LEMs: A Primer On Large Execution Models

Battle Count: 50

This paper is highly relevant to quantitative trading, specifically in the domain of algorithmic execution and order management. It directly addresses the core problem of how to optimally execute large orders in financial markets while minimizing costs relative to benchmarks (VWAP/TWAP). The unified framework handles multiple execution scenarios (buy/sell, volume/notional, fixed/flexible time horizons) in a single model, which is directly applicable to institutional trading desks. The empirical results on both cryptocurrency and equity markets demonstrate practical improvements over traditional TWAP benchmarks. The connection to share buyback contracts and optimal stopping problems further broadens its applicability. The paper builds on and extends prior work by the same authors on neural VWAP execution strategies.

The architecture is highly complex, combining multiple novel components: Temporal Kolmogorov-Arnold Networks (TKANs) with LSTM-inspired memory, Variable Selection Networks with Gated Residual Networks, multi-head attention with causal masking, and a FusedMLP with parallel paths for different execution modalities. The loss function involves sophisticated masking, soft execution decisions, differentiable clipping, and multi-scenario PnL computation. Training requires careful handling of constraint satisfaction, sequential allocation conservation, and temporal masking. However, the authors provide a pip-installable package ('pip install lems') and GitHub repository, which significantly reduces the barrier to adoption. The model processes multi-frequency, multi-asset data simultaneously, adding to the engineering complexity.

Code is available via 'pip install lems' and a GitHub repository is referenced. Data is accessible for reproducing experiments via the GitHub link. The paper provides detailed architecture descriptions, loss function formulations, training configurations (Adam optimizer, learning rate 0.00001, batch size 256, early stopping patience 2, LR reduction patience 1), and feature engineering methodology. However, exact hyperparameter values for all network dimensions (hidden units H, number of heads mH, attention dimension d_attn, number of fused paths M, number of layers L) are not fully specified in the extract. The paper covers two distinct experimental setups (crypto and equity) with clear data splits.