Basket construction via rank, correlation, and locked rules

Membership first, leftover inputs last

The archive builds a basket as a single procedure. Names must first qualify by rank. They must then pass a correlation grouping measured on an earlier window. Only after those membership rules, and the entry and exit that members will use, are written down can leftover inputs be exposed.

Editorial reading: those are construction gates, not optional extras. If membership is still free to move while inputs are searched, the search can rewrite who sits in the book.

A sample spike-fade inside the book

A range-style series can be defined as the larger of the absolute high-versus-prior-close move and the absolute low-versus-prior-close move, leaving out the same-bar high-low width. That reading is a modified-true-range: a gap-based range that keeps the larger of those two distances.

That series can be averaged and given a standard deviation over a chosen lookback, then scored by how often it meets a user-set distance goal.

One sample procedure treats a reading at least three standard deviations above that average as a spike setup, then fades the close with a next-bar market order. That short-horizon fade is a spike-fade.

The same sample exits with a three-bar trailing channel. A long stops if the low undercuts the three-day lowest low. A short covers if the high exceeds the three-day highest high.

Who ranks into the book

A basket can rank members by a five-period relative-strength oscillator and restrict entries to the top or bottom three ranks, exiting when a name leaves that rank band.

That step is rank-rotation. Candidates are ordered by a shared score and only those that stay inside the rank band are admitted for the system holding period.

Who is too correlated to sit together

Pairwise correlation over a multi-month window, including a Pearson-style coefficient, can be stored as a matrix so candidate names are grouped before a basket is specified.

The window used to measure those correlations is kept earlier than the bars reserved for a later system test, so membership is not chosen on the same sample used to evaluate the rules.

Raw pairs can be sorted by correlation, cut by a threshold in the range from minus one to one, and optionally capped by a maximum number of pairs. The list is then reduced so leftover unlinked clusters become separate candidate baskets.

Those leftover unlinked clusters are correlation-islands: names that correlate strongly with each other but not with other clusters after threshold filtering.

A market-context variant averages how index members correlate with a volatility-futures series and only then allows a channel breakout, using a low-correlation regime as permission to act. That is correlation-analysis used so one trade sits inside a regime-aware basket.

Which leftover inputs may later be optimized

After rank, correlation, entry, and exit steps are written as one procedure, lookback, rank-cutoff, and correlation-threshold inputs can be exposed for later optimization and backtest review.

That last step is system-optimization: vary rule inputs only after entry, exit, and abstention are already written as one testable procedure.

Editorial reading: the exposed inputs may change how strict a rank band or correlation cut is. They should not rebuild the membership list on the same bars used to score the rules.