I previously made a stock simulation and made it accessible on the web in a way that I can view.
Since it's initial, only the top 100 KOSPI stocks are loaded.
For use with Kiwoom's real-time account API, as long as the call frequency is less than 5 times per second and less than 1000 times per hour, it's sufficient for simulation (calling 10 stocks simultaneously in parallel like xxxxx, xxxxx, xxxxxx seems to be the most stable).
Currently, I am testing two versions of the qwen3 14b model.
A : I just threw everything in and put in 10,000 pieces of data.
B : As much as possible, I refined textbooks mainly based on Kiwoom and the Korea Exchange's PDFs and Dart's official materials. I matched about 1,000 textbooks tailored to the Korean stock market. About 500 are textbook content, 300 are terms or Korean-specific content, and 200 are Q&As. In the future, it seems like we can increase the number of Q&As. For now, about 1,000 have been trained.
Each model has 10 bots, each with an initial capital of 100 million won, and they proceed based only on the data sent from Kiwoom without any prior information. It's pure performance.
I started a separate simulation last week with a capital of 100 million won and there is profit.
Next week, I plan to proceed by giving prior information. I will provide weekly reports for about a week + the price trends of the top 100 companies. I also plan to give the models an independent memo function and a system that can use RAG.
From now on, fight mode?????
I don't know which is better, the model with 1000 textbooks or the one with 10,000 pieces of data.
