To predict tomorrow’s SPY direction, a model needs inputs that describe market conditions. An independent variable is one such input: a measured value or derived feature supplied to the model, distinct from the direction it is trying to predict. Useful inputs expose an economic signal; a more complex algorithm cannot compensate for weak raw prices. section 10.4
Build inputs around the decision and the data available to support it:
- For market direction, derive momentum, volatility, and volume indicators across ETF assets, then add rolling correlations and cross-asset spreads. The illustrated pipeline expands 40 price-and-volume series into approximately 208 features. section 10.4.2
- For fraud detection, use short-window activity counts, deviations from a user’s behavior, shared-device links, and session or device-trust signals. Real-time decisions require timely feature pipelines. section 7.4.2
- For credit decisions, calculate recent application activity through streaming pipelines and validate critical features for missing values or distribution shifts. Apply masking and other data controls across environments. section 4.3.1
Table 10.3 Feature categories for institutional-grade strategies
|
Category
|
Purpose
|
Examples
|
Quant insight
|
|---|---|---|---|
|
Cross-Asset Signals
|
Capture regime shifts
|
SPY-TLT correlation, CBOE Volatility Index (VIX)-equity spreads
|
Relationships break during crises.
|
|
Momentum Cascades
|
Multiple time frame trends
|
5d versus 20d momentum divergence
|
Short-term reversals cause long-term persistence.
|
|
Volatility Regimes
|
Risk environment
|
VIX term structure, realized volatility clustering
|
Volatility predicts volatility and affects correlations.
|
|
Economic Proxies
|
Macro factor exposure
|
Dollar strength versus emerging markets (EM), commodity momentum
|
ETFs are used as real-time economic sensors.
|
The feature-category table links candidate market inputs to the regimes, trends, and economic relationships they are intended to capture. section 10.4.2
Transaction features for a graph model
A transaction-fraud graph can give each transaction three numerical inputs: amount divided by 1,000, hour divided by 24, and weekday divided by 6. Shared user or device IDs define connections separately. A graph neural network combines each transaction’s features with information from connected transactions to score normal versus fraudulent activity. section 9.4.2
Inputs need not be purely numerical market or transaction records: structured, evidence-backed news analysis can produce quantifiable policy-tone, supply-risk, and novelty features. Keep their evidence auditable, and use evaluation, guardrails, and human oversight rather than treating generated signals as risk-free. section 11.5
