Investigating Conditional Restricted Boltzmann Machines in Regime Detection

Battle Count: 62

The paper provides a novel, interpretable diagnostic tool for systemic risk monitoring via free energy decomposition. The distinction between magnitude shocks (quadratic term) and correlation regime shifts (structural term) offers actionable insight beyond VIX. However, the model is not directly a trading signal generator, lacks numerical benchmarks against standard models, and the generative component (synthetic data) is limited by thin tails. The regime detection capability is most relevant for risk overlay, portfolio rebalancing triggers, and as a supplemental indicator to existing volatility measures.

CRBM implementation requires understanding of energy-based models, PCD training, autoregressive conditioning with dynamic biases (matrices A and B), and free energy computation. The Gaussian-Bernoulli variant is more straightforward than the Bernoulli-Bernoulli (which requires 16-bit encoding). Training stability with PCD, proper initialization, and handling the quadratic energy landscape add complexity. The 19-dimensional asset space with N=5 lag creates ND=95 autoregressive inputs. No reference implementation is provided.

The paper provides detailed mathematical formulations and specifies hyperparameters (N=5 lag, 16-bit encoding, T=1, training 2013-2019, testing 2020-2025). However, no code repository is provided, no numerical results are reported in tabular form (all in figures), and the specific PCD training details (number of steps, learning rates, batch sizes, number of hidden units) are not fully specified. The asset universe is listed but data sources (Bloomberg terminal) are proprietary.