The alert record does not support a performance claim. It compares each signal’s directional hit rate with the unconditional forward behavior of the same instruments over the same period. There is no entry price, spread, slippage, contract selection, or executable fill in the analysis. The useful conclusion is narrower: the data describes what happened, not what can be traded.
The alert record does not support a performance claim. It compares each signal’s directional hit rate with the unconditional forward behavior of the same instruments over the same period. There is no entry price, spread, slippage, contract selection, or executable fill in the analysis. The useful conclusion is narrower: the data describes what happened, not what can be traded.
No signal beat its own baseline by a margin that survives scrutiny. Across seven family-by-horizon cells with a usable sample, only two showed a positive gap and five did not. The largest positive cell was SELL at 36 hours: 83.3% on 12 usable windows against a 55.0% baseline. Its interval clears that baseline, but the sample is too small to establish an edge.
No signal beat its own baseline by a margin that survives scrutiny. Across seven family-by-horizon cells with a usable sample, only two showed a positive gap and five did not. The largest positive cell was SELL at 36 hours: 83.3% on 12 usable windows against a 55.0% baseline. Its interval clears that baseline, but the sample is too small to establish an edge.
Pooled accuracy did not separate from the base rate at any horizon. The 12-hour read was 38.2% on 68 usable windows versus a 49.6% baseline; the 24-hour read was 35.3% versus 49.1%; and the 36-hour read was 54.1% versus 53.5%. A candidate selected from that record is unstable across horizons by definition.
Pooled accuracy did not separate from the base rate at any horizon. The 12-hour read was 38.2% on 68 usable windows versus a 49.6% baseline; the 24-hour read was 35.3% versus 49.1%; and the 36-hour read was 54.1% versus 53.5%. A candidate selected from that record is unstable across horizons by definition.
Path behaviour and checkpoint behaviour also disagree. At 36 hours, the average favorable excursion was 1.5% against 1.5% adverse on 37 usable windows. Reads frequently traveled a long way in the right direction and then were wrong at the graded mark, which means a single binary checkpoint test misrepresents the path.
Path behaviour and checkpoint behaviour also disagree. At 36 hours, the average favorable excursion was 1.5% against 1.5% adverse on 37 usable windows. Reads frequently traveled a long way in the right direction and then were wrong at the graded mark, which means a single binary checkpoint test misrepresents the path.
The binding constraint is data quality, not signal logic. Of 115 graded observations, 27 windows were stale and excluded from every hit-rate calculation. Seven alerts had only partially settled, and 23 alert instances overlapped another alert on the same instrument inside twelve hours. The record cannot repair those missing or non-independent observations.
The binding constraint is data quality, not signal logic. Of 115 graded observations, 27 windows were stale and excluded from every hit-rate calculation. Seven alerts had only partially settled, and 23 alert instances overlapped another alert on the same instrument inside twelve hours. The record cannot repair those missing or non-independent observations.
The one candidate worth carrying forward is SELL at the 36-hour checkpoint. It posted 83.3% on 12 usable windows, a +28.3 percentage-point gap versus the same-instrument baseline. That cell was selected after inspecting seven comparable family-by-horizon cells, so its confidence interval and apparent advantage should be read as a hypothesis, not a conclusion.
The one candidate worth carrying forward is SELL at the 36-hour checkpoint. It posted 83.3% on 12 usable windows, a +28.3 percentage-point gap versus the same-instrument baseline. That cell was selected after inspecting seven comparable family-by-horizon cells, so its confidence interval and apparent advantage should be read as a hypothesis, not a conclusion.
SELL COMBO specifically is less persuasive under forward test: 42.1% at 12 hours, 50.0% at 24 hours, and 83.3% at 36 hours on small samples. At these counts, the strongest positive print is a warning against overconfidence rather than proof of a durable signal.
Every number in the memo is limited by what the records themselves contain. Nothing was repaired, backfilled, or regraded to produce the review. There are no trigger prints missing from tape and no whole-market close prints excluded. The stale-window exclusions and incomplete terms are preserved as they occurred.
The retrospective method has been replaced with a frozen prospective ledger. Rules were fixed before the first signal: checkpoint timing is verified within fifteen minutes, delayed to 120 minutes, stale beyond that and excluded. Every qualifying signal is preserved, including failures, stale marks, and incomplete terms.
The research epoch began on November 14, 2026, with zero prospective signals captured so far. The 36-hour horizon was chosen after this retrospective review because it was the strongest of the three. That decision is disclosed in advance and recorded in the frozen manifest, not presented as a pre-existing result.
This memo does not establish that any signal is profitable or predictive, and it does not compare strategies. A hit rate above a base rate on a small overlapping sample is a hypothesis, not an edge. Read-only analysis only: no alert, observation, settlement window, or score was modified to produce this memo; no cached snapshot is presented as an executable market price.
This memo does not establish that any signal is profitable or predictive, and it does not compare strategies. A hit rate above a base rate on a small overlapping sample is a hypothesis, not an edge. Read-only analysis only: no alert, observation, settlement window, or score was modified to produce this memo; no cached snapshot is presented as an executable market price.