World

I'm excited to say that I'll be presenting a session at PyCon KR 2025!

Documentize all

(Before)
Local file and Notion: https://www.notion.so/Financial-time-series-forecasting-with-AI-22fa35337f588040b81cdb54c78ca3a3

(After)
Github: https://github.com/deveworld/stock-forecast
Website: https://stock.worldsw.dev
Password is '250716'

Financial time series forecasting with AI

First method
→ Classification model: XGBoost (for Baseline), Transformer
Get stock data with pykrx, news with gnews
Sentiment analysis for the news with ‘tabularisai/multilingual-sentiment-analysis’
And add technical features with ‘ta’
Using class_weights in CrossEntropyLoss to unbalance classes

XGBoost: 43.5%, Transformer: 41.7%
⇒ Class Imbalance problem?

Trials (but all failed):
- Data Augmentation (Noise or SMOTE)
- Changing Threshold of Classification
- More technical features or less technical features
- Change model architecture from Transformer to LSTM
- Use FocalLoss Instead of CrossEntropyLoss
- Add EarlyStopping method to prevent Over-Fitting
- Increase dropout value
- Chaning LR
- Gradient Clipping

Okay. I feel like I'm missing something.
Turn a classification problem into a regression problem
→ RMSE: 0.0212

Impressive…. It’s even worse. Okay, regression is much harder than classification.
Roll back into classification

Add macroeconomic scale data using Fred
- KOSPI
- KRW/USD Exchange Rate
- 10-Year Treasury Constant Maturity Rate
⇒ Still low perf…

WHY??????

Compute the FFT on the CPU to simulate the gravitational field.
This is a very computationally intensive operation, and needs to be moved to the GPU.
However, the implementation of FFT on GPU is taking longer than expected because it is much harder than calculating Newtonian N-body gravity.

Decided to change from 3D to 2D (space) + 1D (gravity field) rendering to increase computational efficiency.

No more complains from linter, formatter!!

https://github.com/deveworld/KorT/commit/5f3a71042a6c9b0bb74ff79ab71cc06be7879abb

add typing for pyright

Trying to use online-judge server

FULL FPS with 100k bodies in GPU (OpenGL)
more optimization!!!

Barnes-Hut Simulation
~10 FPS with 100k bodies in CPU
more optimization needed!!