LEMs: A Primer On Large Execution Models
Rémi Genet § , Hugo Inzirillo § · 2025
Read LEMs: A Primer On Large Execution Models by Rémi Genet § , Hugo Inzirillo § (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-25211.txt.
Introduction
Rémi Genet and Hugo Inzirillo introduce Large Execution Models, neural networks built to schedule trade orders. Their paper appeared in 2025. Executing a big order well means slicing it over time to limit market impact. Older neural methods assumed a fixed time window. The new models allow the window to stretch between set bounds, as in share buyback deals.
One shared pipeline reads market data and builds context. Separate small networks then decide allocations for each case: fixed amount or fixed value, buy or sell. Tests on crypto markets within a day and on big United States stocks over days showed the models beat standard benchmarks. The authors argue one unified model can serve many execution desks at once.
Catalog record
| CREATOR | Rémi Genet § , Hugo Inzirillo § |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 72 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-25211
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
LEMs: A Primer On Large Execution Models Rémi Genet § , Hugo Inzirillo § Abstract This paper introduces Large Execution Models (LEMs), a novel deep learning framework that extends transformer-based architectures to address complex execution problems with flexible time boundaries and multiple execution constraints. Building upon recent advances in neural VWAP execution strategies, LEMs generalize the approach from fixed-duration orders to scenarios where execution duration is bounded between minimum and maximum time horizons, similar to share buyback contract structures. The proposed architecture decouples market information processing from execution allocation decisions: a common feature extraction pipeline using Temporal Kolmogorov-Arnold Networks (TKANs), Variable Selection Networks (VSNs), and multi-head attention mechanisms processes market data to create informational context, while independent allocation networks handle the specific execution logic for different scenarios (fixed quantity vs. fixed notional, buy vs. sell orders). This architectural separation enables a unified model to handle diverse execution objectives while leveraging shared market understanding across scenarios. Through comprehensive empirical evaluation on intraday cryptocurrency markets and multi-day equity trading using DOW Jones constituents, we demonstrate that LEMs achieve superior execution performance compared to traditional benchmarks by dynamically optimizing execution paths within flexible time constraints. The unified model architecture enables deployment across different execution scenarios (buy/sell orders, varying duration boundaries, volume/notional targets) through a single framework, providing significant operational advantages over asset-specific approaches. Our results suggest that the deep learning methodologies developed for standard VWAP execution can be successfully extended to more complex execution problems, offering a scalable solution for institutional trading desks facing diverse execution requirements. Figure 1 : Large Execution Models (LEMs) § § footnotetext: These authors contributed equally. 1 Introduction The optimization of transaction execution is a key challenge in quantitative finance, where determining “when” and “how” to trade in the markets is essential to minimize costs and maximize performance. A fundamental aspect of this problem concerns the market impact generated by order execution, particularly for large orders. Market impact refers to the phenomenon whereby executing an order influences the asset’s price, generally creating an adverse movement for the initiator of the transaction and thus an additional implicit cost. Early studies have shown that this impact is a concave function of order size, and this impact has since frequently been decomposed into temporary and permanent components, or represented as impact that dissipates over time. Researchers proposed numerous models to quantify these impacts [ 23 , 15 , 1 , 6 , 16 ] , but two determining factors consistently emerge: the proportion that the order represents relative to the volumes usually traded in the market, and the asset’s intrinsic volatility. In the context of cryptocurrency markets, characterized by generally lower liquidity and higher volatility than traditional financial markets, market impact becomes a particularly critical issue. Rather than seeking instant execution, it therefore appears pertinent to extend the execution duration of the order so that its participation in the market volume becomes negligible, thus reducing its impact. In practice, the trader, or the broker in charge of execution, receives the initial order, which he or she divides into several sub-orders; the aggregate order is then referred to as a metaorder. However, this approach immediately raises an evaluation question: how can one objectively measure the execution quality of an order over a period of time? It is in this context that the VWAP has established itself as an essential reference. The VWAP represents the average price at which the asset actually traded during the period considered, calculated by weighting each transaction price by its corresponding volume. This measure has a unique and particularly relevant mathematical property: over any horizon, the sum of the deviations between individual transaction prices and the VWAP, weighted by the corresponding volumes, necessarily cancels out. This property makes it a natural benchmark for evaluating execution performance. This use of the VWAP as an evaluation metric gave rise to a specific order type: the VWAP order, which today ranks among the most commonly issued execution instructions by financial institutions. Unlike a limit order or a standard market order, the VWAP order executes over a defined time period and is distinguished by its performance objective: to minimize the gap between the average execution price and the market VWAP over that period. This formulation, although seemingly simple, conceals considerable complexity, as it implies the simultaneous consideration of two interdependent variables: price and volume. The evolution of academic literature on VWAP strategies reflects different approaches to this complexity. Early works, such as those by Konishi 2002 and McCulloch and Kazakov 2007 , explicitly integrated the relationship between volatility and volume into their allocation models. This approach recognized the well-documented empirical finance stylized fact that periods of high volatility generally coincide with high trading volumes. However, a subsequent trend developed in the literature, favoring an exclusive focus on modeling volumes at the expense of explicitly accounting for price–volume interaction. This approach, notably illustrated by the works of Bialkowski et al. 2008 , Bialkowski et al. 2012 , rests on a valid theoretical reasoning: if one could perfectly predict the future volume curve, it would be possible to execute a perfect VWAP by allocating orders proportionally to this curve. This simplification allowed the integration of more sophisticated models such as Self-Exciting Threshold AutoRegressive (SETAR) proposed by [ 26 ] , which apply different autoregressive models according to thresholds determined by the data themselves. The use of this type of modeling is particularly relevant because the observation of volume dynamics quickly reveals that a purely linear relationship would be insufficient to capture the observable regime changes. Another major advantage of this simplification is that it greatly facilitates the development of dynamic allocation models, enabling real-time adjustment of order allocation, a particularly desirable property for execution, which by its nature is a dynamic problem. This formalization becomes significantly more complex when one attempts to model simultaneously the interdependent dynamics of prices and volumes. Nevertheless, this simplification rests on two particularly problematic implicit assumptions in the context of cryptocurrency markets: the ability to predict future volumes accurately, and the statistical independence between prediction errors and price volatility. The first assumption runs up against the intrinsic unpredictability of financial markets, exacerbated in the crypto ecosystem by its relative youth and sensitivity to exogenous events. The second overlooks a robust empirical fact: volumes and volatility exhibit a significant positive correlation, particularly during major market moves, precisely the moments when execution quality becomes most critical. The execution of large orders in financial markets has long been recognized as a fundamental challenge in quantitative finance, with implications extending far beyond simple transaction cost minimization. Traditional approaches to optimal execution, rooted in the seminal work of Almgren and Chriss and subsequent optimal control methodologies, have provided valuable theoretical frameworks but often struggle with the practical complexities of modern market microstructure and the diverse execution requirements encountered in institutional trading environments. The landscape of execution strategies has undergone significant transformation with the emergence of deep learning techniques in financial applications. Recent work has demonstrated that neural networks can effectively capture complex temporal dependencies and market patterns in Volume Weighted Average Price (VWAP) execution scenarios [ 10 ] , achieving substantial improvements over traditional volume-based approaches. These advances suggest that the sophisticated pattern recognition