reports in all. The S&P 500 companies are among them; the tests use the S&P 500, and the shares of reports quoted below use all of the large companies. The program finds the sections, and for each of 24 themes it counts the theme's words in the whole report and inside Item 1A. Each theme is a fixed list of words, written before anything was counted. No AI model reads the reports, and the word lists are published with the card. On any date, each S&P 500 company is placed by its latest report filed before that day, so a test only uses what was public at the time.

"The trend is your friend" is an old market rule. Before testing it, a quant has to say what a trend is. With the counts, a theme becomes three numbers for every company and every year: whether the company mentions it, where it mentions it (outside or inside the risk factors), and how often (mentions per 10,000 words). These numbers show that themes rise and fade in different ways. AI spread slowly for eight years, and in 2024 the share of AI-mentioning reports that put it in the risk factors went from about half to 85%. The pandemic was a standard risk line, became universal in 2021, and is now moving back into the risk factors only. Remote work rose and fell within five years. Of the 24 themes, Y2K appears in no report, and regulation was already in almost every risk factors section in 2015. The other 22 all moved toward the risk factors (Figure 4), partly because the risk factors section grew by about 60% in words while the whole report grew by 11%.

A word can also change its meaning. Written in capitals, AI can mean avian influenza in a food company's report, or Autopart International in a car parts company's report; for AI we use a stricter count that checks how each report defines the letters. We added that check after the first counts showed the false hits, before any test ran, and it applies to every year. Tariff means a tax on imported goods, but for a utility, a pipeline or a railroad it is also the price list a regulator approves. Our count includes both meanings: about a quarter of the 2016 tariff talkers were utilities and pipelines, and about one in six in 2025.

We chose AI, tariffs, the pandemic and remote work on purpose. They are the best-known themes of the last decade, and everyone knew they mattered; this paper does not claim to have found them. It shows what the counts can do with words the whole world knows. The full study will not start from famous words: it will use thresholds to find themes as they rise in the reports.

This paper also uses the simplest rules on purpose: one theme, one place in the report, one date a year. The database we are completing goes much further. It covers 20 years of filings, and it will let you build a theme from your own words and groups of words and set your own thresholds: buy a company when its mentions of a theme pass a level, sell it when the theme moves into its risk factors.

The tests here start in April 2016 because the reports in this test start in January 2015: on 1 April 2015, one company in five had not yet filed a report in the data, because its business year does not end in December. Apple, Microsoft and Cisco are examples. Each list is compared with all S&P 500 companies held the same way, with the same amount of money in each company: the equal-weight S&P 500. SPY, which holds the companies by their size, is shown beside it.

A company talks about AI when it sells AI or builds it into its products. In April 2016 only 12 S&P 500 companies mentioned AI outside the risk factors, most of them technology companies: NVIDIA, Microsoft, Meta, Salesforce, Qualcomm and Cisco among them. By 2025 there were 292 talkers, from every industry.

A company warns about AI when a competitor could use it better, when a new law could limit it, or when its own use of AI could go wrong. Until 2023 the warners were a small part of the index, among them Microsoft, Adobe and Mastercard, though already 77 of the 170 S&P 500 companies that mentioned AI at all. In 2024 almost every company that mentioned AI added it to its risk factors, and in 2025 there were 411 warners.

A company stays silent about AI when it has nothing to say about it. In April 2025 only 53 S&P 500 companies did not mention AI in their latest report. Most of them made chemicals, steel or fertilizer, produced oil and gas, ran utilities, owned buildings or ran industrial businesses: ExxonMobil, Nucor, Albemarle, Delta Air Lines and Berkshire Hathaway, for example.

The results (Figure 6, Table 2). $1 put in the talkers on 1 April 2016 became $5.40 by April 2026. In SPY it became $3.99, and in the equal-weight S&P 500 $3.33. The talkers beat the equal-weight S&P 500 in 8 of 10 years, most strongly in the first four years, when few companies used the word. One company carried the first year: in the year from April 2016, NVIDIA gained about 205% and made more than half of the talkers' 31% return, while the other ten talkers gained about 14%, a little less than the equal-weight S&P 500's 16%. In 2017 NVIDIA led again, and the other talkers were ahead too. The warners could be bought from 2018, the first year there were ten of them. From then they beat the equal-weight S&P 500 in 6 of 8 years, and their $1 became $3.50, against $3.01 for SPY. The silent were behind the equal-weight S&P 500 in 9 of 10 years; their $1 became $3.05.

All of the AI lists' lead over SPY came before April 2021. From April 2021 to April 2026, $1 in SPY became $1.75, in the warners $1.70 and in the talkers $1.64.

