Short Borrow Availability
A backtest that shorts whenever the signal says short is assuming a lender always shows up. In practice the names a short signal likes best, small floats and crowded shorts, are exactly the ones most often unborrowable.
Prerequisites: Market vs. Limit Orders
Shorting a stock in a backtest is a one-line operation: flip the sign, hold the position, done. Shorting a stock in real markets requires first locating shares to borrow from a lender, and that borrow is not guaranteed to exist. A backtest that never checks availability is implicitly assuming every short is a "general collateral" name that a prime broker can always source — which is precisely untrue for the names most short strategies want to trade, since low float and high existing short interest are both signals a short-selection model is drawn to and reasons a stock becomes hard to borrow.
Worked example: the crowded short that was never actually available
A short-interest-based signal ranks small-cap names by a mix of high existing short interest and deteriorating fundamentals, and shorts the top decile. Backtested gross short notional wanted at any time: $40 million across roughly 60 names.
Cross-referencing against realistic borrow availability for that universe:
| Name bucket | Share of desired short notional | Typically available to borrow? |
|---|---|---|
| Large-float, low short interest | 45% | Yes, general collateral |
| Moderate short interest | 35% | Usually, at a fee |
| High short interest, small float (the signal's favorites) | 20% | Often unavailable or capped |
Of the $40 million the model wants to be short, roughly $8 million sits in names that are frequently unborrowable at the size required — meaning the backtest is crediting P&L on positions that could never have been fully established. If those unborrowable names are disproportionately the ones with the strongest raw signal (which is common, since crowding correlates with conviction), the realized edge on the achievable portfolio is measurably lower than the backtested edge on the full ranked list.
Backtests of short strategies should apply a borrow-availability filter correlated with the same characteristics — low float, high existing short interest — that drive the signal itself, not a uniform assumption that every name can be shorted at full desired size.
Because scarcity and signal strength are correlated rather than independent, a flat "assume X% of names are unborrowable" haircut understates the damage. The unavailable fraction concentrates in the highest-conviction names, so it should be modeled as a function of the same signal, not subtracted as a constant.
Where a borrow-locate feed isn't available for the backtest period, a workable proxy is to cap position size as a declining function of short interest as a percentage of float, since that ratio is the single best predictor of whether a real desk finds the stock hard to borrow.
Related concepts
Practice in interviews
Further reading
- D'Avolio, The Market for Borrowing Stock