# Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models Last updated: {sub-ref}`today` ## Table of Contents ```{toctree} :maxdepth: 1 :caption: Notebooks 📖 ``` ```{toctree} :maxdepth: 1 :caption: Pipeline Charts 📈 charts.md ``` ```{postlist} :format: "{title}" ``` ```{toctree} :maxdepth: 1 :caption: Pipeline Dataframes 📊 dataframes/P5/crsp_daily.md ``` ```{toctree} :maxdepth: 1 :caption: Appendix 💡 myst_markdown_demos.md apidocs/index ``` ## Pipeline Specs | Pipeline Name | Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models | |---------------------------------|--------------------------------------------------------| | Pipeline ID | [P5](./index.md) | | Lead Pipeline Developer | samoszhu | | Contributors | samoszhu | | Git Repo URL | | | Pipeline Web Page | Pipeline Web Page | | Date of Last Code Update | 2026-02-08 01:38:20 | | OS Compatibility | | | Linked Dataframes | [P5:crsp_daily](./dataframes/P5/crsp_daily.md)
| ## About this project This paper demonstrates the capability of large language models (LLMs) like ChatGPT to predict stock market reactions from news headlines. ## Quick Start The quickest way to run code in this repo is to use the following steps. You must have TexLive (or another LaTeX distribution) installed on your computer and available in your path. You can do this by downloading and installing it from here ([windows](https://tug.org/texlive/windows.html#install) and [mac](https://tug.org/mactex/mactex-download.html) installers). First, you must have the `conda` package manager installed (e.g., via Anaconda). However, I recommend using `mamba`, via [miniforge](https://github.com/conda-forge/miniforge) as it is faster and more lightweight than `conda`. Create and activate the conda environment: ```bash conda env create -f environment.yml conda activate p5 ``` Finally, run the project tasks: ```bash doit ``` And that's it! ### Other commands #### Unit Tests and Doc Tests You can run the unit test, including doctests, with the following command: ``` pytest --doctest-modules ``` You can build the documentation with: ``` rm ./src/.pytest_cache/README.md jupyter-book build -W ./ ``` Use `del` instead of rm on Windows #### Setting Environment Variables You can [export your environment variables](https://stackoverflow.com/questions/43267413/how-to-set-environment-variables-from-env-file) from your `.env` files like so, if you wish. This can be done easily in a Linux or Mac terminal with the following command: ```bash set -a # automatically export all variables source .env set +a ``` On Windows (PowerShell): ```powershell Get-Content .env | ForEach-Object { if ($_ -match '^([^=]+)=(.*)$') { [Environment]::SetEnvironmentVariable($matches[1], $matches[2], 'Process') } } ``` ### Formatting This project uses [Ruff](https://docs.astral.sh/ruff/) for linting and formatting Python code. ```bash # Auto-fix linting issues (e.g., unused imports, undefined names) ruff check . --fix # Format code (consistent style, spacing, line length) ruff format . # Sort imports, then fix linting issues, then format ruff format . && ruff check --select I --fix . && ruff check --fix . ``` - `ruff check --fix` applies safe auto-fixes for linting violations - `ruff format` formats code similar to Black - `--select I` targets only import sorting rules (isort-compatible) ### General Directory Structure - The `assets` folder is used for things like hand-drawn figures or other pictures that were not generated from code. These things cannot be easily recreated if they are deleted. - The `_output` folder, on the other hand, contains dataframes and figures that are generated from code. The entire folder should be able to be deleted, because the code can be run again, which would again generate all of the contents. - The `data_manual` is for data that cannot be easily recreated. This data should be version controlled. Anything in the `_data` folder or in the `_output` folder should be able to be recreated by running the code and can safely be deleted. - I'm using the `doit` Python module as a task runner. It works like `make` and the associated `Makefile`s. To rerun the code, install `doit` (https://pydoit.org/) and execute the command `doit` from the `src` directory. Note that doit is very flexible and can be used to run code commands from the command prompt, thus making it suitable for projects that use scripts written in multiple different programming languages. - I'm using the `.env` file as a container for absolute paths that are private to each collaborator in the project. You can also use it for private credentials, if needed. It should not be tracked in Git. ### Data and Output Storage I'll often use a separate folder for storing data. Any data in the data folder can be deleted and recreated by rerunning the PyDoit command (the pulls are in the dodo.py file). Any data that cannot be automatically recreated should be stored in the "data_manual" folder. Because of the risk of manually-created data getting changed or lost, I prefer to keep it under version control if I can. Thus, data in the "_data" folder is excluded from Git (see the .gitignore file), while the "data_manual" folder is tracked by Git. Output is stored in the "_output" directory. This includes dataframes, charts, and rendered notebooks. When the output is small enough, I'll keep this under version control. I like this because I can keep track of how dataframes change as my analysis progresses, for example. Of course, the _data directory and _output directory can be kept elsewhere on the machine. To make this easy, I always include the ability to customize these locations by defining the path to these directories in environment variables, which I intend to be defined in the `.env` file, though they can also simply be defined on the command line or elsewhere. The `settings.py` is responsible for loading these environment variables and doing some preprocessing on them. The `settings.py` file is the entry point for all other scripts to these definitions. That is, all code that references these variables and others are loaded by importing `config`. ### Naming Conventions - **`pull_` vs `load_`**: Files or functions that pull data from an external data source are prepended with "pull_", as in "pull_fred.py". Functions that load data that has been cached in the "_data" folder are prepended with "load_". For example, inside of the `pull_CRSP_Compustat.py` file there is both a `pull_compustat` function and a `load_compustat` function. The first pulls from the web, whereas the other loads cached data from the "_data" directory. ### Dependencies and Virtual Environments #### Working with `conda` environments This project uses conda for environment management. The dependencies are stored in `environment.yml`. To create/update the environment: ```bash conda env create -f environment.yml # or to update an existing environment: conda env update -f environment.yml ``` To activate the environment: ```bash conda activate p5 ``` To export the current environment: ```bash conda env export > environment.yml ``` **Tip:** Consider using `mamba` instead of `conda` for faster package resolution. Install via [miniforge](https://github.com/conda-forge/miniforge).