Home / Libraries / ML4T Data / Docs
ML4T Data
ML4T Data Documentation
Unified market data acquisition from 19+ providers
Source on GitHub
Skip to content

Synthetic Provider

Provider: SyntheticProvider API Key: Not required Free Tier: N/A (generates data)

Overview

Generates synthetic OHLCV data for testing, demos, and development without requiring network access or API keys.

Best For: Testing, demos, development

Quick Start

from ml4t.data.providers import SyntheticProvider

provider = SyntheticProvider()

# Generate synthetic data
df = provider.fetch_ohlcv("DEMO", "2024-01-01", "2024-12-01", frequency="daily")

print(df.head())
# Synthetic OHLCV data with realistic patterns

provider.close()

Configuration

provider = SyntheticProvider(
    base_price=100.0,
    annual_return=0.08,
    annual_volatility=0.20,
    calendar_mode="equity",
    seed=42,
)

calendar_mode="equity" emits a simplified weekday 09:30-16:00 UTC session. It does not model exchange holidays, early closes, or daylight-saving changes. Intraday timestamps are bar starts and exclude 16:00. Daily bars are labelled at 16:00 UTC. Use calendar_mode="continuous" for 24-hour UTC sessions that include weekends; continuous daily bars are labelled at 00:00 UTC. Weekly and monthly bars use period-end labels. Annual return and volatility use 261 weekdays for equity mode and 365 days for continuous mode, with 52 weekly or 12 monthly periods in either mode.

With a seed, the provider derives a symbol-specific stream using BLAKE2b and NumPy's PCG64 generator. Identical seed, symbol, date, frequency, model, and calendar inputs produce identical results across interpreter processes and supported platforms.

Use Cases

  1. Unit Tests: Test data pipelines without API calls
  2. Demos: Show functionality without credentials
  3. Development: Fast iteration without rate limits
  4. Documentation: Reproducible examples

Generated Data

  • Realistic OHLCV patterns (geometric Brownian motion)
  • Proper OHLC relationships (High >= Open, Close, Low)
  • Volume follows log-normal distribution
  • Equity-session or continuous UTC calendars

Learned Samples

LearnedSyntheticProvider converts pre-generated model samples into the same OHLCV contract. It accepts a non-pickle NumPy array with shape (n_samples, sequence_length, n_features) or an artifact directory containing samples.npy and metadata.json.

from ml4t.data.providers import LearnedSyntheticProvider

provider = LearnedSyntheticProvider.from_samples("timegan_sequences.npy", seed=42)
df = provider.fetch_ohlcv("SYNTH_TIMEGAN", "2024-01-01", "2024-12-31", "daily")

Executable model checkpoints are not loaded. Generate samples.npy during model training before constructing the provider.

See Also