Pre-built Models

See also

See model_layer for more details.

As a software package, part of the innovation of FinOL is the pre-implemented models in its model layer. These pre-built models serve as a significant advantage for researchers in the data-driven OLPS domain, offering them a solid foundation to build upon. By leveraging these models, researchers can streamline the initial phases of their projects, circumventing the need to build complex models from scratch and thereby conserving valuable time and resources.

Time Series Representation Models

The time series representation models within FinOL are implemented to address the unique challenges of sequential data in data-driven OLPS. These models are specifically designed to handle sequential inputs, ensuring that the temporal dependencies inherent in financial time series data are accurately captured and analyzed. By leveraging these models, researchers can gain deeper insights into market trends, price movements, and other critical factors that influence data-driven OLPS task.

AlphaPortfolio

Overall Framework of AlphaPortfolio

Description:

The AlphaPortfolio model is a Transformer-based model for asset scoring and portfolio selection.

Basic Idea:

Leverages attention mechanisms to score assets based on their historical performance.

Construction:

Source code for AlphaPortfolio

Pros:

High accuracy in scoring assets; interpretable results through economic distillation.

Cons:

Computationally intensive; requires careful tuning of hyperparameters.

Details:

The AlphaPortfolio model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the AlphaPortfolio model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

For more details, please refer to the paper AlphaPortfolio: Direct Construction through Reinforcement Learning and Interpretable AI.

Table C.1: Hyperparameters of TE-CAAN-Based AP.

Hyper-parameter	Choice	Hyper-parameter	Choice
Embedding dimension	256	Optimizer	SGD
Feed-forward network	1021	Learning rate	0.0001
Number of multi-head	4	Dropout ratio	0.2
Number of TE layer	1	Training epochs	30

AlphaStock

Key Component of AlphaStock

Description:

The AlphaStock model is a LSTM-based model for asset scoring and portfolio selection.

Basic Idea:

Utilizes LSTM to capture both sequential and global dependencies in asset data.

Construction:

Source code for AlphaStock

Pros:

Effective for time series data; can learn complex patterns over time.

Cons:

May suffer from overfitting.

Details:

The AlphaStock model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the AlphaStock model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

For more details, please refer to the paper AlphaStock: A Buying-Winners-and-Selling-Losers Investment Strategy using Interpretable Deep Reinforcement Attention Networks.

DNN

Description:

A Deep Neural Network model for asset scoring and portfolio selection.

Basic Idea:

Implements multiple fully connected layers to learn representations from input features.

Construction:

Source code for DNN

Pros:

Flexible architecture.

Cons:

Risk of overfitting without regularization.

Details:

The DNN model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

GRU

Description:

A Gated Recurrent Unit model for asset scoring and portfolio selection.

Basic Idea:

Similar to LSTM but with a simpler architecture, focusing on capturing sequential dependencies.

Construction:

Source code for GRU

Pros:

Less computationally intensive than LSTM; faster training times.

Cons:

May not capture long-term dependencies as effectively as LSTM.

Details:

The GRU model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

LSRE-CAAN

Overall Framework of LSRE-CAAN

Description:

The LSRE-CAAN model is a Transformer-based model for asset scoring and portfolio selection.

Basic Idea:

Combines long sequence representations with attention mechanisms to address efficiency issues in AlphaPortfolio.

Construction:

Source code for LSRE_CAAN

Pros:

Efficient in processing long sequences; powerful attention mechanisms.

Cons:

Lacks theoretical guarantees.

Details:

The LSRE-CAAN model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the LSRE-CAAN model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

For more details, please refer to the paper Online portfolio management via deep reinforcement learning with high-frequency data.

Table 7: Hyper-parameters of the LSRE-CAAN framework.

