Feature Engineering

Feature Engineering is the definitive clean build of a feature engineering workbench on the chart: four slots each pick a transform from a ten-entry library, apply an optional backward lag, and pass through a rolling scaler fitted strictly on trailing data. The pane shows what a model would actually see — features sharing one range — while the dashboard tags each slot's stationarity and whether anything still reads the forming bar.

How to Trade the Feature Engineering?

• Stretched readings: a feature crossing the upper or lower threshold (±2 z-units by default) flags an unusual value in the scaled space; both crossings fire alerts.
• Raw-level caution tag: the Price and Volume (Raw Level) transforms drift by design; the dashboard flags them.
• Timing row: “Forming bar live” means a lag-0 slot updates until the close; lag every slot for a strict no-lookahead read.

This is instrumentation, not signals — prototype inputs here before they feed anything from the machine learning family.

Feature Engineering Settings

• Feature 1–4 (all default on): slot toggles with per-slot transform (defaults Log Return, Vol-Normalized Return, Distance From MA (ATR), RSI) and Lag (default 0).
• Method (default Z-Score) and Window (default 100): the rolling scaler and its trailing fit window.
• Source (default close) and Return Horizon (default 1): input series and return span.
• Volatility Window (20), MA Length (50), ATR Length (14), RSI Length (14), Correlation Window (50), Relative Volume Length (20): transform parameters.
• Upper Threshold (2.0) and Lower Threshold (-2.0): event levels in the scaler's display units.
• Alerts on Bar Close Only (default on): evaluates crossings on completed bars.
• Show Dashboard (default on); per-slot colors, Line Width (1) and the Guides color cover styling.

Frequently Asked Questions

How is this different from feature selection?

Engineering manufactures the candidates; selection prunes them. This build creates and normalizes inputs, while Feature Selection ranks a candidate set and discards redundant or uninformative ones.

Why scale the features at all?

Unscaled features let the largest-ranged input dominate distance-based learners. The scaler is fitted on the trailing window and applied to the current value — mirroring the rule that scalers are fitted on training data alone.

What are the raw level transforms for?

Contrast. Price and volume levels drift across history, so their scaled lines wander in ways return-based features do not — making non-stationarity visible rather than theoretical.