Measuring the Short Side: Daily Consolidated Short-Volume
Crowding as a Risk Measure, Not a Tradable Signal

In one line: I tested four ways to trade public short selling data. None worked. It measures risk, not profit.

Abstract. I test whether daily consolidated short-volume, published without charge by FINRA under Regulation SHO, carries cross-sectional information for US large-cap equities. On a panel of 2,006 trading days and 406 names (2018–2026), I define a short-volume ratio (SVR) as short volume divided by total volume and, following the informed-short-seller literature (Boehmer, Huszar and Jordan, 2010), fix the direction a priori: long low-SVR, short high-SVR. Four smoothed transforms are declared in advance and judged against a deflation-adjusted promotion bar of Newey–West t ≥ 2.96. Zero of the four clear it; the statistics range from −2.24 to 0.18. In the nine worst equal-weight market months, however, the most-shorted quintile outperforms the least-shorted by 220 basis points per month (t = 5.34), the signature of crowded shorts covering under stress. Because the universe is current index constituents (heavily-shorted names that survived their squeezes rather than those that collapsed and delisted), that estimate is upward-biased. I classify short-side crowding as a risk measure, not a signal.

1. Introduction

Short sellers are, on average, informed. Heavily shorted stocks subsequently underperform (Boehmer, Huszar and Jordan, 2010), and daily short-sale flow predicts returns at short horizons (Diether, Lee and Werner, 2009). Those results rest on monthly short interest or on proprietary order flow. FINRA publishes a free daily alternative, consolidated short-volume from the Regulation SHO tapes. This paper asks the narrow question whether that public, high-frequency proxy reproduces the same predictability in large-caps, or whether it merely reprices what is already crowded. The study is pre-registered: one measure, a family of four transforms fixed in advance, the trade direction fixed a priori, and the promotion bar set before any return was computed.

2. Data

The panel is FINRA daily consolidated short-volume (Regulation SHO), covering 2,006 trading days and 406 large-cap names from 2018 through 2026. Consolidated (CDN) tape coverage begins in 2018; this is a stated limit that confines the sample to the post-2018 window and excludes every earlier stress episode, so the regime evidence below rests entirely on the drawdowns within this window. For each name and day the short-volume ratio SVR is short volume divided by total reported volume, smoothed before use.

3. Results

3.1 Cross-sectional tests

Four transforms of SVR were declared in advance: the smoothed level over 21 and 63 trading days, the 21-day change, and a sector-neutral level. Each is sorted in the a-priori direction and evaluated against a deflated promotion bar of Newey–West t ≥ 2.96, the hurdle appropriate to a family of four. None clears it. The largest magnitude, the 21-day change at t = −2.24, still falls short of the bar; the two level tests are indistinguishable from zero. Short-volume crowding does not forecast the cross-section of large-cap returns in this window.

3.2 Behaviour in stress

The signal is silent on average but not in the tails. In the nine worst equal-weight market months of the sample, the most-shorted quintile outperforms the least-shorted by 220 basis points per month, with a Newey–West t-statistic of 5.34. The sign is the opposite of a tradable short. The crowded-short leg gains precisely when the market falls, the mechanical signature of forced covering as short positions are unwound under stress.

4. Promotion test and discussion

Only one statistic in the study clears the bar, and it points the wrong way for a signal. The crowded-short leg rises in stress rather than falling. That is a property of the state, not an edge a desk can harvest, and it is contaminated by survivorship in a way set out in Section 5. The finding is therefore held, not promoted. It mirrors the museum’s 13F crowding study, where an effect that was statistically real was retained as a risk measure rather than shipped as product. A risk desk learns from it where covering pressure concentrates; a trading desk learns nothing it can act on.

5. Limitations

The binding limitation is survivorship. The universe is current index constituents, so it retains only the heavily-shorted names that survived their squeezes; the names that collapsed under short pressure and delisted are absent by construction. The +220 bps stress estimate is measured on the survivors of exactly the event it describes and is therefore an upper bound, not a neutral figure. Two further limits are structural: consolidated coverage begins only in 2018, so the sample cannot see earlier crises, and short volume is a flow proxy for positioning, not a direct measure of short interest or of economic exposure through derivatives.