Our Phase I Proof of Concept (POC) project demonstrates how professional traders, using guidance from Gen AI transformer models, can outperform competitors at those stock loan trading desks using regression models. The Gen AI advantage stems from the ability of AI-powered models to detect complex trends from vast datasets and deliver precise, actionable recommendations in highly volatile market sectors.
In our 4Q24 proof of concept (POC), an ensemble of open-source Gen AI transformer models showed phenomenal predictive capabilities, successfully forecasting the a) direction and b) magnitude of closing rates for 55% of tested stock loans, including critical trend reversals. (By comparison, a series of coin tosses would be expected to achieve only 25% accuracy.) ARIMA and similar regression-based models offered little to no accuracy in identifying these pivotal shifts. Indeed, more than half (72%) of the correct forecasts from the AI ensemble were able to accurately predict turning points in rebate rates — an ability far beyond the scope of traditional models.
The transformative potential of Gen AI models lies not only in forecast accuracy but also in the measurable financial impact. When applied to a typical large-scale loan book, the AI models’ recommendations delivered substantial rebate savings for beneficial owners who lend securities. Traders emboldened with predictions of trend reversals by AI models accounted for over 60% of these savings.
Generative AI: Creating an Edge for Traders in the Training Stage
What sets transformer models apart is their ability to integrate historical data with current market conditions, focusing on relevant variables using advanced training methodologies like our Relevant Period Training (RPT).
These models dynamically recommend whether traders should hold or push inventory into the market based on expected rate changes.
These POC results underscore the strategic advantage of adopting AI to guide securities lending decisions. By offering traders a clearer, data-driven pathway to optimize loan pricing, these models can significantly enhance profitability while maintaining policy and regulatory compliance while mitigating risk. The findings validate AI’s potential to reshape the competitive landscape in securities finance, delivering both operational efficiency and financial outperformance.
The performance section of this report focuses on the outcomes of one-day forecasts generated during our Phase I Proof of Concept (POC). It is important to note that these forecasts were made without the benefit of adaptive learning — a capability that would typically be a prerequisite for such predictive modeling. Instead, each day’s predictions were based solely on static transformer models trained on prior data, without iterative updates or real-time recalibration. Despite this limitation, the models demonstrated remarkable accuracy in identifying optimal trading decisions, showcasing their potential to significantly enhance performance when combined with adaptive learning frameworks in future applications.
Pragmatic Accuracy: The Key to ROI in Securities Finance
Accuracy metrics from academia, such as mean squared error (MSE) or R-squared, provide useful statistical validation but often fail to capture the practical impact of AI-driven decision-making in securities finance. In contrast, pragmatic accuracy, as demonstrated in our proof of concept (POC), focuses on actionable outcomes — specifically, the ability to improve rebate pricing and generate measurable returns. Capital markets managers are not investing in AI for theoretical precision; head traders and division heads want models that deliver better trading decisions for tangible profit improvements.
By using RPT training and real-world performance metrics like the success rate of “Correct Holds” and “Correct Pushes,” our approach provides a direct link between AI model performance and ROI. This enables trading desks to justify AI investments, not as an experimental initiative, but as a necessary capital expenditure that enhances profitability, competitiveness, and operational efficiency — key factors for in-house data scientists in securing buy-in from capital markets leadership.
Implications for Performance Measurement in Securities Lending
Performance management in securities lending begins at the trading desk, where decisions on inventory management and loan pricing set the tone for the entire portfolio’s gains and losses. Traders armed with AI-enhanced models gain a critical advantage, as these tools not only refine rebate pricing strategies but also help prioritize opportunities with the greatest potential impact. By leveraging insights from deep learning models, trading desks can ensure optimal resource allocation, mitigate risks tied to suboptimal pricing, and align short-term actions with long-term portfolio goals.