Tariffs are taxes on imported goods. A company talks about them outside the risk factors when they are part of how it competes. Nucor makes steel, and tariffs protect it from imported steel. Whirlpool asked the US government for tariffs on imported washing machines in 2017. Brown-Forman, the maker of Jack Daniel's, wrote about the tariffs Europe put on American whiskey from 2018. Nike and Harley-Davidson wrote about what tariffs cost them. Utilities, pipelines and railroads use the word's second meaning, their regulated price lists.

A company warns about tariffs when they could raise the cost of what it imports, or close a market where it sells. The warners are importers and exporters: retailers that buy from Asia, such as Gap, Ross Stores, Macy's, Hasbro and Dollar Tree; manufacturers whose parts cross borders, such as Eaton, Amphenol, Dover and Hubbell; chip and solar companies, such as Teradyne, Microchip and First Solar.

A company stays silent about tariffs when it sells services, and services pay no tariffs. The silent are banks, insurers, landlords, cruise lines, casinos, health insurers and software companies. In April 2025 they included Wells Fargo, State Street, Mastercard, Marriott, Carnival, Las Vegas Sands and Palantir.

The results (Table 3). The talkers and the warners did about the same: each was ahead of the equal-weight S&P 500 by a point a year or less, in 6 of 10 years. The silent did worst. They were behind in 6 of 10 years, and their $1 became $2.66, against $3.53 for the talkers and $3.55 for the warners.

Most of that difference came in one year. On 1 April 2025, the day before the US announced its new tariffs, 59 S&P 500 companies had not mentioned tariffs anywhere in their latest annual report. In the next twelve months they gained 0.3%. The equal-weight S&P 500 gained 15.7%, and SPY 18.1%. In the other nine years the silent were behind by less than a point a year on average, and ahead in 4 of them.

Before 2020, about 30% of the large companies' reports mentioned a pandemic, almost always inside the risk factors, as one disaster among many. The talkers were companies for which disease is part of the business: drug and vaccine makers such as Merck and Gilead, Zoetis, which makes vaccines for animals, the insurer Hartford, and the travel company Booking. Most companies were silent.

In April 2020, the filing date mattered. The biggest talkers were retailers whose business year ends around 1 February. They filed their reports in late March 2020, while their stores were closing: Ross Stores, TJX, Macy's and Home Depot. Carnival, the cruise company, filed its report on 28 January 2020 without one pandemic word. About a week later, one of its ships, the Diamond Princess, was put in quarantine in Yokohama. From 2021 almost every report mentioned the pandemic, so for four years there were too few silent companies to buy.

Results: the pandemic talkers and warners were each about a point a year ahead of the equal-weight S&P 500, in 6 and 7 of 10 years. The silent were behind in 4 of the 6 years they could be bought, mostly in the year from April 2020.

Remote work rose fast and then faded. Before 2020 it was in about 2% of the large companies' reports. In 2021 it was in 87%, and in 2026 in fewer than half. 2021 was also the first year of an SEC rule that asks companies to describe their workforce in Item 1, and many used it to say how their own staff now worked. The biggest talkers included Jack Henry, HP, Cisco and Microsoft. The warners put remote work in their risk factors: the office landlord Vornado already in 2020, the advertising group Omnicom every year from 2021. In 2021, when almost every report mentioned remote work, the silent were mostly companies whose work cannot be done from home, such as the oil producers Occidental and EOG, the supermarket chain Kroger and the grain trader Archer-Daniels-Midland.

Results: the remote-work talkers were ahead of the equal-weight S&P 500 in 3 of their first 4 years. In the year from April 2025, when 111 companies still talked about it, they were 10 points behind. The warners' average lead comes from one year: in 2020, 11 companies that already warned about remote work, mostly retailers and landlords, gained 23 points more than the equal-weight S&P 500. In 4 of the next 5 years they were behind.

AI: the talkers beat the equal-weight S&P 500 in 8 of 10 years and beat SPY over the ten years; all of the lead over SPY came before April 2021. The warners beat the equal-weight S&P 500 in 6 of 8 years from 2018. The silent were behind in 9 of 10 years.

Tariffs: the talkers and the warners each beat the equal-weight S&P 500 in 6 of 10 years, by a point a year or less. The silent were behind in 6 of 10, most of all in the year from April 2025.

The pandemic: the talkers and the warners were a little ahead, in 6 and 7 of 10 years. The silent were behind in 4 of the 6 years they could be bought.

Remote work: the talkers were ahead in 3 of their first 4 years and 10 points behind in the year from April 2025. The warners' lead comes from 2020 alone.

Only three of the twelve lists beat SPY over their years: the two AI lists, and the remote-work warners because of 2020 (Figure 5, Table 3).

