ts-quant 🚀

GPU-Accelerated Quantitative Feature Extraction for Financial Time Series

Extract 2000+ quantitative features from OHLCV stock data using 5 powerful engines — all running on GPU via PyTorch.

🎯 Why ts-quant?

Feature	tsfresh (CPU)	ts-quant (GPU)
Speed	~60s per stock	~0.3s per stock
Memory	OOM on large data	VRAM-managed, crash-proof
Features	794 (statistical only)	2000+ (statistical + wavelets + rocket + signatures)
GPU	❌	✅ PyTorch/CUDA

📦 Installation

pip install ts-quant

Requirements: Python 3.9+, PyTorch 2.0+ (with CUDA for GPU acceleration)

🚀 Quick Start

import pandas as pd
from ts_quant import generate_features

# Load your OHLCV data
df = pd.read_csv('stocks.csv')
# Expected columns: symbol, date, close, volume, open, high, low

# Extract features (GPU)
df_features = generate_features(
    df,
    device='cuda',              # 'cpu' for CPU-only
    tsfresh_mode='comprehensive',
    window_sizes=[20],
)

print(f"New features: {df_features.shape[1] - df.shape[1]}")

🔧 5 Feature Engines

Engine A: MultiRocket 🎯

Random convolutional kernels with multi-scale dilations.

from ts_quant import RocketEngine

engine = RocketEngine(n_kernels=250)
features, names = engine.extract(x)  # x: [B, T]
# → 1000 features (250 kernels × 4 pooling ops)

Engine B: Catch22 📊

22 canonical time-series features (ACF, DFA, entropy, etc.).

from ts_quant import Catch22Engine

engine = Catch22Engine()
features, names = engine.extract(x)
# → 22 features per window

Engine C: Path Signatures ✍️

Ordered interaction features via iterated integrals.

from ts_quant import SignaturesEngine

engine = SignaturesEngine(depth=3, channels=['close', 'volume'])
features, names = engine.extract(x_multivariate)  # [B, T, d]
# → d + d² + d³ features

Engine D: Wavelets 🌊

Discrete wavelet transform (Haar, db4, db2, sym4).

from ts_quant import WaveletsEngine

engine = WaveletsEngine(wavelet_types=['haar', 'db4'])
features, names = engine.extract(x)
# → 62 features (energy, entropy, mean, std, max per level)

Engine E: Tsfresh Complete 📈

63 mathematical functions, 361 features in comprehensive mode.

from ts_quant import TsfreshEngine

engine = TsfreshEngine(mode='comprehensive')
features, names = engine.extract(x)
# → 361 features (statistics, ACF, FFT, entropy, trend, ...)

⚡ VRAM Management

ts-quant automatically manages GPU memory to prevent OOM crashes:

# Works on any GPU size (8GB - 80GB)
df_features = generate_features(
    df,
    device='cuda',
    max_vram_gb=8,  # Optional: manual VRAM limit
)

🔬 Feature Selection

Built-in redundancy removal:

from ts_quant import auto_select_features

selected, names = auto_select_features(
    features_tensor,
    feature_names,
    correlation_threshold=0.95,  # Remove highly correlated
)

📁 Input Format

Your DataFrame should have this structure:

symbol	date	open	high	low	close	volume
BBCA	2024-01-02	9500	9600	9400	9550	1000000
BBRI	2024-01-02	5200	5300	5100	5250	2000000

🏗️ Architecture

ts-quant/
├── ts_quant/
│   ├── api.py              # Orchestrator — generate_features()
│   ├── core/
│   │   ├── memory_manager.py   # Dynamic VRAM management
│   │   ├── tensor_utils.py     # DataFrame ↔ Tensor conversion
│   │   └── windowing.py        # GPU sliding windows
│   ├── engines/
│   │   ├── rocket.py           # Engine A: MultiRocket
│   │   ├── catch22.py          # Engine B: Catch22
│   │   ├── signatures.py       # Engine C: Path Signatures
│   │   ├── wavelets.py         # Engine D: Wavelets
│   │   └── tsfresh_core.py     # Engine E: Tsfresh Complete
│   └── utils/
│       ├── config.py           # Default configurations
│       ├── validation.py       # Input validation
│       └── feature_selection.py # Redundancy removal
└── tests/

📄 License

MIT License