Alpaca Algo Trader Plus Data Access

Experiment Overview

Context

Default Alpaca free tier limits historical data to ~2 years. The Algo Trader Plus subscription ($99/month) provides:

• 5+ years of historical bars for equities
• Extended crypto history
• Higher API rate limits

This enables training on 4 years of data (1460 days) for more robust models.

Verified Workflow

1. Configure Lookback Period (notebook)

# Data configuration
LOOKBACK_DAYS = 1460              # 4 years of data (Algo Trader Plus subscription)
# LOOKBACK_DAYS = 730             # 2 years (free tier limit)

2. Data Fetcher Configuration

from alpaca_trading.data.fetcher import DataFetcher

# Initialize with API keys
fetcher = DataFetcher(keys_file='API_key.txt')

# Fetch extended history
df = fetcher.get_bars(
    symbol='AAPL',
    timeframe='1Hour',
    lookback_days=1460,  # 4 years
    use_cache=True
)

# Expected: ~10,000 1Hour bars (6.5 hours/day * ~252 days/year * 4 years)
print(f'Fetched {len(df):,} bars')

3. API Key Environment Variables

# Set in environment or notebook
import os
os.environ['APCA_API_KEY_ID'] = 'your_key_id'
os.environ['APCA_API_SECRET_KEY'] = 'your_secret_key'

# Or use keys file
os.environ['ALPACA_KEYS_FILE'] = 'API_key.txt'

4. Cache Configuration for Large Data

# Google Drive cache for Colab
DRIVE_DATA_DIR = '/content/drive/MyDrive/Colab_Projects/training_data'
SELECTION_CACHE_EXPIRY_DAYS = 3   # Shorter for fresh rankings
TRAINING_CACHE_EXPIRY_DAYS = 7    # Longer for stability

# Save 4 years of data to cache
save_to_cache(symbol, df, timeframe, lookback_days=1460)

5. Validate Data Availability

# Check date range
print(f'Date range: {df.index[0]} to {df.index[-1]}')

# Expected for 4 years:
# Start: ~2021-01-XX
# End: 2024-12-XX

# Verify bar count
expected_bars = 6.5 * 252 * 4  # ~6,552 for 1Hour
print(f'Expected ~{expected_bars:,.0f} bars, got {len(df):,}')

Data Source Comparison

Failed Attempts (Critical)

Final Parameters

# Recommended settings for Algo Trader Plus
lookback_days: 1460             # 4 years
min_bars_required: 500          # For selection
min_bars_training: 2000         # For training
cache_expiry_days: 7            # Weekly refresh
timeframe: '1Hour'              # Base timeframe

# Expected bar counts (1Hour, equities)
1_year: ~1,638 bars
2_years: ~3,276 bars
4_years: ~6,552 bars
5_years: ~8,190 bars

Key Insights

• 4 years is the sweet spot: Enough data for robust training, available for most symbols
• yfinance is unreliable for crypto: Zero-volume bars break training
• Cache aggressively: 4-year fetches are slow without caching
• Not all symbols have full history: Check bar count after fetch
• Cost-benefit: $99/month for quality data saves debugging time

Subscription Tiers

References

• notebooks/training.ipynb: LOOKBACK_DAYS configuration
• alpaca_trading/data/fetcher.py: DataFetcher class
• Alpaca Subscription Plans