Technology / Layer 5 · Enrichment

Retrieval, not
prediction

Alpha Echo does not forecast. It finds the states that resembled this one, reports what followed them, and says nothing when the record is too thin to support a claim.
Image
E1 / Conditioning

Aggregate statistics hide the case in front of you

A detector with strong lifetime expectancy can be sharply negative in the conditions present right now. Averaging over all instances erases that. Conditioning on resemblance surfaces it.
Image
E2 / Levels

Targets and stops come from distributions, not ratios

The neighbor set records how far price ran in favor and how far it went against first. Those two distributions define the levels. A fixed reward-to-risk ratio is a preference. An excursion quantile is a measurement.
Image
E3 / Calibration

A confidence layer that is wrong about its own confidence is worse than none

Say 70%, be right about seven times in ten. Overconfidence corrupts every downstream sizing decision and does it hardest at the top of the range,  where size is largest.
Image
E4 / Abstention

Silence is a supported answer

Two conditions produce it: too few analogs, or analogs that disagree. Both are measurable before any estimate is issued.
Image
E5 / Learned components

One constraint governs everything added to this layer

Learned similarity, generated candidate features, and natural-language explanation of a neighbor set are all live questions here. Each faces the same test: if it degrades calibration, it does not ship. A more expressive estimator that is less honest about its own uncertainty is a downgrade.
Image

Alpha Echo outputs are research estimates. Past performance is not indicative of future results.
Full disclosures.