Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Representation learning for regime detection in financial markets
We investigate financial market regime detection from the perspective of deep representation learning of the causal (reflexive) information geometry underpinning complex (multi-scale) dynamical traded asset systems using an emergent hierarchical correlation structure to characterise evolving macroec...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Department of Statistical Sciences
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | We investigate financial market regime detection from the perspective of deep representation learning of the causal (reflexive) information geometry underpinning complex (multi-scale) dynamical traded asset systems using an emergent hierarchical correlation structure to characterise evolving macroeconomic market phases. Specifically, we assess the robustness of three toy models: SPD Matrix Network (SPDNet), SPD Matrix Network with Riemannian Batch Normalisation (SPDNetBN) and U-shaped SPD Matrix Network (U-SPDNet) whose architectures respect the underlying Riemannian manifold of input block hierarchical Symmetric Positive Definite (SPD) correlation matrices by employing Log-Euclidean Metric (LEM)s. Market phase detection for each model is carried out using three data configurations: i.) Randomised Johannesburg Stock Exchange (JSE) Top 60 data, ii.) synthetically-generated block hierarchical SPD matrices, and iii.) chronology-preserving block-resampled JSE Top 60 data. We show that using a singular performance metric is misleading in our financial market use cases. We confirm that U-SPDNet performs improved latent feature extraction with better classification performance in stressed and rally market phases, despite achieving lower Out-of-Sample (OOS) backtest scenario accuracy than that of the benchmark SPDNet. The SPDNet-based models fail in capturing latent reflexive spatio-temporal block hierarchical correlation dynamics and deliver corner solutions across all input data sets. The U-SPDNet is promising in terms of its utility in regime dependent portfolio optimisation strategy generation as a model better-suited to capturing latent block hierarchical correlation structures arising from lead-lag causal feedback information loops that often drive the evolution of evolving market regimes |
|---|