# Detecting Reversal Points in US Equities — Data

来源：https://www.kaggle.com/competitions/detecting-reversal-points-in-us-equities/data

采集时间：2026-10-19T07:13:02Z

# Time Series Classification Competition Dataset

## Overview
This dataset contains time series data from 6 anonymized financial instruments for classification tasks.

## Dataset Statistics
- **Total Samples**: 2,760
- **Features**: 68504 (excluding metadata and target)
- **Time Period**: 2023-04-03 to 2025-01-31
- **Unique Tickers**: 6

## Target Variable
- **class_label**: The target variable for 3-class classification

Multi-class classification (3 classes): ['H', 'L', 'None']
  Class H: 82 samples (3.0%)
  Class L: 81 samples (2.9%)
  Class None: 2597 samples (94.1%)

Class Conversion Mapping:
  Original -> New:
  HH, LH -> H (High patterns)
  HL, LL -> L (Low patterns)
  Others -> None (No clear pattern)
  Missing/NaN values converted to None: 2597 samples

## Problem Definition
This is a **3-class classification problem** based on financial time series patterns:

- **Class H**: High patterns (originally HH: Higher High, LH: Lower High)
- **Class L**: Low patterns (originally HL: Higher Low, LL: Lower Low)  
- **Class None**: No clear pattern or missing data

The original 5-class labels have been consolidated to create a more balanced and interpretable classification task.

## Files Description

### train.csv
- Contains all features including the target variable 'class_label'
- Use this data to train your models
- Rows: 1932 (approximately 70% of total data)

### test.csv  
- Contains all features EXCEPT 'class_label'
- Make predictions on this data for submission
- Includes 'id' column for matching predictions
- Rows: 828 (approximately 30% of total data)

### sample_submission.csv
- Shows the expected format for your predictions
- Columns: id, class_label
- Submit your predictions in exactly this format
- **Important**: Predictions must be one of the valid class values: H, L, None

### solution.csv (Competition Setup)
- Ground truth data for scoring submissions
- Columns: id, class_label, Usage
- Usage column determines public/private test split:
  - "Public": Used for public leaderboard (visible during competition)
  - "Private": Used for final private leaderboard (determines winners)
- Approximately 30% public, 70% private split

## Column Descriptions

### Metadata Columns
- **id**: Unique identifier for test set predictions (test.csv only)
- **ticker_id**: Anonymous identifier for the financial instrument (001, 002, etc.)
  - Each ticker represents a different financial time series
  - Use this to understand which samples belong to the same instrument
  - Consider ticker-specific patterns in your modeling
- **t**: Timestamp (YYYY-MM-DD format)
  - Time series data spanning from 2023-04-03 to 2025-01-31
  - Data is chronologically ordered within each ticker

### Feature Columns (68504 total)
The dataset contains various technical and fundamental features derived from financial time series data.
Note: Sensitive columns containing price information, returns, and PnL data have been removed to ensure anonymization.



### Target Column
- **class_label**: Classification target (present in train.csv only)
  - 3-class classification: H, L, None
  - **H**: High patterns (HH: Higher High, LH: Lower High trends)
  - **L**: Low patterns (HL: Higher Low, LL: Lower Low trends)  
  - **None**: No clear pattern or missing data
  - Target to predict for the test set

## Data Characteristics

### Ticker Distribution
- Average samples per ticker: 460
- Min samples per ticker: 460
- Max samples per ticker: 460
- Each ticker has time series data spanning the full date range

### Data Types
- Numeric columns: 5
- Categorical columns: 3

## Evaluation Metric
Your submissions will be evaluated based on classification accuracy.

## Submission Format
Submit a CSV file with exactly 828 predictions in the format:
```
id,class_label
0,H
1,L
2,None
3,H
4,L
...
```

**Important**: Ensure your predictions only contain valid class values: H, L, None

## Important Notes About ticker_id
- **ticker_id** is crucial for understanding the data structure
- Each ticker (001, 002, etc.) represents a different financial instrument
- Time series patterns may vary significantly between tickers
- Consider both:
  - **Within-ticker patterns**: How each instrument behaves over time
  - **Cross-ticker patterns**: Common behaviors across different instruments

## Data Preprocessing Notes
- All financial instruments have been anonymized as ticker_id (001, 002, etc.)
- The dataset has been cleaned and preprocessed
- Sensitive price and return information has been removed
- Time series data maintains chronological order within each ticker
- Features have been normalized/standardized where appropriate

## Tips for Participants
1. **Temporal Modeling**: Explore time-based patterns within each ticker using the 't' column
2. **Ticker-Specific Features**: Consider creating features that capture each ticker's unique behavior
3. **Cross-Ticker Analysis**: Look for patterns that generalize across different instruments
4. **Feature Engineering**: The temporal dimension and ticker grouping offer rich opportunities
5. **Ensemble Methods**: Often work well on financial time series data
6. **3-Class Strategy**: Consider the semantic meaning of H/L patterns vs None when designing features
7. **Class Imbalance**: Handle the imbalanced nature of pattern detection (many None vs fewer H/L)
8. **Time Series Validation**: Use time-aware cross-validation strategies
9. **Ticker Leakage**: Be careful not to use future information when creating features
10. **Pattern Recognition**: Focus on features that can distinguish trend patterns (H vs L) from noise (None)

## Data Usage Strategy
Consider these modeling approaches:
- **Global Model**: Train one model on all tickers
- **Ticker-Specific Models**: Train separate models for each ticker
- **Hierarchical Models**: Combine global and ticker-specific approaches
- **Transfer Learning**: Use patterns from well-represented tickers for sparse ones

Good luck with your predictions!

Dataset prepared on: 2025-09-10 21:56:14
Total tickers: 6
Date range: 2023-04-03 to 2025-01-31
Class distribution: {'H': 82, 'L': 81, 'None': 2597}