| name | signal-audit |
| description | Measure predictive information content of signals using mutual information, correlation, and optimal lag analysis. Use before building models, when adding data sources, debugging signal decay, or selecting features. Covers MI estimation, regime-conditional analysis, and decay tracking. |
| requires | ["measurement-infrastructure"] |
| user-invocable | false |
Signal Audit Skill
Purpose
Systematically measure the predictive information content of all available signals before building models. This prevents wasting effort on low-value features and identifies high-value signals you might be ignoring.
When to Use
- Before building any new predictive model
- When adding a new data source (cross-exchange feed, new API field)
- Quarterly review of signal value decay
- Debugging why a model stopped working
- Deciding which features to include in a model
Prerequisites
measurement-infrastructure implemented (for outcome data)- Historical market data with candidate signals
- Defined prediction targets
Core Concept: Mutual Information
Mutual information measures how many bits of information signal X provides about target Y:
I(X; Y) = H(Y) - H(Y|X)
Where H is entropy.
Key properties:
- I(X; Y) ≥ 0 (always non-negative)
- I(X; Y) = 0 if and only if X and Y are independent
- Works for non-linear relationships (unlike correlation)
- Units are bits (or nats if using natural log)
Signal Catalog
Book-Derived Signals
- Basic:
spread_bps, mid_price
- Imbalance:
microprice_imbalance (L1), book_imbalance_l5 (top 5), book_pressure (weighted depth asymmetry)
- Depth:
depth_at_1bps, depth_at_5bps, depth_at_10bps
- Shape:
book_slope_bid, book_slope_ask (depth increase rate away from mid)
Trade-Derived Signals
- Volume imbalance:
trade_imbalance_{1s,10s,60s} (net signed volume)
- Intensity:
trade_arrival_rate, volume_rate
- Size distribution:
avg_trade_size, trade_size_std, large_trade_count_1m (> 2 sigma)
- Aggression:
aggressor_imbalance
Hyperliquid-Specific Signals
- Funding:
funding_rate, funding_rate_change_{1h,8h}, predicted_funding_rate, time_to_funding_settlement_s
- Open Interest:
open_interest, open_interest_change_{1m,5m,1h}, oi_momentum
- Vault:
hlp_vault_position
Cross-Exchange Signals
- Binance:
binance_mid, binance_spread_bps, binance_hl_basis_bps
- Lead indicators:
binance_return_{100ms,500ms,1s}
- Volume ratio:
binance_volume_ratio (Binance / HL)
Composite Signals
- Interactions:
funding_x_imbalance, oi_x_funding, basis_x_imbalance
- Momentum:
price_momentum_{1m,5m}, volume_momentum
See implementation.md for full struct definitions.
Prediction Targets
- Direction:
PriceDirection{1s,10s,60s} -- sign of future return
- Magnitude:
AbsReturn{1s,10s}, Volatility1m
- Fill-related:
FillWithin{1s,10s}, TimeToNextFill
- Adverse selection:
AdverseOnNextFill, InformedFlow
- Regime:
RegimeTransition (will regime change in next minute?)
Mutual Information Estimation
k-NN Estimator (Kraskov et al.)
For continuous variables, use the k-nearest-neighbor estimator (estimate_mutual_information). Normalizes inputs to [0,1], builds k-d trees in joint and marginal spaces, then computes MI via digamma functions on neighbor counts. Typical k = 3-10.
For Binary Targets
Use the simpler binned estimator (estimate_mi_binary_target). Bins the continuous variable, counts joint/marginal frequencies, and computes MI from the frequency table.
See implementation.md for full code of both estimators.
Signal Analysis Framework
SignalAnalysisResult captures per-signal metrics:
| Field | Description |
|---|
mutual_information_bits | Raw MI in bits |
mutual_information_normalized | MI / H(Y), fraction of target entropy explained |
correlation, correlation_abs | Linear relationship (for comparison) |
auc_roc | ROC AUC for binary targets |
optimal_lag_ms, mi_at_optimal_lag | Best lag and MI at that lag |
mi_by_regime, regime_variance_ratio | Per-regime MI and max/min ratio |
mi_trend_30d | Stationarity trend |
analyze_signal() computes MI, correlation, optimal lag (candidate lags: -500ms to +500ms), and per-regime MI (min 100 samples per regime). find_optimal_lag() sweeps candidate lags and returns the one with highest MI.
See implementation.md for full code.
Signal Audit Report
generate_signal_audit_report() analyzes all signals against a target, sorts by MI descending, and produces a table with actionable insights. The report flags: (1) highest-MI signal, (2) regime-conditional signals (variance ratio > 2x), (3) signals with 20%+ more MI at a lag, (4) high-correlation but low-MI signals (possibly spurious).
See implementation.md for full code.
Example Report Output
=== Signal Audit Report ===
Target: PriceDirection1s
Signal MI (bits) Corr Opt Lag Regime Var
─────────────────────────────────────────────────────────────────────
binance_return_100ms 0.0890 0.31 -150ms 2.3x
trade_imbalance_1s 0.0670 0.24 0ms 1.4x
microprice_imbalance 0.0450 0.19 0ms 1.2x
funding_x_imbalance 0.0410 0.15 0ms 3.1x
open_interest_change_1m 0.0230 0.08 0ms 1.1x
book_pressure 0.0180 0.11 0ms 1.3x
funding_rate 0.0120 0.05 0ms 1.8x
ACTIONABLE INSIGHTS:
1. binance_return_100ms has highest MI (0.089 bits) - prioritize if not already used
2. funding_x_imbalance has 3.1x higher MI in some regimes - consider regime conditioning
3. binance_return_100ms has 20%+ more MI at -150ms lag - incorporate lag in feature
Signal Quality Thresholds
Typical thresholds:
| Parameter | Typical Value | Purpose |
|---|
min_mi_bits | 0.01 | Minimum MI to include in model |
min_samples | 1000 | Minimum samples for reliable MI estimate |
max_regime_variance | 3.0 | Above this, require regime conditioning |
min_correlation | 0.05 | Sanity check floor |
filter_signals() keeps signals above MI and correlation thresholds. flag_regime_conditional() flags signals with regime variance ratio above threshold.
See implementation.md for full code.
Tracking Signal Decay
Signals lose value over time as:
- Other participants discover them
- Market structure changes
- Regime shifts
compute_signal_decay() fits a linear regression on historical MI values over time and computes a half-life (days until MI drops 50%). Action thresholds:
- < 30 days: URGENT -- signal decaying rapidly, investigate or replace
- < 90 days: WARNING -- monitor closely
- >= 90 days: OK -- signal stable
See implementation.md for full code.
Dependencies
- Requires: measurement-infrastructure (for outcome data), historical market data
- Enables: All model skills (by identifying which features to use)
Common Mistakes
- Using correlation instead of MI: Correlation misses non-linear relationships
- Not checking lag: Some signals lead the target and are more valuable at a lag
- Ignoring regime conditioning: A signal useless overall might be gold in specific regimes
- Not tracking decay: Signals that worked last year might be worthless now
- Too few samples: MI estimation needs 1000+ samples for reliability
Next Steps
After signal audit:
- Select top signals for your target (MI > 0.01 bits)
- Flag regime-conditional signals for special handling
- Incorporate optimal lags into feature engineering
- Read the relevant model skill to build the predictor
- Set up decay tracking for production monitoring
Supporting Files
- implementation.md -- All Rust code: signal structs, MI estimators, analysis framework, report generation, quality thresholds, decay tracking
