Skip to main content

Encrypted Features

SignalNet provides encrypted, obfuscated feature datasets for each tournament round.

What Are Encrypted Features?​

We source raw financial data (price, volume, fundamentals, sentiment, etc.) and transform it into an anonymized feature set:

  • ~98 features per stock
  • ~503 stocks (S&P 500 universe)
  • Features are normalized to [0, 1] range
  • Feature names are obfuscated (feature_001, feature_002, ...)
  • Random monotonic transformations applied (preserves rank order)
  • A few noise features included (for baseline testing)

Why Obfuscate?​

  1. Level playing field — success depends on modeling skill, not data access
  2. IP protection — prevents reverse-engineering of our data pipeline
  3. Anti-gaming — can't exploit known feature semantics
  4. Legal safety — no licensing issues with redistributing transformed data

Downloading Features​

Python SDK​

from signalnet import Tournament

t = Tournament(api_key="your_key")

# Current round features
df = t.get_features()
# DataFrame: index=ticker, columns=feature_001...feature_098

# Historical features (for backtesting)
df_hist = t.get_features(round_id=42)

# Feature metadata
meta = t.features.metadata()
print(f"Features: {meta.num_features}")
print(f"Stocks: {meta.num_stocks}")
print(f"Updated: {meta.last_updated}")

REST API​

# Download as CSV
curl -H "Authorization: Bearer YOUR_API_KEY" \
  https://api.signalnet.io/v1/features/current \
  -o features.csv

# Download as Parquet (smaller, faster)
curl -H "Authorization: Bearer YOUR_API_KEY" \
  https://api.signalnet.io/v1/features/current?format=parquet \
  -o features.parquet

Feature Statistics​

Each round's feature dataset includes basic statistics:

features.describe()
#        feature_001  feature_002  ...  feature_098
# count      503.0       503.0         503.0
# mean         0.50        0.49          0.51
# std          0.29        0.28          0.30
# min          0.00        0.00          0.00
# max          1.00        1.00          1.00

Tips for Modeling​

  • Don't assume feature meaning — feature_042 might be momentum this round and volatility next round
  • Use rank-based methods — since features are monotonically transformed, rank information is preserved
  • Watch for noise features — some features are pure random noise. Good feature selection helps
  • Ensemble your own models — combine multiple approaches for robustness
  • Don't overfit — practice round performance ≠ live performance