Across six European countries in March 2020, roughly 13 basis points of wider quoted spreads bought roughly 310 basis points of relief from maximum drawdown. Della Corte, Kosowski, Papadimitriou and Rapanos express both sides of the regulator's tail-risk trade in comparable units.
Where does the ban bite?
The theory builds on Diamond and Verrecchia (1987), using a Glosten-Milgrom sequential trade setup. Informed traders know terminal value, noise traders trade for liquidity, and a competitive market maker chooses bid and ask prices that leave expected profit at zero.
Two changes drive the result. First, the regulator decides whether to ban short sales while uncertain about how much the noise traders need to sell. The objective is to keep the probability of price falling below threshold c within confidence level x. Second, informed and noise traders own the stock with different probabilities, h_I and h_N. This ownership difference produces the cross-section.
A ban allows sales only by investors who already own the stock. When informed traders are over-represented among owners (h_I/h_N > 1), a sell order becomes more likely to carry information. Adverse selection rises, the market maker lowers the bid, and spreads widen. The ban also turns some intended short sales into no-trade events, reducing the likelihood of a very low print.
With a uniform prior on noise-trader selling, the paper derives a crash probability under the ban of exactly h_I^2/h_N times its unconstrained level. Tail risk falls when h_I^2 < h_N. That condition can coexist with the condition for wider spreads.
Institutional ownership proxies for h_I/h_N. The measure comes from Bloomberg as of December 2019, drawing on evidence (Boehmer and Kelley, 2009; Bai, Philippon and Savov, 2016) that stocks with high institutional ownership have more informative prices. The authors openly call the proxy imperfect. Their interpretation requires institutional ownership to be positively associated with sophisticated investors, while allowing individual institutions to remain uninformed.
The estimates
The sample uses Datastream daily bid/ask prices and total return indices across 17 European markets from 2 January to 2 June 2020. Stocks below $250m in market cap and observations with negative spreads are removed. The result is 1,922 stocks and 207,418 daily observations: 1,126 small, 545 mid and 251 large cap.
Austria, Belgium, France, Greece, Italy and Spain form the treated group. These countries enacted bans on 17 and 18 March and all lifted them on 18 May. The pre-period runs from 17 February to 16 March, followed by a post-period from 17 March to 15 April.
The headline Ban x Country estimate for spreads is 0.116pp without fixed effects (s.e. 0.040). It reaches 0.130pp with stock and time fixed effects, then 0.120pp after adding firm size and sovereign CDS controls (s.e. 0.048). The regression contains 77,667 stock-days. The overall Ban coefficient is 0.247pp, indicating that most spread widening in March 2020 occurred across markets, with the ban accounting for only part of the increase. Average spreads rose from 0.72% to 1.09% in ban countries and from 0.62% to 0.87% elsewhere.
Institutional ownership changes the picture sharply.
For the bottom tercile, covering ownership up to 41%, estimates range from 0.058 to 0.079pp and are insignificant in all four specifications. In the top tercile, above 68%, they run from 0.229 to 0.285pp and remain significant at 1% in all four. Each group contributes roughly 20,900 stock-days.
Mean returns showed no shift: the Ban x Country estimate is -0.073, with an s.e. of 0.070. Median returns fell 22bp at 1% significance. Maximum drawdown improved by 2.757pp before controls and 3.105pp after controls, both at 1%, while volatility declined by 1.20 to 1.26pp.
The tail result is strongest where the model predicts. The low-ownership tercile records +4.300pp (s.e. 0.987), compared with +2.077pp (s.e. 1.116) for the high-ownership tercile. The high-ownership estimate is significant only at 10%, with a t around 1.9, and the paper prints the standard error plainly. In the matched sample, the corresponding estimates are +5.502pp and +2.522pp on 552 and 550 observations.
Beber and Pagano found around 198bp of spread widening during covered bans in 2008-09, reduced by 65bp when short positions required disclosure. This paper estimates 11 to 13bp. The authors explain the gap through their exclusion of micro and nano caps below $250m, which are less liquid and likely to amplify the effect, along with Europe's prohibition of naked shorting since 2012.