AI is the one theme where the talkers clearly beat the equal-weight S&P 500, and they did it while the theme was young: their lead was largest from 2016 to 2019, when few companies used the word, in the first year mostly thanks to NVIDIA, and all of the lead over SPY came before April 2021. Remote work shows a theme on its way out: the companies still talking about it after it had faded did worst of the three remote-work lists. For tariffs and the pandemic, the talkers and the warners were a little ahead of the equal-weight S&P 500, and the silent were behind.

Lawyers write the risk factors to protect the company, so a warning could be read as fear. For AI, the warnings came mostly from companies that also talked about AI: in 2025, 273 of the 411 warners were also talkers. That fits a brag better than fear, but this test cannot tell the two apart, because the warners and the talkers are mostly the same companies; the companies that only warn are the test, and the full study will run it. The pattern is familiar outside the stock market too, though our data does not measure it: Anthropic, one of the leading AI developers, is also one of the companies that speaks most about the risks of AI.

The silent lists show which companies a theme does not reach. In 2025 the companies with nothing to say about AI were mostly in materials, energy, utilities, property or industry. The companies with nothing to say about tariffs were mostly banks, insurers and other service companies.

A theme also picks industries. In ten years when technology companies led the market, a theme that picks technology companies looks good, and a theme that leaves them out looks bad. These tests cannot separate a company's words from its industry. The 20-year database allows longer tests, other filings, and themes defined down to single words and sections.

Everything in this paper is in the lab, on the Annual Report Words card, for the S&P 500 on any day since 2016. Choose a word from 24 themes. Choose where to look: anywhere in the report, inside the risk factors, or outside them. Choose which companies to keep: the ones that use the word, use it a lot, use it for the first time, dropped it, moved it into the risk factors, or do not use it. Connect the card to the Held Forward Test to see what those companies did in the next year, next to SPY.

Some settings to try. Tariffs, anywhere in the report, does not use it, on 1 April 2025: the 59 silent companies of the tariff year. Climate and ESG, outside the risk factors, uses it a lot: the companies that make climate part of their business. The pandemic, inside the risk factors, for the first time, on 1 April 2020: 121 companies added it in the reports they filed in the first three months of that year. Remote work, dropped it, on 1 April 2023: 86 companies.

Twenty years of filings, your own words and your own thresholds come next, on the same card.

We count words; we do not read their meaning. A mention can be one line in a list of a hundred risks, and our tariff count also includes the price lists of utilities and pipelines. The word lists were fixed before counting and are published with the card. The risk factors sections grew by about 60% in these years, which pulls every theme toward them. In the first three years some companies had no prices and were left out: 33 of the 494 companies of the equal-weight S&P 500 in 2016, fewer in 2017 and 2018, and almost none from 2019. The further back, the more are missing, which is what happens when companies that were later bought or delisted have no prices in our data; Yahoo, one of the 12 AI talkers of 2016, is one of them. One company can also carry a small list: NVIDIA made more than half of the AI talkers' return in the year from April 2016. Each list is also a list of industries: the early AI talkers were mostly technology companies, and the silent lists came mostly from a few industries, so these tests cannot separate a company's words from its industry. Ten one-year holds are few, and twelve tests ran on the same ten years. The ten later tests were chosen after the two AI results were known. Some results rest on one year: the year from April 2025 for the tariff silent, the year from April 2020 for the pandemic silent and for the remote-work warners. In April 2020 the pandemic lists also sorted companies by the date they filed. A list with fewer than ten companies in a year was not bought that year, and that year is left out of its comparison: the AI warners before 2018, the remote-work warners before 2020 and talkers before 2021, and the pandemic silent from 2021 to 2024. The counts of reports cover the largest US companies. In each month from 2016, 98 to 100 percent of the S&P 500's members could be matched to the company that filed the reports; the others were spin-offs that had not yet filed an annual report and two banks that file with the FDIC instead of the SEC. 2026 runs to 6 October.

1. Cohen, Malloy and Nguyen (2020), Lazy Prices, Journal of Finance 75(3): 1371-1415. https://doi.org/10.1111/jofi.12885
2. Dyer, Lang and Stice-Lawrence (2017), The Evolution of 10-K Textual Disclosure: Evidence from Latent Dirichlet Allocation, Journal of Accounting and Economics 64(2-3): 221-245. https://doi.org/10.1016/j.jacceco.2017.07.002
3. Loughran and McDonald (2011), When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks, Journal of Finance 66(1): 35-65. https://doi.org/10.1111/j.1540-6261.2010.01625.x
4. US Securities and Exchange Commission (2020), Modernization of Regulation S-K Items 101, 103, and 105, Release No. 33-10825, 26 August 2020.
5. US Securities and Exchange Commission (2024), SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence, press release 2024-36, 18 March 2024.