Paper: arXiv 2512.21823
Authors: Siddhartha Srinivas Rentala
Abstract
This study investigates the efficacy of Conditional Restricted Boltzmann Machines (CRBMs) for modeling high-dimensional financial time series and detecting systemic risk regimes. We extend the classical application of static Restricted Boltzmann Machines (RBMs) by incorporating autoregressive conditioning and utilizing Persistent Contrastive Divergence (PCD) to incorporate complex temporal dependency structures. Comparing a discrete Bernoulli-Bernoulli architecture against a continuous Gaussian-Bernoulli variant across a multi-asset dataset spanning 2013-2025, we observe a dichotomy between generative fidelity and regime detection. While the Gaussian CRBM successfully preserves static asset correlations, it exhibits limitations in generating long-range volatility clustering. Thus, we analyze the free energy as a relative negative log-likelihood (surprisal) under a fixed, trained model. We demonstrate that the model’s free energy serves as a robust, regime stability metric. By decomposing the free energy into quadratic (magnitude) and structural (correlation) components, we show that the model can distinguish between pure magnitude shocks and market regimes. Our findings suggest that the CRBM offers a valuable, interpretable diagnostic tool for monitoring systemic risk, providing a supplemental metric to implied volatility metrics like the VIX.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 6.5/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper features dense statistical physics derivations, energy-based formulations, and complex probabilistic calculus, indicating high mathematical sophistication. It also includes concrete empirical analysis with multi-asset datasets (2013-2025), comparisons between Bernoulli and Gaussian architectures, and implementation-focused techniques like Persistent Contrastive Divergence, demonstrating solid empirical implementation.
Research Flowchart
flowchart TD
A["Research Goal:<br>Model financial time series<br>and detect systemic risk regimes<br>using Conditional RBMs"] --> B["Data & Models:<br>Multi-asset dataset<br>2013-2025<br>Discrete Bernoulli CRBM<br>Continuous Gaussian CRBM"]
B --> C["Methodology:<br>Autoregressive Conditioning<br>Persistent Contrastive Divergence PCD"]
C --> D{"Computation:<br>Analyze Free Energy<br>as relative negative log-likelihood"}
D --> E["Decomposition:<br>Quadratic Component<br>Structural Component"]
E --> F{"Findings/Outcomes"}
F --> G["Generative Fidelity<br>Gaussian CRBM preserves static correlations<br>but struggles with volatility clustering"]
F --> H["Systemic Risk Metric<br>Free energy serves as robust regime stability metric<br>distinguishing magnitude shocks vs market regimes"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes Conditional Restricted Boltzmann Machines with autoregressive conditioning and uses the model's free energy as a robust unsupervised metric for systemic risk regime detection.
Source sentence
We demonstrate that the model’s free energy serves as a robust, regime stability metric.
Assumes 3
- The system temperature is set to T = 1, treating the model as a standard Energy-Based Model. model
Source sentence
In this study, we assume the system temperature T = 1 [5], treating the model as a standard Energy-Based Model (EBM) where learning corresponds to minimizing the energy of observed data samples while maximizing the energy of unobserved configurations.
- A symmetric autoregressive lag of N = 5 is used for both visible and hidden connections. model
Source sentence
While the autoregressive lag order can theoretically differ for visible and hidden connections, in this study we set a symmetric lag of N = 5.
- Conditional independence across time is assumed given a finite lag window. model
Source sentence
Several limitations must be emphasized. The model assumes conditional in‑dependence across time given a finite lag window and does not possess an internal latent state capable of generating true dynamical persistence.
Data
- multi-asset dataset market-prices unknown
Source sentence
This study utilizes a multi-asset dataset spanning from January 2013 to January 2025, curated to capture a diverse range of market factors including Equities, Rates, Volatility, and Credit spreads.
Costs
Not applicable
Finds 5
- The Gaussian CRBM preserves static asset correlations but fails to generate long‑range volatility clustering. negative
Source sentence
While the Gaussian CRBM successfully preserves static asset correlations, it exhibits limitations in generating long-range volatility clustering.
- Free energy spikes during market crashes, indicating low probability for out‑of‑sample states and distinguishing high‑volatility regimes. performance
Source sentence
We can clearly observe that the free energy spikes during market crashes (as the model assigns low probabilities to these out-of-sample states).
- The Bernoulli‑Bernoulli CRBM suffers from structural limitations such as information bottlenecks and unstable likelihood estimates driven by bit‑level noise. negative
Source sentence
The Bernoulli CRBM was shown to suffer from structural limitations inherent to binary representations, including information bottlenecks and unstable likelihood estimates driven by bit-level noise.
- The Gaussian CRBM produces a stable likelihood (free‑energy) signal over time, though its quadratic term leads to thin‑tailed synthetic generation. statistical
Source sentence
In contrast, the Gaussian CRBM preserved the empirical correlation structure of the market and produced a stable likelihood signal over time, albeit at the cost of thin-tailed synthetic generation due to its quadratic term.
- Decomposing free energy into magnitude‑dependent and interaction‑dependent components reveals heterogeneous stress episodes not captured by volatility measures alone. theoretical
Source sentence
Decomposing this surprisal into magnitude‑dependent and interaction‑dependent components revealed that periods of market stress are not homogeneous: some episodes are dominated by large return realizations with relatively stable dependence structure, while others coincide with pronounced deviations in the model’s learn
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 39 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (6)
- Modeling human motion using binary latent variables (2006)
- The market generator (2019)
- A practical guide to training restricted Boltzmann machines (2010)
- Improving the robustness of trading strategy backtesting with boltzmann machines and generative adversarial networks (2020) — link
- Temperature based restricted Boltzmann machines (2016)
- A tutorial on energy-based learning (2006)
0 of 6 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.