Hidden Markov Model (HMM)

The foundational assumption that the probability of transitioning to the next hidden state depends only on the current state, not the full history. This first-order Markov property drastically reduces computational complexity.

• Transition Matrix (A): An NxN matrix where a_ij = P(q_{t+1} = S_j | q_t = S_i)
• Practical Implication: A customer's next latent state (e.g., 'High Engagement') depends solely on their current state ('At-Risk'), not how they got there.
• Higher-Order Extensions: While standard HMMs use first-order, second-order models can capture longer dependencies at increased parameter cost.

Each hidden state generates observable actions according to a probability distribution. In retail, these emissions are the purchase events, clicks, or browsing patterns we actually see.

• Emission Matrix (B): Defines b_j(k) = P(o_t = v_k | q_t = S_j), the probability of observing symbol v_k while in state S_j.
• Discrete vs. Continuous: Discrete emissions model categorical actions (product category viewed); continuous Gaussian emissions model monetary values (order amounts).
• Example: A 'Churn-Ready' state might emit 'No Purchase' with 95% probability and 'Low-Value Purchase' with 5%.

The probability vector π_i = P(q_1 = S_i) specifying the likelihood of a customer starting in each hidden state at time t=1. This anchors the entire temporal sequence.

• Cold-Start Relevance: For new customers with no history, π provides the prior probability of their initial latent state based on acquisition channel or demographics.
• Stationary Distribution: For ergodic HMMs, the long-run state probabilities can be derived from the transition matrix, representing the steady-state composition of the customer base.
• Constraint: Must sum to 1.0 across all N states.

Standard HMMs implicitly assume a geometric distribution for state duration, which may not reflect real customer lifecycles where 'Loyalty' states persist for predictable periods.

• Limitation: The probability of remaining in state S_i for d steps decays exponentially as a_{ii}^{d-1}(1-a_{ii}).
• Explicit Duration HMMs: An advanced variant that models state sojourn time with an explicit duration distribution (e.g., Gamma or Poisson), capturing more realistic customer retention curves.
• Application: Accurately modeling the expected duration of a 'High-Value' state directly informs CLV projections.

To capture the multi-faceted nature of customer behavior, HMMs can be composed into more expressive architectures.

• Hierarchical HMMs (HHMM): States themselves contain sub-HMMs, modeling behavior at multiple temporal scales—e.g., a 'Shopping Session' state containing sub-states for 'Browsing', 'Cart Addition', and 'Checkout'.
• Factorial HMMs: Represent the hidden state as a collection of independent state variables evolving in parallel. A customer might have a latent 'Price Sensitivity' state and a separate 'Category Interest' state.
• Switching State-Space Models: Combine a discrete HMM with continuous linear dynamics within each state, useful for modeling evolving spend trajectories.

A statistical approach that estimates individual customer parameters by borrowing strength from the population distribution. In HMM contexts, this enables robust estimation of transition and emission probabilities even with sparse individual data. Key benefits include:

• Partial pooling prevents overfitting to noisy individual sequences
• Naturally handles heterogeneity across customer segments
• Provides full posterior distributions rather than point estimates

The broader discipline of capturing temporal patterns in user actions to predict next-click intent and session outcomes. HMMs serve as a foundational generative model within this field, explicitly modeling the latent psychological state (browsing, comparing, ready-to-buy) that drives observable clickstream data. Modern approaches often combine HMMs with deep learning for richer representations.

A model-based clustering technique used to identify unobservable subgroups within a customer base that share similar longitudinal purchasing patterns. While HMMs model dynamic transitions between latent states over time, Latent Class Analysis identifies static, time-invariant segments. The two are often used together: LCA defines the initial state distribution for an HMM.

A statistical framework for analyzing the expected duration of time until a specific event occurs, such as customer churn. HMMs complement survival analysis by modeling the hidden state trajectory that precedes the terminal event. A customer in a 'disengaged' latent state exhibits a sharply elevated hazard rate, making HMMs powerful early-warning systems for churn prediction.

A predictive score representing the probability that a customer will make a subsequent purchase. HMMs enhance this binary classification task by conditioning the repurchase probability on the current inferred latent state. A customer in a 'loyal' state has a fundamentally different propensity than one in a 'trial' state, even if their recent observable behavior appears similar.