StepanTita / llm-trading

Financial U-Net FFwD

Modelling

For this project, we will use the following models:

• DistilBERT to extract the features from the set of buy/sell sides of the order book

For the final layer, we will use the following models:

• A simple feedforward neural network to predict the price movement based on the previous prices + features from the previous step
• Convolutional Neural Network to predict the price movement based on the previous prices + features from the previous step
• Recurrent Neural Network to predict the price movement based on the previous prices + features from the previous step

Prediction target:

• We will predict the next action (buy/sell)
• The price of this action
• The amount traded

Architecture of the Neural Network

Configuration

epochs = 1
batch_size = 512
max_seq_len = 512
learning_rate = 1e-5
max_grad_norm = 1000
hidden_dim = 3

Results Table:

Model	Val loss	Val acc	Val f1	Val precision	Val recall	Val Price RMSE	Val Price MAPE	Val Amount RMSE	Val Amount MAPE
FFwD	87461085.5548	0.8120	0.8115	0.8119	0.8113	16853.8339	2619650877685760.0	8509.8906	5010660245110784.0
Financial U-Net FFwD	102821550.3097	0.7988	0.7988	0.7991	0.7995	16416.7109375	329857212874752.0	8655.359375	409164085985280.0

Ablation Study