Hyper-parameter	Choice	Description
Depth of net (L)	1	The number of process layers in LSRE.
Number of latents (M)	1	The number of latents.
Latent dimension (D)	32	The size of the latent space.
Number of cross-heads	1	The number of heads for cross-attention.
Number of latent-heads	1	The number of heads for latent self-attention.
Cross-attention dimension	64	The number of dimensions per cross-attention head.
Self-attention dimension	32	The number of dimensions per latent self-attention head.
Dropout ratio	None	No dropout is used following Jaegle et al. (2022).
Embedding dimension	None	No Embedding layer is used, as illustrated in Section 4.1.
Optimizer	LAMB	An optimizer specifically designed for Transformer-based models.
Learning rate	0.001	Parameter of the LAMB optimizer.
Weight decay rate	0.01	Parameter of the LAMB optimizer.
Training steps	104	Training times.
Episode length (T)	50	The length of an episode.
G	m/2	Half of the assets are identified as winners.
W	100	The look-back window size.

LSTM

Description:

Long Short-Term Memory model for asset scoring and portfolio selection.

Basic Idea:

Designed to remember information for long periods, ideal for time series data.

Construction:

Source code for LSTM

Pros:

Excellent for capturing temporal dependencies; widely used in time series field.

Cons:

More complex than traditional RNNs; can be computationally expensive.

Details:

The LSTM model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

RNN

Description:

Recurrent Neural Network model for asset scoring and portfolio selection.

Basic Idea:

Processes sequences of data by maintaining a hidden state that captures information from previous inputs.

Construction:

Source code for RNN

Pros:

Simple architecture; effective for short-term dependencies.

Cons:

Struggles with long-term dependencies; can suffer from vanishing gradients.

Details:

The RNN model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

TCN

Description:

Temporal Convolutional Network model for asset scoring and portfolio selection.

Basic Idea:

Utilizes convolutional layers to capture temporal dependencies across sequences.

Construction:

Source code for TCN

Pros:

Can handle long-range dependencies; parallelizable training.

Cons:

May require careful design of the network architecture; less interpretable than RNNs.

Details:

The TCN model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

Transformer

Description:

Transformer model for asset scoring and portfolio selection.

Basic Idea:

Employs self-attention mechanisms to process sequences in parallel.

Construction:

Source code for Transformer

Pros:

Highly effective for a wide range of tasks; captures complex relationships.

Cons:

Requires significant computational resources.

Details:

The Transformer model takes an input tensor x of shape (batch_size, num_assets, num_features_augmented), where num_features_augmented represents the number of features (including any preprocessed or augmented features) for each asset. The final output of the model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

Image Representation Models

In addition to time series data, FinOL also provides image representation models. These models are designed to extract meaningful representations from financial images, such as charts, graphs, and other visual data. By doing so, it enables researchers to harness the wealth of information contained in visual formats, which can be crucial for tasks like pattern recognition, trend analysis, etc. The inclusion of this model in FinOL further solidifies its position as a versatile tool for data-driven OLPS research.

CNN

Description:

Convolutional Neural Network model for asset scoring and portfolio selection.

Basic Idea:

Applies convolutional layers to extract features from images of assets.

Construction:

Source code for CNN

Pros:

Effective for image classification and feature extraction.

Cons:

Requires significant computational resources.

Details:

The CNN model takes an input tensor x of shape (batch_size, num_assets, height, width), where height and width are the dimensions of the image for each asset. The final output of the model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

CNN-JF

Overall Framework of CNN-JF

Description:

The CNN-JF model is a CNN-based model for asset scoring and portfolio selection. It leverages CNN to analyze historical stock price data represented as images.

Basic Idea:

Leverages CNN architecture to capture patterns in financial time series represented visually.

Construction:

Source code for CNN_JF

Pros:

Combines advantages of visual data analysis with financial modeling.

Cons:

Requires significant computational resources.

Details:

The CNN-JF model takes an input tensor x of shape (batch_size, num_assets, height, width), where height and width are the dimensions of the image for each asset. The final output of the model is a tensor of shape (batch_size, num_assets), where each element represents the predicted score for the corresponding asset.

For more details, please refer to the paper (Re-)Imag(in)ing Price Trends.

Hyper-parameters of (Re-)Imag(in)ing Price Trends.

Hyper-parameter	Choice	Hyper-parameter	Choice
Kernel Size Height	5	Kernel Size Width	3
Stride Height	3	Stride Width	1
Dilation Height	2	Dilation Width	1
Padding Height	12	Padding Width	1
Dropout Rate	0.5