When a company's biggest customers matter more, does the stock follow?

1. Events. Every 10-K and 20-F of a US-listed company that discloses at least one customer worth 10% of revenue. Concentration is the sum of those shares (named and anonymous customers; sales channels and group companies excluded).
2. Change. Both years come from the same filing (current year and prior-year comparative), so nothing is used before it was public. Rising: +5 points or more. Falling: −5 points or more.
3. Returns. Stock minus S&P 500 (SPY), buy-and-hold, from trading day +2 after the EDGAR filing date to day +61. Prices under $1 excluded; extremes winsorised at 1% and 99%.
4. Chance. Labels are shuffled only among companies that filed in the same month (5,000 times), so a market move that hit everyone filing in March cannot create a result. 95% intervals resample whole months.
5. Rule. A signal only if p < 0.01, the sign is the same before and from 2021, and the spread is at least 1 point, enough to survive about 1% round-trip costs.

Group A is rising concentration (or the most concentrated third in the level row), group B falling (or least concentrated). Returns are log returns relative to the benchmark. Rows with fewer than 20 events per group get no spread.

• Survivorship: prices come from current tickers, so companies that were delisted or acquired since are missing; this flatters average returns but affects both groups.
• Only companies that disclose at least one 10% customer are in the sample, so the level test compares concentrated companies with less concentrated ones, not with unconcentrated ones.
• Market-adjusted returns do not remove size, industry or momentum effects; the IWM row shows the result against small caps.
• Gross of trading costs; many filers are small and illiquid.
• Returns are log returns, which sit below simple returns for volatile stocks (a 70% annual volatility costs about 25 points a year). Both groups carry this, so the spread compares like with like, but the absolute averages look worse than an investor's experience.

Changelog

• cc-v12026-10-18

  First run. Implementation error: groups of customers in the filings ("top ten customers", "two customers") were added to single-customer shares, which pushed 854 of 6,575 filings to 100%. Result: spread −0.19 points, p 0.88, no signal.

• cc-v1.12026-10-18

  Same design and rule. Customer groups excluded, and a year that a filing reports only as groups counts as unknown instead of zero (an intermediate run without that second fix gave −1.13 points, p 0.42). This is the published result.

Sample: 6,575 annual filings, 6,401 with prices (1,290 of 1,321 tickers found), 338 excluded under $1; 2,795 steady and 1,401 without a prior-year column. Prices: Yahoo Finance adjusted closes. Filings: SEC EDGAR XBRL. Run 2026-10-18.

Window: trading days +2 to +250 after the filing. Volatility check: the high group averaged 64% annualised volatility and the low group 69% after matching (73% vs 65% before). Concentration: 74% vs 13% of revenue from 10% customers.

Inside each volatility band

In calendar time

Monthly long-short mean +0.12% (Newey-West t 0.52); alpha after SPY and IWM-minus-SPY +0.29% per month (t 1.30), betas -0.12 and 0.13. On an average day 114 stocks long and 114 short, 2014-04 to 2026-07.

• Survivorship: companies delisted or acquired since are missing. If highly concentrated companies fail more often, the missing failures make the high group look better than it was.
• The short side is small, often illiquid stocks; borrow fees can be several percent a year and some names cannot be borrowed at all.
• Volatility and dollar volume are controlled; industry is not. Single-customer biotech and defence contractors cluster in the high group.
• Market cap is not in our data; dollar trading volume stands in for size.

Sample: 4,325 filings with a full year of prices after the filing and a year of history before it, 2,852 in the high or low third. Left out: 136 without prices, 282 under $1, 77 with less than 120 days of history, 954 without a price 250 days later (mostly filed in the last year). Run 2026-10-18.