Measuring the Crowd: Excess Co-movement and Crisis-State
Underperformance in 13F-Inferred Hedge-Fund Crowding
In one line: Stocks crowded by hedge funds do fall harder in bad months, but not reliably enough to trade.
Abstract. I test whether stocks jointly held by hedge-fund-like 13F filers exhibit excess co-movement and elevated underperformance in market stress, using 53 quarters of Form 13F data (2013–2026). Both effects are present and statistically significant: residual pairwise correlation among co-held stocks is elevated relative to matched controls (Newey–West t = 4.09), and the most-crowded quintile underperforms the least-crowded by 142 basis points per month in the worst decile of market months (t = 3.90). Neither effect clears a margin-of-safety promotion hurdle calibrated near a raw t of 5. I therefore classify crowding as a risk measure rather than a tradable signal, and treat the non-promotion as the paper’s central result.
1. Introduction
The August 2007 “Quant Quake” demonstrated that positions held in common across leveraged managers can unwind together, transmitting one fund’s deleveraging into every co-holder’s drawdown (Khandani and Lo, 2011). This paper asks whether that channel is measurable ex ante in institutional holdings, and whether the resulting signal is strong enough to trade. The study was pre-registered before any filing was retrieved: one crowding measure, two claims, a hedge-fund filter fixed in advance, and the family of tests declared at n = 2 so that the multiple-testing correction was specified before results existed.
2. Data
The Securities and Exchange Commission publishes every institutional holdings report (Form 13F) as a structured data set. I take 53 quarters (2013 through mid-2026; 2.9 GB) and identify, among 11,750 filers, the 1,216 that behave like hedge funds: fewer than 250 positions and quarterly name-turnover above 25%, both thresholds fixed in advance. A stock’s crowding in a given quarter is the number of such funds holding it, applied only after the 45-day filing deadline to preserve a point-in-time information set.
3. Results
3.1 Excess co-movement
If many funds hold the same two stocks, those stocks should co-move beyond what sector, size, and the six standard factors explain (Anton and Polk, 2014). Across 3,312 co-held pairs, each matched to a control of the same sector and size decile that the crowd does not co-hold, residual correlation is higher by approximately 0.9 correlation points, with a Newey–West t-statistic of 4.09. The effect is real but economically small.
3.2 Crisis-state underperformance
In the worst decile of market months since 2014, the most-crowded quintile of the panel underperforms the least-crowded by 142 basis points per month, factor-adjusted (t = 3.90). Across all months the spread is −33 basis points. The crowd pays a small toll in ordinary conditions and a heavy one in stress, consistent with the deleveraging channel documented in 2007.
4. Promotion test and discussion
Both effects are statistically significant; neither is promoted. The promotion rule requires a claim to survive a margin-of-safety re-audit (gross return halved, costs doubled) at approximately a raw t-statistic of 5, a threshold in the spirit of the multiple-testing corrections of Harvey, Liu, and Zhu (2016). Neither 4.09 nor 3.90 qualifies. The stress result rests on fourteen months, the small sample the pre-registration anticipated, and a small sample remains a small sample however well it agrees with the prior. The finding therefore stands as pre-registered. Crowding is a risk measure, not a signal. It indicates to a risk desk where liquidity is narrow; it does not indicate to a trading desk what to buy.
| Claim | Effect | NW t | Outcome |
|---|---|---|---|
| Co-held pairs co-move (3,312 pairs) | +0.9 corr pts | 4.09 | held, not promoted |
| Crowded underperform in stress (14 months) | −142 bps/mo | −3.90 | held, not promoted |
5. Limitations
Three, recorded before and during the study rather than after. First, the panel is survivorship-biased (current constituents only), so crowding is counted among survivors. Second, the pre-registration specified industry matching; the available data support sector matching, and the deviation was recorded in code before results existed. Third, 13F filings disclose neither short positions nor swaps; an Archegos-type exposure is invisible by construction. The crowding that can be measured here is a floor, not a ceiling.
References
- Anton, M., and Polk, C. (2014). Connected Stocks. Journal of Finance, 69(3), 1099–1127.
- Harvey, C. R., Liu, Y., and Zhu, H. (2016). …and the Cross-Section of Expected Returns. Review of Financial Studies, 29(1), 5–68.
- Khandani, A. E., and Lo, A. W. (2011). What Happened to the Quants in August 2007? Evidence from Factors and Transactions Data. Journal of Financial Markets, 14(1), 1–46.