Track Record

The same calls, measured four ways

Holding period changes the answer, so all four readings are shown. “As stated” honours each call’s own timeframe but only short-horizon setups have closed yet, making it a slice of momentum trades from one early window rather than a cross-section. The fixed rulers grade every closed call the same way, at the cost of cutting longer theses short.

By direction

Bearish calls shipped long-shaped levels until 2026-09-06, so short history is thin by construction.

By conviction tier

Scores normalised to 0–100 across the scale change of 2026-05-13; without that the early era misfiles into <50.

By market cap

Today's market cap applied backwards — point-in-time cap is in no snapshot.

By sector

By tactical tag

Tags entered the published payload only in September 2026; validate_system Layer 3 replays a longer history.

By alpha signal

By market structure

By action priority

How trades ended

Method

unit

One trade per signal episode: consecutive run dates carrying the same verdict for a ticker collapse into a single trade, entered on the day the call flipped and managed to that day's levels.

entry

Next session's OPEN after the run date.

exit

First touch of target or stop on daily bars, else the close of the final session in the stated timeframe.

horizon

Fixed at 20 sessions.

same bar rule

When target and stop both fall inside one daily bar the STOP is taken. Daily data cannot order intraday touches.

control

For every trade, the identical percentage target, percentage stop and holding window applied to every other name in that day's universe. The reported edge is the paired difference.

clustering

Calendar month. All names share one market on a given day.

r multiple

(exit - entry) / |entry - stop|, signed by trade direction.

costs

None modelled. No commission, spread, slippage or borrow cost.

Limitations

• SURVIVORSHIP — stockList.csv is today's watchlist. LAZR, NKLA, CFLT and DM were removed because they died, so calls that preceded a collapse are systematically absent and every figure here is biased upward.
• COMPOSITION — the universe grew from 5 to 103 names over the sample. Early months are a different, far narrower measurement than recent ones.
• OPEN CALLS — a call is graded only once its holding window has fully closed, so the most recent stretch of calls is still running and absent from every figure here. Grading each call on its own stated timeframe instead would shrink the sample from 152 to 50 and select for short-horizon momentum setups, because horizon length is assigned by setup type — which is why the primary view uses one fixed ruler and shows the as-stated reading beside it.
• NO COSTS — gross of commission, spread, slippage and borrow. A real book keeping this record would have returned less.
• SHORT-SIDE HISTORY — until 2026-09-06 every SHORT shipped long-shaped levels (target above spot, stop below), so bearish calls before that date cannot be graded as trades and are excluded as SHORT_LEVELS_INVERTED.
• PRICE ADJUSTMENT — bars are dividend- and split-adjusted at fetch time (auto_adjust=True, shared with validate_system), while each call's target and stop are the raw prices published on the day. A dividend paid after a call therefore shifts that name's whole history a little relative to its own levels, and a recent bar can still be revised by the vendor — so a trade sitting within cents of a barrier can flip between runs. Observed once: one trade moved TIMEOUT to STOP across two runs four days apart, with no other figure changing. The nightly job regrades from scratch, so this settles rather than accumulates.
• MARKET CAP is today's, applied backwards — point-in-time cap is in no snapshot and is unrecoverable for delisted names.
• SAMPLE — the paired edge needs 12 monthly clusters before it can be called significant. Below that the honest verdict is INSUFFICIENT_EVIDENCE regardless of how good or bad the point estimate looks.