Technical Summary: Learning Time-Inhomogeneous Markov Dynamics in Financial Time Series via Neural Parameterization

1. Problem Statement

The paper addresses the challenge of modeling non-stationary stochastic systems, specifically financial time series, where classical Markov models and modern deep learning approaches each suffer from significant limitations.

• The Sparsity Barrier: Classical Markov modeling relies on empirical counting of transitions between discrete states. To capture the nuance of heavy-tailed financial returns, a fine discretization (large number of bins, $n$) is required. However, the number of possible transitions scales quadratically ($n^2$). In finite, noisy financial datasets, this leads to severe data sparsity where the vast majority of transition cells are unobserved (e.g., 99.9% of cells have fewer than 5 observations at $n=55$). This causes empirical estimators to collapse into degenerate operators filled with zeros and high-variance noise, rendering them useless for structural analysis.
• The Interpretability Gap: While neural sequence models can handle complex dependencies, they typically operate via opaque latent representations, preventing the application of classical operator-theoretic diagnostics (e.g., transition heatmaps, row heterogeneity, Chapman–Kolmogorov consistency).

The authors propose a middle ground: a framework that utilizes the flexibility of neural networks to parameterize explicit, time-varying Markov transition operators, thereby overcoming the sparsity of empirical counting while maintaining mathematical transparency.

2. Methodology

The core contribution is a neural parameterization of time-inhomogeneous Markov operators.

2.1 Framework Formulation

The system is modeled as a feature-conditioned, time-inhomogeneous Markov chain.

• State Space: Returns are discretized into a finite state space $S = \{s_1, \dots, s_n\}$ via quantile binning.
• Feature Conditioning: Transition probabilities depend on observable covariates $F_t$ (macroeconomic indicators, firm fundamentals) and the current state $X_t$.
• Operator Construction: Instead of counting transitions, a neural network $f^{(h)}_\theta$ learns the mapping from $(X_t, F_t)$ to a probability distribution over the next state.
  $$\hat{A}^{(h)}_t(i, \cdot) = f^{(h)}_\theta(s_i, F_t)$$
  To construct the full transition matrix $\hat{A}^{(h)}_t$ at time $t$, the network is evaluated for every possible state $s_i \in S$ under the same feature snapshot $F_t$. This ensures the output is a valid, smooth stochastic operator even for unobserved transitions.

2.2 Model Architecture and Training

• Architecture: A Multi-Layer Perceptron (MLP) takes the concatenation of a one-hot state encoding and the feature vector. It outputs a probability distribution via a softmax layer.
• Training Objective: Models are trained using Adam optimization to minimize negative log-likelihood. To handle noise in discretized returns, a smoothed-target objective is optionally employed, allocating partial mass to neighboring bins.
• Baselines:
  • State-Free Ablation: A model conditioned only on features $F_t$ (removing $X_t$), producing row-invariant operators.
  • Empirical Counters: Classical estimators (marginal, conditional, backoff-smoothed) to demonstrate the failure of counting at high resolution.

2.3 Diagnostic Tools

Because the output is an explicit operator matrix, the framework enables specific diagnostics:

• Row Heterogeneity: Measures the total-variation distance between rows to quantify state dependence.
• Row Entropy: Measures the dispersion/uncertainty of the transition distribution.
• Chapman–Kolmogorov (CK) Consistency: Compares the directly estimated $h$-step operator against the composition of $h$ one-step operators. This serves as a localized diagnostic to identify temporal windows where the first-order Markov assumption breaks down.

3. Key Results

The experiments focus on a single equity (JPM) with $n=55$ bins and various forecast horizons.

3.1 Failure of Empirical Counting

At the target resolution ($n=55$), empirical transition matrices are effectively empty. Over 57% of cells are entirely unobserved, and 99.9% contain fewer than 5 observations. This confirms that tabulation cannot support operator-level analysis at this granularity.

3.2 Structural Validity and State Dependence

• State Conditioning: The state-conditioned model produces operators with non-trivial structure, achieving a mean row heterogeneity of $\bar{\rho} = 0.0073$. In contrast, the state-free ablation collapses to exactly $\rho = 0.0000$ (identical rows), confirming that the discretized state $X_t$ carries unique transition information.
• Volatility and Homogenization: A counter-intuitive finding is that high-volatility regimes homogenize transition dynamics rather than diversify them.
  • Row entropy correlates strongly and negatively with realized variance ($r = -0.62, p \approx 10^{-251}$).
  • During stressed periods (top 20% of realized variance), transition rows collapse toward a similar, concentrated distribution. During tranquil periods, rows are more diverse.

3.3 Chapman–Kolmogorov Diagnostics

The CK diagnostic reveals that the first-order Markov assumption is not uniformly violated but fails in specific temporal windows.

• The state-conditioned model exhibits higher CK discrepancy (mean KL divergence 0.140 vs. 0.022 for the baseline) because it learns genuine state-dependent dynamics that a simple product of one-step operators cannot perfectly reconstruct.
• CK violations concentrate in specific windows (e.g., $t \in [2120, 2200]$), correlating with periods of concentrated, low-entropy dynamics. This allows the model to pinpoint where "memory" assumptions break down.

3.4 Predictive Performance

The paper reports modest predictive gains, consistent with the weak-form efficiency of single-asset daily data and the limitations of low-frequency features.

• At horizon $h=10$, the state-conditioned model achieves a $\Delta$NLL of $+0.025$ relative to a marginal baseline (95% CI: $[-0.020, +0.028]$).
• Shorter horizons show performance near or below the marginal baseline.
• Calibration is reasonable, with the state-conditioned model achieving an Expected Calibration Error (ECE) of 0.079 for negative returns.

4. Significance and Claims

The paper claims its primary significance lies not in achieving state-of-the-art predictive accuracy, but in restoring structural interpretability to complex time series modeling.

• Bridging the Gap: It demonstrates that neural networks can be constrained to output mathematically coherent structural forms (explicit stochastic operators), making rigorous classical operator analysis viable for high-resolution, non-stationary data.
• Diagnostic Capability: By exposing the transition operator, the framework enables the use of classical tools (heatmaps, heterogeneity metrics, CK checks) that are impossible with black-box sequence models.
• Insight into Market Dynamics: The framework reveals specific structural behaviors, such as the homogenization of transition dynamics during high volatility and the localized breakdown of Markov assumptions, providing a "transparent, calibrated microscope" for studying market regimes.

The authors remain modest regarding the immediate predictive utility, noting that the current single-asset, daily-granularity setting limits signal-to-noise. They suggest that scaling to multi-asset, high-frequency universes is the necessary path to realizing material predictive improvements while retaining these interpretability benefits.