Feature Selection

Feature Selection runs the screen most stacks skip: deciding which inputs survive. Ten standard candidates are scored by correlation against the forward return over a training window, then a greedy filter applies the discipline of feature selection as written — rank by training score, enforce the minimum, discard near-duplicates, stop at the cap. Survivors are re-scored on a held-out window selection never saw, and the pane plots the selected set's mean held-out correlation against the candidate pool's.

How to Trade the Feature Selection?

• Selected vs pool line: the selected set holding above the gray pool line means selection adds value out of sample; crossings both ways are alerted.
• Train r vs Test r: a training score that collapses in the held-out column was likely a false discovery — the dashboard shows the pair per feature.
• Stab %: the share of recent walk-forward snapshots selecting a feature. A flickering pick is an arbitrary survivor among near-duplicates; set changes fire an alert.

Feature Selection Settings

• Prediction Horizon (default 10) and Label (default Forward Return): what candidates must predict; Forward Direction uses the sign only, muting outliers.
• Max Selected Features (K) (default 4), Minimum |Score| (default 0.05) and Redundancy Threshold (default 0.7): the screen's three gates.
• Training Window (default 500), Held-Out Window (default 100), Walk-Forward Step (default 50) and Stability Memory (default 20): the evaluation scheme.
• Candidate pool toggles with lookbacks — RSI (14), Stochastic %K (14), CCI (20), Rate of Change (10), Bollinger %B (20), SMA Distance (50), MACD Histogram (12/26/9), ATR % (14), Return Volatility (20), Volume Z-Score (20) — all on.
• Show Dashboard (on) with Location (Top Right) and Size (Small); Gradient Fill (on).

Frequently Asked Questions

How does this relate to Feature Engineering?

Feature Engineering creates candidate inputs; this build decides which of those candidates a model keeps. The pool here is taken as given — several candidates are near-copies precisely so the redundancy filter has real work to do.

What does "Redundant → RSI" mean in the Status column?

The candidate correlated with an already-selected feature beyond the Redundancy Threshold and was discarded as a near-duplicate; the column names the survivor it duplicates.

Why is nothing selected sometimes?

No enabled candidate cleared the Minimum |Score| on the training window, and a dedicated alert says so. When many features are screened, some correlate with returns by pure chance — the floor keeps those out.