Time Series Forecasting by fine-tuning a LLM model in replicate

Replicate provides a inference and fine-tuning LLM services. We try to use the TimeLLM paper base as a to fine-tune llama-2-7b for forcasting best gaining value token on ethereum network.

Installation

Use poetry install dependencies:

git clone https://github.com/tumaysem/replicate.git
cd replicate
poetry install

Usage

We used our (agentc.xyz)[https://agentc.xyz] prices clickhouse database to generate fine tune prompts and push it to replicate for fine-tune.

Create your .env file ,see example.

Prices data

Collect hourly price data fro the last constants.YEARS.

poetry run prices

Train prompts

Converts prices data into propmts file for fine tuning. Replicate uses publicly accessible files for fine-tuning.

poetry run prices

Initiate a fine-tune training on replicate

Uses the prompts to fine tune llama-2-7b-chat for forecasting

poetry run train

Forecast best value gaining token

Uses fine-tuned LLM to forecast most value gaining token based on statistic data of previous token prices.

poetry run forecast