This is the existing v8 version stock simulation.
Two local models for stock simulation were fine-tuned at 14b and 9b, and each model was given 10+1 bots. So, each bot has a different personality, attack balance, defense, short-term, mid-term, long-term, etc., with basic characteristics + slightly different personalities to create over 100 combinations randomly.
The bots buy, hold, and sell. One bot (+1) only buys and sells without holding.
Since it runs on one GPU as a local model, the process takes about 5 minutes per model. Within 5 minutes, time is distributed to 10+1 bots, with 70% based on the model's judgment and 30% based on each bot's personality for stock trading.
There are some bots with particularly high returns. Buying at the end of Thursday's market and selling at the opening of Friday's market results in unexpectedly good returns, especially for the +1 bots of each model.
This is the v9 version.
I customized the most popular one from GitHub for the existing v8 version, but since there were functions I didn't use, I started making it from scratch with an agent in June. The agent handles everything except orders and judgments. It's almost complete today.
I'm making a system that monitors news and stock price data 24 hours a day in real time. Since it's still early, data accumulation is not yet complete. I am currently accumulating data and doing some fine-tuning on the monitoring bot.
While making a stock simulator, I studied models and faced frustrations, but it seems like I've created a version that can run all night tonight.
