A static Bayesian network models a system at one point in time. A dynamic Bayesian network models how that system evolves.

From static to dynamic

For a static BN, building the model once is building it forever — the conditional probability tables do not change. This is appropriate when the system being modeled is stable: a fixed causal mechanism, a constant operating environment, a population that does not change over time. It fails when any of those assumptions break. Component condition changes as an asset ages. Environmental exposure changes with climate. Counterparty risk changes as leverage ratios move. In each case, the relationships between variables at time t depend on what happened at time t−1 — and a static model cannot represent that dependency.

For a static BN, building the model once is building it forever — the conditional probability tables do not change. This is appropriate when the system being modeled is stable: a fixed causal mechanism, a constant operating environment, a population that does not change over time. It fails when any of those assumptions break. Component condition changes as an asset ages. Environmental exposure changes with climate. Counterparty risk changes as leverage ratios move. In each case, the relationships between variables at time t depend on what happened at time t−1 — and a static model cannot represent that dependency.

A Dynamic Bayesian Network (DBN) extends the static BN by adding a temporal dimension. The model represents the same set of variables at multiple time steps, with two types of edges: intra-slice edges (causal relationships within a single time period — the same as a static BN) and inter-slice edges (how the value of a variable at time t affects the value of a variable at time t+1).

The simplest DBN is a first-order Markov model: the state at time t+1 depends only on the state at time t, not on earlier history. The transition distribution P(X_t+1 | X_t) encodes how each variable evolves given the current state. Hidden Markov Models and Kalman filters are special cases of this structure.

The key advantage over a time series model: the DBN preserves the causal structure at each time step while adding the temporal evolution. Interventions — changing a control variable at time t — propagate through the model to affect the entire future trajectory, not just the next period. This is what a time series model cannot do.

The transition structure

The transition structure requires specifying, for each variable X, the conditional distribution P(X_t+1 | parents(X_t), X_t). This includes both the same-time parents (intra-slice) and the lagged values that affect the next period (inter-slice). The combination of intra-slice and inter-slice edges determines the model’s temporal dynamics.

Example: in an operational risk DBN, Control Effectiveness_t affects Loss Event Frequency_t (intra-slice edge, same as the static model), but Loss Event Frequency_t also affects Control Effectiveness_t+1 (inter-slice edge — a spike in loss events triggers a control review, improving control effectiveness in the next period). This feedback loop cannot be represented in a static BN.

The transition CPTs are more demanding to elicit than intra-slice CPTs because they require expert judgment about dynamics — how quickly does the system respond to shocks? How persistent are regime changes? These questions are often harder to answer than the static causal questions, and the resulting model uncertainty should be reflected in sensitivity analysis over the transition parameters.

Stress testing and scenario simulation

Static BN stress testing sets evidence at one time point and propagates forward: “what is the loss distribution if the macroeconomic regime is recession?” This is useful but incomplete — it produces a snapshot rather than a path. For capital planning, liquidity stress testing, and strategic scenario analysis, the path matters as much as the terminal state.

A DBN stress test: set Macroeconomic Regime_t=0 = recession as evidence. The model propagates forward through the transition structure to produce a distribution over the full trajectory — how does default frequency evolve over the following 12 quarters? What is the expected time to recovery? At what quartile of the distribution does the system not recover within the planning horizon?

These are the questions that Basel stress testing (DFAST, CCAR) and ORSA insurance stress testing require answers to. A DBN provides a principled structural framework for generating those answers rather than applying ad hoc multipliers to historical patterns.

Early warning indicators

An early warning indicator is a variable that is observable today and predicts an adverse outcome in the future. In a DBN, this is a natural inference: for each observable variable X_t and target adverse outcome Y_t+k, compute the mutual information I(X_t; Y_t+k) — how much information does observing X today provide about the adverse outcome k periods from now?

The DBN adds two capabilities that a purely statistical early warning system lacks: (1) the causal interpretation — X is an early warning indicator because it causes (or is caused by a common cause of) Y, not because they happen to be correlated in historical data; (2) the intervention assessment — if X is an early warning indicator and X is controllable (e.g., a specific control metric), then acting on X when it deteriorates has a predictable effect on the probability of Y, which can be quantified through the model.

Early warning systems built on pure correlation are vulnerable to Goodhart’s Law: once a measure is used as a target, it ceases to be a good measure. A causal early warning indicator — one that is causally upstream of the adverse outcome — is more robust to this dynamic because the causal relationship is maintained even when the indicator is acted upon.

Applications and limitations

Well-suited applications: credit portfolio stress testing over regulatory horizons; operational risk trend detection with feedback loops between loss events and control responses; supply chain disruption propagation where a shock at one node reduces the resilience of downstream nodes; climate physical risk evolution as temperature pathways unfold over decades; insurance claim development over multiple reporting periods.

Limitations: the transition structure requires a modeling commitment about how the system evolves that is often harder to justify than the static causal structure; first-order Markov assumptions may be too restrictive for systems with long memory; computational cost grows linearly with the number of time steps and exponentially with the treewidth of the intra-slice graph; and the model requires ongoing validation that the transition dynamics remain stable.

The practical entry point: start with a static BN that is well-validated and used in a context where temporal evolution matters (stress testing, trend monitoring). Add the inter-slice edges one at a time, starting with the most important feedback loops, and validate each addition against historical trajectories. The full DBN is built incrementally rather than specified from scratch.