Hazard Function

The hazard is not a probability but a rate with a range from 0 to infinity. It represents the event's propensity to occur in an infinitesimally small interval [t, t + Δt), given survival to t.

• Formula: h(t) = lim (Δt→0) P(t ≤ T < t + Δt | T ≥ t) / Δt
• Key Distinction: A hazard of 2.0 means the event is twice as likely to occur in the next instant compared to a baseline, not a 200% probability.
• Example: In churn analysis, a spiking hazard at day 30 might indicate a cohort reaching the end of a free trial.

The hazard function is fundamentally conditional. It only considers individuals still in the risk set at time t, automatically excluding those who have already experienced the event or been censored.

• Dynamic Risk Set: As customers churn, the denominator shrinks, isolating the true risk among active survivors.
• Survivor Bias Correction: This conditioning prevents the dilution of risk estimates by individuals no longer at risk.
• Practical Use: For a subscription service, the hazard at month 12 analyzes only users who maintained their subscription through month 11.

The hazard can adopt any functional shape over time, revealing critical behavioral patterns in the underlying process.

• Increasing Hazard: Risk grows over time (e.g., mechanical wear-out failures, long-term customer fatigue).
• Decreasing Hazard: Risk is highest early and declines (e.g., infant mortality in hardware, early-stage user churn after a bad onboarding experience).
• Bathtub Curve: A combination of decreasing, constant, and increasing phases, common in lifecycle analysis.
• Constant Hazard: The memoryless exponential distribution where risk is uniform, implying the future is independent of the past.

The hazard function and the survival function S(t) are mathematically linked. The survival probability is a direct function of the cumulative hazard H(t).

• Cumulative Hazard: H(t) = ∫₀ᵗ h(u) du, which aggregates all risk experienced up to time t.
• Conversion: S(t) = exp(-H(t)). A high cumulative hazard forces the survival probability toward zero.
• Interpretation: If the cumulative hazard reaches 1.0 by month 6, the probability of surviving past month 6 is approximately 36.8% (e⁻¹).

In the Cox Proportional Hazards model, the hazard function is decomposed into a non-parametric baseline hazard h₀(t) and a parametric covariate effect.

• Formula: h(t|X) = h₀(t) × exp(β₁X₁ + β₂X₂ + ...)
• Proportional Hazards Assumption: The effect of a covariate is multiplicative and constant over time. A feature like 'premium support tier' might multiply the baseline hazard by 0.5 at all time points.
• Baseline Flexibility: h₀(t) is left completely unspecified, allowing the model to capture complex temporal patterns without parametric constraints.

When covariates are not considered, the hazard can be estimated directly from data using life-table or kernel-smoothing methods.

• Nelson-Aalen Estimator: Provides a direct non-parametric estimate of the cumulative hazard H(t) by summing the ratio of events to the risk set size at each distinct event time.
• Discrete Hazard: In discrete-time settings, the hazard is the conditional probability of failure in an interval, estimated via life tables.
• Smoothing: Kernel smoothers can be applied to raw hazard increments to visualize the underlying continuous trajectory and identify inflection points.

The hazard function identifies the precise moments when a customer's instantaneous risk of churn spikes. By monitoring the hazard rate in real-time, systems can trigger retention offers exactly when the probability of defection is highest, rather than relying on static, calendar-based campaigns.

• Example: A streaming service detects a user's hazard rate triples after a billing failure; an immediate retry with a grace period is triggered.
• Key Metric: Time-varying hazard ratio.

Using the Cox Proportional Hazards Model, analysts can quantify how specific customer attributes (covariates) multiply the baseline hazard. This allows CLV models to adjust future revenue projections based on risk factors like support ticket frequency or login recency.

• Example: A SaaS company finds that a customer who submits 3+ support tickets in a week has a hazard ratio of 2.5, effectively accelerating their expected churn timeline.
• Key Metric: Hazard Ratio (exp(β)).

In retail settings where customers don't formally cancel, the hazard function models the latent dropout process. By estimating the probability that a customer becomes permanently inactive after a specific transaction, it provides a probabilistic end-of-life marker for CLV calculation.

• Example: A fashion retailer models a sharp increase in the hazard rate after 90 days of inactivity, defining the point of 'silent churn' for CLV write-off.
• Key Metric: Cumulative Hazard.

The shape of the hazard function dictates the optimal frequency of marketing touches. A bathtub-shaped hazard suggests high initial risk, a stable middle period, and late-life fatigue, guiding a strategy of heavy onboarding, maintenance, and win-back campaigns.

• Example: An insurer uses a Weibull hazard model to space policy renewal reminders, increasing frequency as the policy end-date approaches and the hazard rate climbs.
• Key Metric: Shape Parameter (k).

Clustering customers by their estimated individual hazard trajectories reveals distinct risk archetypes. This moves beyond static RFM segments to dynamic groups like 'rapid burners,' 'stable loyalists,' and 'late defectors,' each warranting a different CLV strategy.

• Example: A telco identifies a 'price-sensitive' segment whose hazard rate spikes predictably at the 12-month contract end, allowing for preemptive loyalty offers.
• Key Metric: Individual Hazard Function.

The area under the survival curve, which is the integral of the hazard function, directly yields the expected customer lifetime. This provides a more granular and statistically robust duration estimate for the CLV formula than a simple historical average.

• Example: A B2B platform calculates that a customer with a decreasing hazard function has an expected lifetime of 4.2 years, compared to 1.1 years for one with an increasing hazard.
• Key Metric: Mean Survival Time.