Identification stays thin
Six countries supply the full treatment variation. The 1,922-stock sample produces many stock-days from very few independent policy decisions, while baseline standard errors are clustered by stock and calendar date.
Country-time clustering appears in the appendix (Table A.8). Within the high-ownership panel, significance remains at 1% in the fixed-effects columns. Errors are 0.069 and 0.074, versus 0.079 and 0.083 under the baseline clustering. The no-fixed-effects column drops to 5% when its error increases from 0.066 to 0.090.
The 18 March enactment date coincides with the ECB's EUR 750bn PEPP announcement. For its instrumental variables exercise, the paper uses lagged sovereign CDS and the Duprey, Klaus and Peltonen (2017) Financial Stress Index. First-stage coefficients are 0.005 and 0.261. The Kleibergen-Paap Wald F is 40.9, with Hansen J p-values of 0.59, 0.25 and 0.31. Second-stage ban estimates range from 0.126 to 0.135pp.
The diagnostics pass. Even so, the instruments measure country-level stress during a period when that stress was directly shifting country-level spreads. Their exclusion restriction depends on contemporaneous spreads having already incorporated all of that stress.
A chart and a placebo division of the control countries support parallel trends. For spreads, placebo estimates range from 0.041 to 0.050 in the low-ownership panel and from -0.070 to -0.076 in the high-ownership panel. Both become -0.014 after controls, and none is significant. Drawdown estimates are +2.896 and +2.200, with standard errors of 2.84 and 2.87, based on 200 and 202 observations. We did not find a formal pre-trend coefficient plot. Given that placebo n, the exercise has power against large placebo effects and little beyond them.
Hypothesis 1 says regulators impose bans where institutional ownership is low. A bar chart supports it, though the paper describes this as suggestive evidence and identifies its own three counterexamples: Switzerland, Germany and Denmark. The authors checked ownership data quality in those markets and found no explanation.
Treatment belongs to the country level across six countries, while ownership terciles are pooled across countries. The split may therefore capture country composition alongside informedness. An ownership split that holds country fixed would settle the issue, and that is the single result that would change my reading. The model predicts outcomes from each stock's owner mix. At present, the cross-section identifying that mechanism remains entangled with the cross-section determining treatment. The December 2019 snapshot cannot react to the shock.
For the desk
Liquidity is measured through end-of-day quoted spread. The paper has no depth, no effective spread, no volume and no intraday data. It leaves unanswered the sizing question a desk faces during a ban: how much can trade before the book moves? The authors cite Goyenko et al. (2009) for treating quoted spreads as closely related to actual transaction costs. Their static, single-asset model also stays silent on spread dynamics, cross-asset effects, portfolio rebalancing and what happens when bans are lifted.
The tradeable claim is narrower than the broad trade-off framing. For low-ownership names in the bottom tercile, with ownership up to 41%, a ban is associated with substantial drawdown relief (+4.3pp) and a spread cost statistically indistinguishable from zero. For high-ownership names in the top tercile, above 68%, the cost is 23 to 29bp of quoted spread for about 2pp of drawdown relief, clearing only a 10% threshold. The paper never reports the size composition of either tercile.
We could not test this ourselves. The design requires a European cross-country universe, country-level ban rules, stock-level quoted bid-ask data, stock-level institutional ownership and sovereign CDS. We have US OHLCV bars. No US adaptation recovers the experiment because its identifying regulatory variation is European.
The drawdown result is the paper's contribution. The authors frame it similarly: earlier evidence showed that bans damaged liquidity and failed to support prices, while their results add lower tail risk in ban countries, especially among low-ownership stocks. Beber and Pagano (2013) studied spreads and prices. Bessler and Vendrasco (2021) found the same deterioration in liquidity under these 2020 bans, while Beber, Fabbri, Pagano and Simonelli (2021) found that bans raise default probability and volatility. Bris et al. (2007) reported less negative skewness under short-selling restrictions, mainly at the market level and with limited stock-level support.
This paper's step is to show the left tail moving stock by stock: 310bp on average and 430bp where the model predicts the largest effect.