dynamic_vwap 0.1.0

Deep-Learning for Optimal VWAP Execution

Recurrent Neural Networks for Dynamic VWAP Execution: An Adaptive Trading Strategy with Temporal Kolmogorov-Arnold Networks

This repository presents Aplo's latest research on dynamic VWAP execution and contains the code discussed in the paper Recurrent Neural Networks for Dynamic VWAP Execution: An Adaptive Trading Strategy with Temporal Kolmogorov-Arnold Networks.

Note: This version includes only the TKAN-based dynamic VWAP model for simplicity in serialization. An LSTM-based version can be implemented by modifying the code accordingly.

Model Overview

The dynamic VWAP model is implemented as a Keras model compatible with any backend (TensorFlow, JAX, or PyTorch). We recommend using JAX for optimal performance.

Installation

1. Clone the repository:

   git clone <repository-url>
   cd <repository-directory>
   

2. Install the package:

   pip install .
   

   Alternatively, using Poetry:

   poetry install
   

Usage

Below is a minimal example demonstrating how to use the dynamic VWAP model:

from dynamic_vwap import DynamicVWAP, quadratic_vwap_loss, absolute_vwap_loss, volume_curve_loss
from dynamic_vwap.data_formater import full_generate
import pandas as pd

# Parameters
lookback = 120   # Number of past time steps used as input
n_ahead = 12     # Number of future time steps to predict
target_asset = 'AAPL'
BATCH_SIZE = 128
N_MAX_EPOCHS = 1000

# Load your data (here using Parquet files)
volumes = pd.read_parquet('path_to_your_volume_data.parquet')
notionals = pd.read_parquet('path_to_your_notionals_data.parquet')

# Generate training and testing datasets
X_train, X_test, y_train, y_test = full_generate(
    volumes, 
    notionals, 
    target_asset,
    lookback=lookback, 
    n_ahead=n_ahead, 
    test_split=0.2, 
    autoscale_target=True
)

# Initialize the dynamic VWAP model
model = DynamicVWAP(
    lookback=lookback, 
    n_ahead=n_ahead, 
    hidden_size=100, 
    hidden_rnn_layer=2
)

# Compile the model with a VWAP-specific loss function
model.compile(optimizer='adam', loss=quadratic_vwap_loss)

# Train the model
history = model.fit(
    X_train, y_train, 
    batch_size=BATCH_SIZE, 
    epochs=N_MAX_EPOCHS, 
    validation_split=0.2, 
    shuffle=True, 
    verbose=False
)

# Make predictions
predictions = model.predict(X_test, verbose=False)

Model Parameters

• lookback: Number of past time steps used as input.
• n_ahead: Number of future time steps for which the dynamic volume curve is predicted.
• hidden_size: Number of units in the hidden layers of the internal RNN.
• hidden_rnn_layer: Number of TKAN layers in the internal RNN.

Note: The input data matrix must have a sequence length of lookback + n_ahead - 1 time steps. This format ensures that the model receives the necessary ahead inputs during training. For real-time applications, appropriate padding may be needed.

Loss Functions

The package provides the following loss functions to effectively minimize the deviation between the achieved VWAP and the market VWAP:

• quadratic_vwap_loss
• absolute_vwap_loss
• volume_curve_loss

Data Formatting

The model expects inputs in matrix format rather than a dictionary. The expected shapes are:

• Features Input: A NumPy array of shape (num_samples, lookback + n_ahead - 1, num_features).
• Targets: A NumPy array of shape (num_samples, n_ahead, 2), where the first element (along the last dimension) corresponds to volume allocations and the second corresponds to prices.

The provided helper function full_generate in data_formater.py facilitates the creation of training and testing datasets. For example:

import pandas as pd
from dynamic_vwap.data_formater import full_generate

volumes = pd.read_parquet('path_to_your_volume_data.parquet')
notionals = pd.read_parquet('path_to_your_notionals_data.parquet')

X_train, X_test, y_train, y_test = full_generate(
    volumes, 
    notionals, 
    target_asset='AAPL', 
    lookback=120, 
    n_ahead=12, 
    test_split=0.2, 
    autoscale_target=True
)

Example and Results

For detailed examples, including reproduction of experimental results and graphs from the paper, please refer to the example_and_results folder.

License

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