--- license: mit tags: - finance - llm - research-replication --- # Research Track: Replicating "Can ChatGPT Forecast Stock Price Movements?" **Paper:** Alejandro Lopez-Lira and Yuehua Tang, *Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models*, Journal of Financial Economics (2026). First posted on arXiv as 2304.07619. The paper was approved by the instructor. **Core idea:** a large language model reads a firm-specific headline and answers YES (good news), NO (bad news) or UNKNOWN. That score predicts the stock's next *tradable* return, beats conventional sentiment scores, and works best for small stocks and bad news. ## Repository layout ``` . ├── README.md ├── requirements.txt ├── paper_review/ │ ├── One_Page_Paper_Review.pdf # one-page review with pseudocode │ └── Framework_Flowchart_and_Replication_Guide.pdf # flow chart + implementation notes ├── notebooks/ │ └── 01_research_replication.ipynb # real-data replication ├── data/ │ ├── raw/ # headlines.csv, prices.csv (created by the notebook) │ └── processed/ # llm_scores.csv, scored_events.csv └── results/ # summary.json, long_short.png ``` ## What the replication does 1. Collects timestamped, firm-linked headlines and daily prices for about 100 U.S. stocks, from large to small caps. Prices come from `yfinance`; headlines come from the `ashraq/financial-news` archive on Hugging Face (see Limitations). 2. Scores each headline with the paper's exact prompt using an open LLM (`Qwen/Qwen2.5-3B-Instruct`). 3. Maps each headline to its first tradable return, using the paper's 06:00 / 16:00 ET rule. 4. Estimates a firm/date fixed-effects regression with two-way clustered standard errors. 5. Runs an equal-weight long-short portfolio, before and after costs. 6. Compares against a FinBERT sentiment baseline, including a head-to-head regression. ## How to run Open `notebooks/01_research_replication.ipynb` in Google Colab with a T4 GPU runtime and run all cells. Adding `OPENAI_API_KEY` as a Colab secret switches scoring to `gpt-4o-mini`. ## Results Full results are in `results/summary.json`, in notebook Sections 9-11 (hit rates, fixed-effects regressions, long-short returns, and examples where the LLM and FinBERT disagree), and in `results/long_short.png`. ## Limitations - **Data source:** Yahoo Finance's news API was unavailable from Colab at run time, so headlines come from the `ashraq/financial-news` archive (2020). These predate the model's training cutoff, so memorization of the outcomes can't be ruled out. - **Sample size:** the LLM scored about 1,460 headlines from a roughly two-month window. That is far smaller and shorter than the paper's multi-year CRSP/RavenPack sample, so the t-statistics are indicative only. - **Model:** I used a 3B-parameter open model, not the paper's GPT models. The paper finds that smaller models predict returns poorly, so weak results here are consistent with its findings. - **Filters and costs:** the archive's firm tags replace RavenPack relevance filters. Transaction costs are a flat bps charge, with no slippage or short-borrow costs. ## AI-use disclosure I used ChatGPT (OpenAI) to help find and verify the paper and to draft an initial code scaffold. I used Claude (Anthropic) to check my submission against the assignment requirements, to tighten the paper review, and to restructure the replication notebook around real data. I selected the paper, checked the prompt and timing rule against the original, and ran and interpreted the replication myself.