llm-etf-prediction

ℹ️ About llm-etf-prediction

Traditional sequence models, like RNNs and LSTMs, struggle to incorporate related data beyond time series for accurate forecasting. Time-LLM addresses this limitation by combining natural language-embedded domain knowledge with time series data, utilizing a pre-trained LLM for downstream tasks.

In this project, Time-LLM is applied to predict the closing prices of Exchange Traded Funds (ETFs). By embedding sector-specific information about the ETFs, such as the industries they track or their inherent market characteristics, the model aims to improve predictive accuracy and make more reliable ETF close price forecasting.

✅ How to run

1) Prerequisites

To run this model, the following prerequisites are required:

• CUDA Version: 11.8
• Nvidia Driver Version: 470.103.01

conda create -n etf python=3.10.0
conda activate etf
conda install nvidia/label/cuda-11.8.0::cuda-toolkit
pip install torch==2.5.0+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
pip install numpy==2.1.3
pip install neuralforecast
pip install transformers
pip install matplotlib
pip install ipywidgets
pip install yfinance

2) Clone

Clone this repository to your local computer using:

https://github.com/JeonHaeseung/llm-etf-prediction.git

3) Run

In the time-llm.py file, you can select your desired ETF to create the model.

python3 time-llm.py

To check the result graph, you can run result.py.

python3 result.py

📊 Dataset

Stock price datasets can be retrieved via API calls using Python's yfinance library. Natural language description data for the LLM is hard-coded in time-llm.py. If you want to fetch descriptions for a new stock, search for the stock ticker on https://money.usnews.com and use the description provided.

📈 Results

Metric	IYR	KBE	SMH	IBB	SPY	GDX
MAE	0.99	0.66	25.23	4.58	6.96	2.93
MSE	1.61	0.65	649.68	21.77	57.92	11.15
MAPE	0.01	0.01	0.09	0.03	0.01	0.07

Time-LLM performed well with IYR and KBE, which follow relatively stable sectors. These ETFs exhibited the lowest error rates, likely due to their low variance and regular cycles, indicating that the model effectively utilized the domain knowledge provided. Conversely, SMH and GDX showed higher error rates, with SMH displaying the highest among all predictions. The results appear to stem from these sentiment-based judgments since sentiment analysis of the domain knowledge texts revealed instances where the tone was interpreted as positive or negative rather than completely neutral.