Our POC results emphasize the importance of integrating advanced analytics into daily trading desk workflows. The findings show that AI-boosted recommendations outperform traditional models in both accuracy and financial impact, directly contributing to traders’ ability to make better-informed decisions. The use of these models fosters a proactive approach to performance management, enabling trading desks to anticipate market shifts, maximize returns on inventory, and ultimately deliver measurable value to both lending agents and beneficial owners. This strategic alignment at the trading desk level creates a ripple effect, improving the performance and competitiveness of the entire securities lending operation.
The Trader’s Role in Outperforming Benchmarks
Performance above the benchmarks starts — and ultimately ends — at the trading desk, where real-time decisions on whether to hold or push inventory shape the profitability of securities lending programs. A correctly advised trader not only captures better rebate rates but also sets up a sequence of downstream decisions that amplify performance. These include adjusting collateral allocations, optimizing recall strategies, and aligning loan pricing with broader market conditions. The impact of these decisions compounds over time, reinforcing the importance of informed execution at the trader level. Given that a purely random decision-making approach would yield only a 25% success rate in predicting the magnitude of changes for each hold/push scenario, achieving a 55% success rate through AI-driven forecasts is nothing short of remarkable. This demonstrates that traders equipped with advanced forecasting tools can significantly improve returns — not by replacing intuition and expertise but by enhancing them with precise, data-backed insights. By working with traders rather than against them, AI-driven models can help securities lending desks optimize performance without disrupting the fundamental role that skilled professionals play in the process.
These Phase I POC results show the potential of Gen AI forecasting models to deliver actionable insights and significant value to trading desks and securities lending operations. We’d like to thank all the lending agents and beneficial owners who provided suggestions during the study.
Beyond Big Data: Relevance Over Volume in AI-Driven Securities Lending
The notion that more data always leads to better models is a common misconception in AI applications for capital markets. While big data enables broad analysis, precise pricing in securities lending requires a more selective approach. Training models on vast but irrelevant historical data can dilute accuracy rather than enhance it. Instead, we emphasize the need for selecting relevant prior periods in RPT training — a pragmatic methodology that ensures AI models learn from the most meaningful market conditions. Our proof of concept results demonstrate that applying RPT-based training significantly improves model performance by focusing only on the data that best reflects current pricing dynamics. This fine-tuning process allows for more reliable and actionable predictions, making AI not just a theoretical exercise but a practical tool for trading desks.
The structured modeling approach of data engineers at Advanced Securities Consulting LLC can provide capital markets managers with accuracy metrics that give tangible proof of AI investments translating directly into competitive advantage. Our results make a strong case: desks that integrate AI-driven forecasting consistently outperform those relying solely on traditional regression models. This is no longer a question of whether to adopt AI—it is a question of how fast firms can implement it to avoid falling behind. To put it into perspective, ask yourself: Would you want to be the only Native American tribe migrating across western America in the 18th century without horses?
Just as horses revolutionized mobility and survival, Gen AI platforms are becoming essential for modern trading desks — empowering both traders and data scientists with the tools they need to navigate and dominate an increasingly complex market landscape.
Background of the Proof of Concept: A Rigorous Test Beyond Academic Standards
The public POC used our curated set of AI transformer models, in conjunction with our Relevant Training Period (RPT) methodology, to generate rebate rate forecasts that were publicly distributed via emails each morning. The POC demonstrated substantial success in a wide range of scenarios.
The 103/105 screening process used to select names for the POC ensured that only the most relevant securities entered our test pool. This methodology identified stocks exhibiting significant spread deviations between OBFR and loan rates, highlighting names where AI-driven pricing could have the greatest impact. After this initial screening, we applied an additional filter to isolate the most dynamic loans — those with the highest volatility in rebate rates and borrowing demand. The result was an extremely volatile sample set, far more challenging than the static, normally distributed datasets used in most academic research. Unlike controlled studies that test AI models under idealized conditions, our POC subjected AI forecasts to real-world market turbulence, where sudden reversals and liquidity shifts occur without warning. That our models consistently outperformed regression-based approaches in this environment proves their robustness and practical value for traders managing high-stakes loan books.
