Quant / 04

Unsupervised market-state analysis

An unsupervised study of market observations without predefined regime labels. The analysis first separates the main body from tail events, then examines continuous transitions, feature gradients in L01 and local groupings within L00. The resulting structure provides a reference for comparing factor behaviour across market conditions.

Trading days, 2010–2026: 3,615
Methods & tools: Bisecting K-Means · PCA geometry · Feature gradients

Research visuals

Initial clustering of all observations · Unsupervised clustering initially separates 3,615 market observations into C00 MainBody (3,360) and C01 EventTail (255). The dashed line marks the initial boundary fixed for this analysis. This split is the starting point for further hierarchical analysis, not a final classification of market regimes. — **Initial clustering of all observations**
Unsupervised clustering initially separates 3,615 market observations into C00 MainBody (3,360) and C01 EventTail (255). The dashed line marks the initial boundary fixed for this analysis. This split is the starting point for further hierarchical analysis, not a final classification of market regimes.
View full image ↗

Continuous transitions within MainBody · MainBody transitions from the L00 core through a mixed boundary region to the L01 core. The two ends are distinguishable, but their boundary overlaps. This weak separation and continuous structure do not support treating the groups as mutually exclusive, stable market regimes. — **Continuous transitions within MainBody**
MainBody transitions from the L00 core through a mixed boundary region to the L01 core. The two ends are distinguishable, but their boundary overlaps. This weak separation and continuous structure do not support treating the groups as mutually exclusive, stable market regimes.
View full image ↗

Feature gradients within L01 · Along the percentile axis, recovery, market breadth and drawdown-related features strengthen monotonically, while time below a previous high, choppiness and volatility decrease. Together, these features describe the continuous variation within L01. — **Feature gradients within L01**
Along the percentile axis, recovery, market breadth and drawdown-related features strengthen monotonically, while time below a previous high, choppiness and volatility decrease. Together, these features describe the continuous variation within L01.
View full image ↗

Stress and gap-tail observations within L00 · The most robust two-group split within L00 (k = 2) separates Mixed Body from the Stress / Gap-tail Pocket. This result describes a local distinction within L00 and does not establish the latter as an independent, stable market regime. — **Stress and gap-tail observations within L00**
The most robust two-group split within L00 (k = 2) separates Mixed Body from the Stress / Gap-tail Pocket. This result describes a local distinction within L00 and does not establish the latter as an independent, stable market regime.
View full image ↗