Skip to content
StockMarketAgent
Methodology · Historical analogs

Historical analog explorer

A nearest-neighbor search across multi-decade index history that surfaces prior regimes most similar to the current month and reports the subsequent 12-month outcomes.

Specification

Historical analog explorer — operational spec

Grounds the regime read in actual outcomes rather than theory. A cluster of analogs delivering double-digit drawdowns is a meaningful prior even if no single analog is conclusive.

Inputs
  • Monthly z-scores of the regime composite, breadth health, valuation dispersion, revision diffusion, and a sector-rotation pattern signature
  • Multi-decade history of the same z-scored series
  • Subsequent 12-month total return and maximum drawdown for every prior month-end
Computation
  • Distance between the current month and each prior month is the weighted Euclidean distance across the z-scored sub-components.
  • Weights are equal by default; the explorer publishes a sensitivity table for alternate weighting schemes.
  • The five lowest-distance analogs are surfaced.
  • Analog-implied 12-month return is the inverse-distance-weighted mean of the five subsequent outcomes.
  • Confidence band is the standard deviation of the five outcomes, scaled by the spread of distances.
Outputs
  • Top-five nearest analogs with date, distance, and subsequent 12-month return / drawdown.
  • Analog-implied 12-month return with confidence band.
  • Cluster diagnostic flagging cases where the five analogs cluster on the same outcome (high-conviction prior) versus spread (low-conviction prior).
Limitations
  • Small samples. Five analogs is not a forecast; it is a distribution.
  • Market structure has changed materially over the multi-decade lookback; a 1980s analog of a 2020s regime is loose at best.
  • The distance metric is a heuristic. Different sub-component weights produce different analog sets, and no weighting is empirically optimal across all regimes.
  • The explorer is most useful when paired with the scenario probability tree; in isolation it can mislead.

Frequently asked

Is the analog-implied return a forecast?
No. With a sample of five, the implied return is a distribution-of-outcomes summary, not a forecast. It is most useful when paired with the scenario probability tree — when the tree's bear-branch fair value lines up with a tight cluster of bearish analogs, the conviction on the bear case rises.
Why nearest-neighbor instead of a regression?
A regression assumes a stable functional form across regimes; multi-decade index history has multiple structural breaks (gold standard, fixed exchange rates, deregulation, monetary regime changes) that defeat that assumption. Nearest-neighbor is non-parametric — it makes no assumption about the relationship between regime sub-components and subsequent returns. That is a lower-power method but a more honest one.
How are the sub-component weights chosen?
Equal by default. The explorer publishes a sensitivity table that re-runs the analog search under alternate weights so the reader can see how robust the top-five list is to that choice. A robust top-five (same date stamps under multiple weight schemes) is materially more informative than a fragile one.
See the full methodology hub for the rest of the model registry, or open the glossary entry for the headline definition.