Companion repository for:

de Kok, T. (2025). ChatGPT for textual analysis? How to use generative LLMs in accounting research. Management Science.

Paper link: SSRN - Management Science

Table of contents

• 1. Online Appendix
  • SSRN
• 2. Code examples
  • OpenAI API - zero-shot with ChatGPT/GPT-4
  • OpenAI API - fine-tuning
  • OpenAI API - ChatGPT/GPT-4 speed optimization
  • Replicate - zero-shot with LlaMa2 70b
  • Vast.AI + Huggingface - zero-shot
  • Runpod.ai + LlaMA-Factory + Deepspeed - fine-tuning
  • Runpod.ai + VLLM -> Fast Inference
• 3. Paper dataset
  • Available in the datasets folder, see description below.
• 4. Full paper code
  • Supplemental Material

Dataset description

Notes

• Located in the datasets folder, available as .dta or .parquet.
• Every row represents an earnings call, keyed by symbol, call_fyear, and call_fquarter.
• Identifiers beyond symbol are provided for convenience, but only on an as-is basis (they are not used in the paper).

Screenshot of dataset

The columns map to Appendix A.2 (Non-answer dimensions) in the following way

Dimension	Outcome	Column Name in Dataset
Justification	Cannot give	num_na_just_cant_give
Justification	Do not know	num_na_just_dont_know
Non-answer type	Complete refusal	num_na_type_compl_refuse
Non-answer type	Qualitative	num_na_type_qualitative
Non-answer type	Range or percentage	num_na_type_range_or_perc
Question type	Breakdown	num_na_qtype_breakdown
Question type	Forward-looking	num_na_qtype_forwardlooking
Question type	Related party	num_na_qtype_related_party
Question type	R&D or regulation	num_na_qtype_rd_or_reg
Question type	Other	num_na_qtype_other
Sentiment spin	Optimistic	num_na_senspin_optimistic
Sentiment spin	Neutral or pessimistic	num_na_senspin_neut_or_pessim