Spectral graph topology with technical indicators for stock market regime detection

The financial markets generate an abundance of technical signals, such as price, momentum, volatility, volume and S/R levels, that are used to capture the activity of traders across time. Although
spectral graph clustering of sliding-window feature vectors has been shown to be better at detecting market regime than traditional baselines, existing methods limit feature space to mean return
and realised volatility and neglect much of the information contained in popular technical indicators. This paper considers the Indicator-Weighted Similarity Graph (IWSG) a principled extension of
the Temporal Similarity Graph (TSG) where each 8-day window is represented by a 14-dimensional
feature vector containing: Bollinger Band Width, %B, Relative Strength Index (RSI), MACD histogram, Stochastic %K/%D, Commodity Channel Index (CCI), Average True Range (ATR), On-Balance
Volume (OBV), and fractal support/resistance proximity score. A feature-adaptive Gaussian kernel
on this space produces the IWSG adjacency matrix whose symmetric normalised Laplacian is proved
positive semi-definite by means of an extended Dirichlet form argument, which provides richer spectral embedding. In experiments using ten NIFTY 50 constituents (January 2019–December 2023), the
IWSG performs better than the baseline TSG model in clustering by silhouette, Davies–Bouldin, time
consistency, and Calinski–Harabasz.