Evaluating weekday-in-month filters for index day trades

A monthly slot rather than a yearly map

Older seasonal tests pair many entry and exit dates inside a year. This evaluation instead keys permission to trade off weekday position inside the month, because scheduled reports and similar events land on those monthly slots rather than on a fixed day of the year.

The weekday-in-month code is a two-digit label for the nth weekday inside a calendar month. It is used to permit or block a trade rather than to pick a day of the year.

How each session is coded

Each session is coded as a two-digit weekday-in-month code, with the first digit the occurrence in the month and the second a weekday from 1 to 5, running as high as a fifth-week 55. Fifth-week codes are pooled into one greater-than-50 gate because those sessions are scarce.

Eight simple open-relative, same-session entry patterns were each screened against those codes instead of being left to fire on every date. The calendar-position filter allows an otherwise simple open-relative entry only on selected weekday-in-month codes.

The hold-out split as a keep-or-drop test

The continuous index-futures series was split so April 1982 through December 1989 is the fitting window and January 1990 through March 1996 is the hold-out split. That reserved later window is used to keep a rule only when win rate and average trade stay comparable to the fitting window.

A pattern-and-filter pairing was dropped when win rate or average trade diverged across that split. Three of the eight patterns were removed for large average-trade gaps even when the hold-out window improved.

What the kept rules share

Kept rules use a same-session close exit, leaving the position at that session close so entry, calendar gate, and exit stay one day-trade procedure. They issue the signal the day before so the weekday-in-month code names the prior session, allow only one position in a given direction at a time, and carry no protective stop in the base procedure.

Stability checks and leftover tilt

Matching quality on both sides of the hold-out split is treated as a check against an after-the-fact target, the error of circling the densest historical hits and then treating that circle as a planned method. Unseen live data is named as the stronger check, and the whole construction is still allowed to be an elaborate fit.

After two short patterns and one long pattern were removed, the retained evaluation book is long-heavy. A closer fixed-dollar stop reduced combined net and raised drawdown versus the same-session close exit, while larger stops scored better in the optimizer run. Editorial reading: that long-side tilt and the stop-size contrast are leftovers of the evaluation book, not a claim that the screen is a finished edge.