lems 0.1.1

Large Execution Models

LEMS: Large Execution Models for VWAP/TWAP Optimization

This repository provides the full implementation of Large Execution Models (LEMS), a deep learning framework for optimal VWAP/TWAP execution across multiple assets and market conditions. LEMS leverages modern Keras3 tooling and is compatible with TensorFlow, JAX, and PyTorch backends. It integrates time-aware architectures, multi-objective allocation, and efficient fused layers to offer scalable and expressive modeling for high-performance trading execution.

---

📦 Installation

Install via pip:

pip install .

Or using Poetry:

poetry install

Requirements:

• Python 3.9–3.12
• keras>=3.0.0
• tkan

Optional (for training and testing):

• jax, torch, or tensorflow
• numpy, pandas, matplotlib (for data and visualization)

---

🚀 Features

• Backend-agnostic with full support for JAX, TensorFlow, and PyTorch.
• Flexible target modeling: simultaneous optimization of VWAP/TWAP for both volume and notional allocation.
• FusedMLP layer for highly parallelized MLP computation across model variants (volume/notional, VWAP/TWAP, buy/sell).
• Custom constraints and initializers for learnable allocation curves with sum-to-one and positivity constraints.
• Soft clipping mechanisms to enforce trading limits while maintaining differentiability.
• Multiple execution-aware loss/metrics to track and optimize performance across multiple trading strategies.
• Full training pipeline with evaluation-ready data preparation tools.

---

🧠 Model Components

The main model, LargeExecutionModel, consists of:

• EmbeddingLayer: learns per-feature representations.
• VariableSelectionNetwork: soft attention over input variables.
• TKAN RNN layers: efficient temporal modeling with attention kernels.
• MultiHeadAttention: causal self-attention for cross-step dependency modeling.
• FusedMLP: efficiently computes multiple allocation predictions in parallel (buy/sell × VWAP/TWAP × volume/notional).
• Custom constraints: such as PositiveSumToOneConstraint and EqualInitializer for normalized allocation curves.

---

🧪 Example Usage

from lems.models import LargeExecutionModel
from lems.loss import lem_loss
from lems.metrics import (
    buy_vwap_avg_imp, buy_twap_avg_imp,
    sell_vwap_avg_imp, sell_twap_avg_imp,
    buy_vwap_avg_risk, buy_twap_avg_risk,
    sell_vwap_avg_risk, sell_twap_avg_risk,
    unsigned_vwap_avg_risk, unsigned_twap_avg_risk
)

model = LargeExecutionModel(
    lookback=120,
    n_ahead=12,
    hidden_size=200,
    num_embedding=20,
    num_heads=8,
    fused_mlp_hidden_dim=100
)

model.compile(
    optimizer="adam",
    loss=lem_loss,
    metrics=[
        buy_vwap_avg_imp, buy_twap_avg_imp,
        sell_vwap_avg_imp, sell_twap_avg_imp,
        buy_vwap_avg_risk, buy_twap_avg_risk,
        sell_vwap_avg_risk, sell_twap_avg_risk,
        unsigned_vwap_avg_risk, unsigned_twap_avg_risk
    ]
)

history = model.fit(X_train, y_train, validation_data=(X_val, y_val), epochs=30, batch_size=64)

---

📊 Data Preparation

A full pipeline is included via data_formatter.py. Example:

from lems.data_formater import full_generate, add_config_to_X, add_config_to_y

X_train, X_test, y_train, y_test, dates = full_generate(
    volumes_df, notionals_df,
    target_asset='AAPL',
    lookback=120,
    n_ahead=12,
    split_date='2022-01-01'
)

X_train = add_config_to_X(X_train, minimum_traded_per_period=0.0, maximum_traded_per_period=1.0)
y_train = add_config_to_y(y_train)

---

✅ Testing & Backend Compatibility

Tests are defined using pytest and validate:

• Consistency before and after saving the model
• Compatibility across JAX, TensorFlow, and Torch backends
• GPU compatibility on NVIDIA (currently not compatible with AMD GPUs due to fused ops)

---

📚 Example Results

Two notebooks in example_and_results/ show results on:

• Dow Jones equities (e.g. AAPL, MSFT)
• Cryptocurrencies (e.g. BTC, ETH)

They demonstrate significant performance improvements over baseline TWAP/VWAP curves using LEMS.

---

📖 Key Files

• models.py: Full model definition
• metrics.py: Custom metrics (VWAP/TWAP performance, risk-adjusted returns)
• loss.py: LEM loss function for efficient and risk-sensitive execution
• data_formatter.py: Preprocessing pipeline from raw time series to training arrays
• example_and_results/: Notebooks with full training code and benchmarks

---

📈 Metrics

Each allocation path (Buy/Sell × VWAP/TWAP) is evaluated using custom Keras metrics:

• *_avg_imp: average improvement in execution price vs market benchmark
• *_avg_risk: risk-adjusted reward relative to baseline
• unsigned_*_avg_risk: neutral performance regardless of trade direction

These are designed for both interpretability and reward shaping.

---

📝 License

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

---