Testing the shorting premium: what a rebalancing bug taught us
What this does not establish That the shorting premium does not exist anywhere. In this large-cap, survivors-only sample, with borrow fees measured on listed stocks only, the expensive decile did not underperform; small caps and delisted names, where the original study found the effect, were not in the data.
How it is usually sold
As a squeeze play or as a contrarian long: "short the most crowded shorts", or the mirror image, "buy the most hated names and wait for the squeeze". The pitch is usually a monthly swing strategy in US stocks, illustrated with two or three famous squeezes as proof. Those episodes are real. They are also exactly the variance that keeps the average from ever being distinguishable from noise. The usual instruments of the pitch are the short interest percentage (how much of a company's stock is currently sold short), the days-to-cover ratio (how many days of normal trading it would take the shorts to buy back), and paid "squeeze score" scanners that combine both into a single number with a proprietary recipe.
Usually sold with Short interest percentDays to coverPaid "squeeze score" scanners
Where we cast
A monthly portfolio long the cheapest-to-borrow stocks and short the most expensive-to-borrow ones, within US large caps. The idea comes from the literature on stock lending fees: when borrowing a share to sell it short is expensive, the market is saying that many people are betting against it and few are willing to lend. Ten years of data, about 650 stocks, positions bought or shorted in equal dollar amounts at the start of each month and held until the next rebalance, transaction costs on what is actually traded and, above all, the real borrow fee charged to the short leg day by day on the value of the position.
Why it looked promising
The premium compensates a concentrated risk: whoever shorts a hard-to-borrow stock is exposed to recalls and squeezes. That risk gets paid. If so, expensive-to-borrow stocks should underperform cheap ones, and a long-short portfolio should capture the gap. The loser on the other side is whoever buys the hard-to-borrow name without looking at the lending market. The authors who documented the effect measured it net of the borrow fee, so the question was never whether the premium exists gross, but whether it survives once the short leg pays its fee.
What we found
First, what we published in the original entry, and why it was wrong. Our first simulation reported that the short leg made money net of the borrow fee, more than twelve percent a month on average, and that only the statistics failed. An independent review of our code found that the simulation did not do what the rule said: instead of holding positions for the month, it averaged the daily returns of the basket with equal weights every day, which quietly rebalances every day for free, and it compounded the short leg in a way that turns a collapsing stock into a gain. One month did the damage: in July of 2026 a bankrupt ticker in the expensive decile fell by ninety percent day after day, and the faulty short accounting turned those collapses into a gain of more than two thousand percent for that month. Remove that month and the old numbers already told the story below.
Then, what the test says once positions are held for the month as the rule states. The long-short portfolio lost about 82 basis points a month, with a t statistic of minus 2.7 against a bar of plus two. The short leg, net of the borrow fee, lost 289 basis points a month. (Figures from the ledger-based accounting of the second independent review, 2026-09-08; the first corrected run said about ninety and three hundred and five: same verdict, same five failed criteria.) Under the randomized-ranking benchmark, every one of the two thousand random rankings did as well as ours or better: sorting by borrow fee did worse than sorting at random. The stretch after 2020 and the run with costs doubled were both negative. Every one of the five pre-registered criteria failed. The long leg was almost identical under both accountings, which is how we know the difference is the short leg and nothing else.
The original minutes, in Spanish (secondary evidence)
| Gate | v1 (acta, anulada) | v2 (enmienda 1) | v3 (libro mayor) |
|---|---|---|---|
| 1. L/S neto t_NW ≥ 2 | +699,5 · t 0,69 | −89,5 · t −2,89 | -81,6 pb · t -2,7 🔴 |
| 2. mono de ranking p < 0,025 | 0,281 | 1,000 | 1,000 🔴 |
| 3. pata corta neta de fee > 0 | +1.269,9 | −304,7 | -288,8 🔴 |
| 4. 2021→ > 0 | +1.269,0 | −112,2 | -103,3 🔴 |
| 5. costes ×2 > 0 | +694,0 | −95,1 | -87,4 🔴 |
Gate names, results and thresholds as written in the original test record. The English summary above is derived from the same rows. Units: pb = basis points per month; t_NW = t statistic with Newey-West standard errors.
What we learned
Three things. First: a simulator must do what the rule says, and that has to be checked, not assumed. The rule said "hold for the month"; the code rebalanced every day for free and compounded losses into gains. A single synthetic test, five stocks doubling and coming back, would have caught it in a minute. We now run that test, and others like it, before any entry is published. Second: one month can be the whole result. A bankrupt ticker with daily collapses is not a premium, it is a corner case, and the old average sat on it. The median month told the truth in both versions. Third: the universe matters as much as the idea. The original effect was measured across thousands of stocks, small and illiquid ones included, with data for the names that died. In large caps that survived, with the borrow fee measured only on listed stocks, the expensive decile did not underperform. That is not evidence against the original study; it is evidence about what this sample can and cannot show.
Questions you may ask
It loses, once the positions are held for the month as the rule states. The earlier number came from a simulation that rebalanced every day without paying for it and turned one bankrupt stock's daily collapses into a gain. We publish the correction because the point of this logbook is the method, and the method includes saying so when the machinery was wrong.
No. It shows that in ten years of US large caps that are still listed, with our costs and our rule, the expensive-to-borrow decile did not underperform. The original study found the effect mostly in small, illiquid names, including the ones that later disappeared. We do not have that data; until we do, this family stays closed for us and open for anyone with better data.
We did not test them, so we cannot say they fail. They read the same crowding the borrow fee reads, with a different thermometer. What this entry shows is that in this universe the thermometer did not order returns at all; a different thermometer would need its own pre-registered test before anyone should trust it.
Check your reading
Three questions to tell what this entry established from what it did not.
What was wrong with the first version of this entry?
The simulation did not hold positions for the month as the rule said; it rebalanced daily for free and compounded a bankrupt stock's collapses into a gain. The corrected numbers use the rule as written.
What does "every random ranking did as well or better" mean?
That under the benchmark, sorting by borrow fee did worse than sorting at random in this sample. It is a statement about outcomes under that procedure, not the probability that the idea is false.
Did this test show the shorting premium does not exist?
No. It shows the expensive decile did not underperform in large caps that are still listed. Small caps and delisted names, where the original study found the effect, were not in the data.
What would bring us back
Nothing, in this form. The cross-sectional bet on borrow fees, in large caps that survived, does not order returns even before the statistics are asked; the family of signals built on the lending market stays closed.
Data we do not have today: a small-cap universe with borrow fees for the stocks that were delisted. Without the ones that died, any measurement of this premium is biased, and the original effect lives exactly where that data is missing. If that data ever exists, it would be a new pre-registered test, judged against the attempts this family has already paid for.