k-datasoft/Multimodal-test-dataset

KRX Investment Warning Prediction Dataset (OHLCV + Korean News)

Dataset Summary

This dataset is a test dataset for predicting Investment Warning (투자주의종목) designations in the Korean stock market (KRX). It contains raw daily OHLCV price data and Korean news text (title + body), designed for multimodal anomaly detection / binary classification.

Important: No technical indicators are included, and no normalization/scaling is applied.

• —Date range: 2025-07-01 ~ 2025-09-30
• —Prediction horizon: whether a stock will be designated as an investment warning within the next 1 trading day

Task

Binary classification:

• —Label 0: Normal trading (no investment warning designation within the next 1 trading day)
• —Label 1: Investment warning designation (within the next 1 trading day)

Label Alignment

For each (ticker, date=t), set label=1 if the stock is designated as an investment warning on t+1 (the next trading day).

Data Sources

Dataset Format

This dataset is structured to be used directly with Hugging Face datasets, and consists of three columns:

• —`labels`: Binary label (0 or 1)
• —`time_series`: Price time-series information (OHLCV)
• —`texts`: Korean news text mapped to the corresponding stock (title + body)

The exact internal structure of time_series and texts (e.g., list/dict formats, sequence length, date ordering) follows the dataset schema. In general, time_series is provided as a fixed-length historical window, and texts contains news from the same date (or window period), either concatenated or stored as a list.

Example (Conceptual)

• —labels: 0 or 1
• —time_series: [[open, high, low, close, volume], ...]
• —texts: ["article1 ...", "article2 ..."]

Feature Details

Price (OHLCV) — time_series

• —OHLCV is provided as raw daily bars.
• —No technical indicators (e.g., RSI, MACD) are included.
• —No normalization/scaling is applied.
• —Currency unit: KRW
• —Volume: number of shares (not value)

News — texts

• —News is mapped to tickers via an exact ticker-code mapping.
• —Deduplication has been applied.
• —Each news item includes title + body (stored as a single string or list depending on schema).

Recommended Metrics

Because investment warning events are likely to be rare (class imbalance), the following metrics are recommended:

• —ROC-AUC, PR-AUC
• —F1 (positive class), precision/recall
• —Precision/recall at Top-k (useful for practical detection scenarios)
• —(Optional) probability calibration