Key trading desk signals such as “Correct Holds” and “Correct Pushes” (explained below) outperformed their respective penalties due to “Incorrect Holds” and “Incorrect Pushes.” The test results indicate that lending agents using advanced models to guide their traders’ decision-making would have achieved better than average outcomes
These results validate the accuracy and utility of our AI-boosted forecasting methods for Push or Hold decision-making in maximizing rebate savings.
Key Insights from the Phase I Proof of Concept
- Correct Holds: Our model successfully forecasted closing rebates in scenarios where retaining a buffered position proved beneficial. Correct HOLD forecasts exceeded incorrect ones by a wide margin, highlighting decision accuracy, as illustrated in the chart above. In addition to forecasting the correct direction, the magnitude of successful forecasts in relation to the reported market averages underscored the models’ ability to maximize returns.
- Correct Pushes: Our ensemble of AI models provided accurate predictions for scenarios that warranted decisive action or upward rebate adjustment. Correct pushes further enhanced revenue streams, contrasting sharply with the penalties observed for incorrect pushes.
The models’ correct advice to Hold or Push far outweighed the minimal instances of Incorrect Hold or Push recommendations.
The Trader’s Role: Tuning AI for Market Realities
Gen AI tools and models, no matter how advanced, will require expert traders to act as the tuning fork that aligns predictive analytics with real-world market dynamics. AI models can process vast amounts of data and detect patterns that human intuition might miss, but they lack the living experience to interpret those signals within the broader context of market structure, liquidity conditions, and competitive flows. Without trader expertise guiding model adjustments, AI risks being just another overfit statistical exercise — technically sound but operationally flawed.
Traders understand the nuances of short squeezes, collateral stress, and shifting demand in ways that raw data cannot fully encapsulate. Their insights fine-tune AI models, ensuring that RPT training periods are selected with relevance, features are weighted with precision, and outputs are actionable, not just accurate. The result? A trading desk that isn’t just automated, but augmented — where AI amplifies trader instincts rather than attempting to replace them.
Summary of the POC Process: Phases I and II
Our Proof of Concept (POC) was designed to test the real-world efficacy of AI-driven securities lending forecasts through a two-stage process. For the first three months, we emailed daily forecasts for a single stock to senior practitioners, providing Hold/Push (H/P) recommendations alongside forecasted closing averages. This phase validated the models’ ability to anticipate rate movements in isolation, while engaging market experts in assessing forecast quality.
Predictive Framework: Our RPT-trained ensemble of Gen AI transformer models was able to leverage historical and recent data trends to focus their attention on the key variables that we, as subject matter experts, believed from experience would influence rebate rates.
Public Testing: A new security was selected each day from those that were screened for specific unit or price movements and appeared on our BWL each morning. Forecasted actions (Hold or Push) were shared with senior managers at lending agents through daily emails to validate the model’s real-world efficacy.
LUCY
- Our Watchlist identified a rapid increase in demand with new loan shares almost double their 20-day average. Returns also declined 50% compared to the average.
- New loans were 2-times LUCY’s return volume, signaling increased demand.
EVOK
- Our Watchlist GLr highlighted a skewed rebate distribution identifying significantly steeper rebate levels being paid by willing borrowers.
- New loans were 8x their 20-Day trailing average. New loans outpaced returns by over 15 to 1. At 100% capacity, there were little remaining shares available for loan.
In the second phase, we applied AI-generated forecasts to a test portfolio selected by a professional trader. Each morning, the trader received predictions on multiple securities, integrating them into his decision-making workflow. Unlike the first phase, this step simulated real market conditions, where multiple assets compete for attention, and capital allocation decisions are dynamic. By tracking how forecasts influenced trading strategies, the POC demonstrated that AI tools could deliver actionable insights — not just theoretical accuracy, but measurable improvements in loan pricing and execution.
Performance Evaluation: Outcomes were evaluated by comparing correct Holds and Pushes against incorrect actions, as reflected in the results.
The Beneficial Owner’s International Securities Finance & Collateral Management Conference featuring keynote speaker Ed Blount
Event details: https://events.imn.org/event/SecuritiesFinance/agenda