Using Kiwoom API to Fetch Stock Volume Data for ML Forecasting

Background

To understand which stocks are surging or falling, I started working with real-world stock trading data. My primary data source: Kiwoom’s REST API. The goal was to:

• Fetch real-time trading volume
• Compare it with previous day’s volume
• Include foreign investor trading data

This information would serve as the foundation for identifying rapid movements in stocks and eventually training a model to forecast those changes.

---

Why Volume Data?

Volume is a leading indicator in technical analysis. If a stock is suddenly seeing abnormal volume spikes, it usually signals an important movement. I focused on:

• Real-time volume (현재 거래량)
• Previous day volume (전일 거래량)
• Foreign trading volume (외국인 순매수/매도)

---

Technical Challenges

Using the Kiwoom API was not straightforward. Key difficulties:

• Each data type (real-time, foreign volume, etc.) uses a different request code
• Response formats are not standardized
• Some values are only accessible through event-based callbacks (not ideal for batch fetching)
• Correlating data across types was hard due to inconsistent indexing

This meant that fetching all related data and merging it correctly required custom logic and frequent debugging.

---

Example: Fetching Real-Time Trading Volume (Python + PyKiwoom)

from pykiwoom.kiwoom import Kiwoom  
import time  

kiwoom = Kiwoom()  
kiwoom.CommConnect(block=True)  

# Example: Fetch top trading volume stocks  
sector_code = "001"  # 코스피 업종 코드 예시  
tickers = kiwoom.GetCodeListByMarket("0")  

for code in tickers[:10]:  
    name = kiwoom.GetMasterCodeName(code)  
    data = kiwoom.block_request("opt10001",  
                                종목코드=code,  
                                output="주식기본정보",  
                                next=0)  
    volume = data.get('거래량')  
    print(f"{name} ({code}): 거래량 {volume}")  
    time.sleep(1)  # Too many requests can trigger rate limiting  

---

Next Steps: Toward ML Forecasting

Once I clean and align these volume datasets, my plan is to:

• Merge volume, price, and investor types into a structured time series
• Use this data to label rising/falling patterns
• Apply binary classification (e.g. will rise: yes/no) using ML models like LightGBM or XGBoost

---

What I Learned

• Financial APIs often sacrifice consistency for flexibility
• Preprocessing and integrating data matters more than model selection at this stage
• Reliable volume data is not just useful for trading — it’s a signal-rich input for ML models

---

Challenges Still Remaining

• Handling API throttling and asynchronous requests
• Designing a clean data schema for training
• Creating a feature pipeline that updates in real-time

I’m continuing to iterate on the pipeline, and I expect the real edge will come from how I preprocess and label the data.

---