dl14666 / voltage

VolTAGE: Volatility forecasting via Text-Audio fusion with Graph convolution networks for Earnings calls

VolTAGE is a mutli-task volatility prediction model that uses cross-modal attention fusion across the text and speech modalities while learning the relational encodings between inter-dependent stocks. It contains the following four modules:

• Speech and text feature extraction
• Cross-modal attention fusion for text and audio features
• Semi-supervised graph convolutional neural network
• Multi-task volatility prediction using "conditioned" LSTMs

Requirements

• For being able to use the code from the files run the command pip3 install requirements.txt

Speech and text feature extraction

• The data in the folder ./audio-feature_extraction/OGdataset is incomplete and is only for illustration. The entire data is available here
• For the textual feature extraction we use the code here
• For the audio feature extraction the code is available in the following script which can be run using python3 ./audio-feature_extraction/audioExtractionPraat.py

Cross-modal attention fusion of text and audio

• The audio and text features extracted from the above step can be made using the key:value format <text_file_name_for_call>:feature_vector
• To get the multi-modal attentive embeddings trained on the volatility prediction task run the script using python3 ./cross-modal-attn/stockspeech-3-at-attn.py

Semi-supervised graph convolutional neural network

• To extract the WikiData company based relations save all your required entity IDs (or Q values) as in ./semi-supervised-gcn/relation_extraction/earnings_wiki.csv
• Then to generate the heterogenous graph and save the adjacency matrix run the script python3 ./semi-supervised-gcn/relation_extraction/graph_script.py
• Save the graph at ./semi-supervised-gcn/gcn/data and run the python script using the command python3 ./semi-supervised-gcn/gcn/StockGCN.py to get the saved relational embeddings trained on the attentive multi-modal feature representations. These embeddings will be next used to multi-task volatility prediction using conditioned LSTMs

Multi-task volatility prediction using "conditioned" LSTMs

• The saved embeddings from the previous step must be saved at ./multi-task-lstm-conditioning. These are used as the hidden state for the conditioned LSTM trained on the past volatility values which can be generated from the data here
• Then to get the final model performance results run the python script using the command python3 ./multi-task-lstm-conditioning/stockspeech-final--lstm-cond.py

Cite this Paper

Consider citing our work if you use our codebase.

@inproceedings{sawhney2020voltage,
  title={VolTAGE: volatility forecasting via text-audio fusion with graph convolution networks for earnings calls},
  author={Sawhney, Ramit and Khanna, Piyush and Aggarwal, Arshiya and Jain, Taru and Mathur, Puneet and Shah, Rajiv},
  booktitle={Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  pages={8001--8013},
  year={2020}
}