Rethinking Short Interest in the Age of AI

By: Mel Sutton, EquiLend Interview with ALTALLO - Founder

Published: 2026-03-10 · Read time: 5 min · Category: Interview

Nancy Allen of EquiLend explains how machine learning is transforming short interest data from a lagging regulatory snapshot into a daily, actionable signal for investors and risk teams.

[HOST]: Traditional US short-interest data is biweekly and backward-looking. How material is that lag in today's markets, and what risks does it create for investors?

[GUEST]: The traditional delay in short interest reporting is a vulnerability in markets built for real-time execution and rapid capital rotation. In environments shaped by dislocations like the Flash Crash and positioning shocks such as the GameStop short squeeze, short exposure can build or unwind in hours, meaning backward-looking data can materially misstate crowding and squeeze risk. The result is real downside: mispriced volatility, underestimated gap risk into catalysts, flawed risk models, and investors entering trades based on positioning that may have already reversed.

[HOST]: For years, market participants have used securities lending balances as a proxy for short interest. Where does that approach work, and where does it break down?

[GUEST]: Securities lending balances are a strong real-time proxy for short interest; as most short sales require borrowing shares, making lending data a near real-time reflection of short positioning rather than a delayed regulatory snapshot. Lending on loan quantities are especially valid for hard-to-borrow securities where borrow demand is primarily driven by directional short sellers and loan balances closely track true bearish positioning. Unfortunately, this approach becomes less reliable in easy-to-borrow securities where abundant supply and low fees allow ETF hedging, market making, and arbitrage activity to disassociate loan balances without signaling strong conviction.

[HOST]: You've described the challenge as separating "directional short demand" from the mechanics of financing and settlement. What does that noise actually look like in practice?

[GUEST]: The "noise" in securities lending balances comes from activity that has little to do with outright bearish conviction. Prime brokers frequently borrow to optimize balance sheet usage, manage capital charges, or facilitate internalization, moving inventory across entities to net exposures. This activity can inflate loan balances without any new directional short being initiated, or in the case of internalization, lending activity can be muted entirely.

ETF creation/redemption hedging, options market maker delta hedging, index and merger arbitrage, settlement fails coverage, and collateral transformation trades all generate borrow demand tied to financing or operational mechanics rather than a fundamental view on the stock.

In the data, this noise shows up as borrow spikes around rebalance dates, corporate events, or quarter-end balance sheet windows. Elevated balances in easy-to-borrow securities with low fees, or temporary utilization increases that reverse quickly are patterns that reflect plumbing and capital efficiency, not true short conviction.

!Visa Inc (V US) Short Interest - Predicted SI vs On Loan and Exchange SI Quantities

[HOST]: Without revealing proprietary detail, what does the machine-learning model actually analyse? What signals matter most in predicting real short exposure?

[GUEST]: Over 10 years of first in class securities financing data from the largest participants in the industry act as the backbone of the Predicted Short Interest data. By leveraging over $4.2 trillion in daily loan balances and $47.2 trillion in daily inventory across 215,000+ securities, our machine-learning model highlights correlations within the traditional short interest data. From there, we've created multiple models that analyze securities that track well with existing securities finance metrics and those that have a higher dependence on external datasets, like publicly available daily short volumes, security prices and public float changes, to generate accurate predictions.

[HOST]: In names with heavy financing activity, borrowed shares can exceed reported short interest. How does your model distinguish between bearish conviction and routine balance sheet activity?

[GUEST]: Securities involved in financing transactions are generally straightforward to identify - typically major index constituents with low fees or borrowing costs, high market capitalization, and low utilization. By using a dedicated model tailored for high market capitalization securities, our forecasts capture security-level loan balance movements around key securities lending events, such as dividend record dates and quarter-end balance sheet activity. This approach improves accuracy by avoiding the over- or understatement of short interest forecasts.

[QUOTE]: Based on our testing, all of our models indicate an R-squared score of 96% and up, indicating the observed securities lending and external data sources have a strong linear relationship with the predictions.

[HOST]: How closely does Predicted Short Interest track exchange-reported data once it's published? And where do you see the largest divergences?

[GUEST]: Based on our testing, all of our models indicate an R-squared score of 96% and up, indicating the observed securities lending and external data sources have a strong linear relationship with the predictions. The Mean Absolute Error (MAE) for larger capitalization, easy-to-borrow securities was 1.33 million shares while many of the securities in this pool exceed 1 billion shares outstanding. Within the hard-to-borrow, or small capitalization model, our predictions had an MAE of 314,000 shares where the security pool generally ranges from 10 million to 100 million shares outstanding. While we have not observed major divergences in the predictions, we are regularly retraining our models as new traditional short interest data is available and will continue iterating our inputs in conjunction with our client partners.

[HOST]: If investors can see AI-derived short exposure daily, rather than waiting for official data, how does that change trading behaviour? Does it reduce information asymmetry? Accelerate squeezes? Compress alpha?

[GUEST]: Providing daily, ML-derived short exposure gives investors earlier insight into market positioning, reducing information gaps. This can speed up reactions to crowding and squeeze risk as participants adjust more quickly. Over time, however, the advantage may diminish as these signals are priced in faster, shifting the edge from access to interpretation and execution speed.

[HOST]: Who stands to benefit most from this kind of real-time insight: hedge funds, long-only managers, prime brokers, or risk teams?

[GUEST]: Daily visibility into short-interest dynamics enables market participants to better capture alpha and manage risk by identifying shifts in market sentiment earlier. For active investors, this data can highlight crowding, changing borrow demand, and emerging squeeze risk, informing positioning and timing decisions. Risk managers gain an earlier signal of concentration and liquidity stress, while intermediaries can proactively manage exposures and balance sheet usage. Long-only investors benefit from improved awareness of sentiment extremes that often precede volatility. While the specific use cases vary by role, timely short-exposure insight serves as an early indicator of market sentiment, enhancing decision-making, risk control, and capital efficiency across the securities lending ecosystem.

[HOST]: There's a lot of scepticism around financial AI models. What gave you confidence that this approach could meaningfully outperform raw lending data?

[GUEST]: EquiLend Data & Analytics has been creating multi-factor models for the securities lending community for over a decade. As with any of the fields we've created in the past, our focus is always building accurate and actionable data. Using modern practices like AI and Machine Learning allows us to reduce human bias and process far more information than prior approaches while maintaining the same disciplined governance and risk controls users can expect from the Data & Analytics team.

[HOST]: Do you see AI-derived short interest as complementing regulatory reporting, or eventually challenging the relevance of traditional exchange-reported figures altogether?

[GUEST]: Traditional exchange-reported figures will always act as the golden source of short interest as the information originates from short sales reported by broker-dealers. But as our Predicted Short Interest solution continues to improve, EquiLend plans to provide the most timely and accurate insights that the industry demands.