Is the positive/negative tone in S-4 and DEFM14A associated with litigation risk?
- Loughran-McDonald as positive/negative word dictionary
- Analyze their presence, order, and dispersion
Are certain topics in S-4 and DEFM14A associated with litigation risk?
- Latent Dirichlet Allocation to find the topics
- Analyze their presence, order, and dispersion
Embed the documents and analyze them
- Does the content across the years tends to change or it remains the same (use cosine similarity)
- How boilerplate are the documents? Since you cannot directly tell if for a specific company the embeddings are becoming more similar (mergers do not happen every year), you combine how similar are the embeddings relative to the others documents and get a boilerplate score from that. Try multiple things:
  - Split the documents into specific sections and test the similarity on the whole document and on the specific sections
  - Find the average similarity across the whole corpus of the documents
  - Find the average similarity only across the k-nearest neighbors on the embeddings (this prevents averaging scores of for example medical - tech - insurance field which might be really different)
- Visualize in a 2D dimensional space the embeddings to see if there are cluster with documents with
  - litigations/non-litigations
  - public acquiror/private acquiror
  - specific topics
- Train different model to do feature extraction to see if they can predict the litigations
  - Train Random Forest
  - Train XGBoost
  - Train FFN
- Combine the feature extraction from the embeddings with the quantitative data of the merger to see if the predictive quality of the model increases