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