Methodology
TAPE is a research assistant that reads a chart, the news, and institutional flow, then publishes a verdict with explicit price levels. This page explains exactly how every read is recorded and graded — before you're asked to trust any of it.
Timestamps come first
Every read is written to an append-style ledger the moment it is issued, timestamped before its outcome exists. Grades are attached later by comparing the recorded read against price data that arrived after it. Nothing is back-filled, re-worded, or deleted after the fact.
Levels are published with the read
Each read ships three concrete prices alongside its verdict: an upside trigger, a downside trigger, and an invalidation level — the price at which the read is simply wrong. Publishing the invalidation up front is the difference between a falsifiable call and a horoscope. — stored reads carry persisted levels.
Grading is deterministic code, not judgement
Outcomes are resolved by a fixed price rule evaluated over the bars that followed the read, on the read's own timeframe. The same inputs always produce the same grade. The resolution rules, plainly:
Correct if price reaches the trigger in the called direction before hitting the invalidation level within the outcome window. Incorrect if invalidation is hit first.
The read claims neither level deserves trust yet. Correct if price closes through neither trigger within the window, or breaks both (whipsaw), or breaks one and immediately fails back (fakeout) — all cases where standing aside won. If one level breaks cleanly and holds, that is logged as a missed move — see below.
Partial-data and conflicting-evidence verdicts are graded by the same level rules where levels exist, and refused otherwise — they are never guessed into a bucket.
A close through the invalidation level always ends the read's window immediately, whatever else happened.
Missed moves are not hidden inside "wrong"
When a NO EDGE read stands aside and the market then moves cleanly through a level, we count it as a missed move — a separate bucket that never enters the accuracy denominator in either direction. It is the cost of conservatism, reported as its own number (— so far) instead of being laundered into "correct" (we said wait, and waiting was safe) or buried in "wrong".
Synthetic reads are excluded and flagged
Engineering verification runs (re-running stored reads to test the pipeline) are written to the same ledger but permanently marked as synthetic and excluded from every learning and accuracy surface. — such reads are currently flagged. They appear in this count precisely so you know they exist.
Incomplete windows are refused, not guessed
A read cannot be graded until its full outcome window of price data exists. The auto-grader checks every 30 minutes and refuses to grade any read whose window has not matured — — reads are pending right now for exactly that reason. No partial-window grades, ever.
What is gradeable and what is not
- Gradeable: any read carrying levels — directional calls (by trigger-vs-invalidation) and NO EDGE calls (by the level-break rule above).
- Not gradeable: reads issued before levels were a required output, and reads whose instrument has no usable bar history in the window. These stay inconclusive and are reported as such — an inconclusive read is a data limitation, not a win.