Markov Chain Funnel Modelling

Description

Models user journeys as states with transition probabilities, including absorbing states like conversion and exit, to compute completion probabilities and step importance.

Linear funnels assume users move forward step by step, but real journeys loop, skip and backtrack. A Markov chain estimates the probability of moving from each state (page, screen, step) to every other. With absorbing states for success and abandonment, the fundamental matrix N = (I - Q)^-1 gives expected visits to each state and absorption probabilities from any starting point. Removal effects, setting a state's transitions to exit, quantify each state's contribution to conversion. Higher-order chains capture memory. This gives a principled, whole-journey model for diagnosing where loops waste effort and which steps matter most.

Attributes

Loss type
Latent
Stage
Post-launch
Effort
Days

Steps

  1. Define states at an appropriate granularity and absorbing states (conversion, exit, timeout).
  2. Estimate the transition matrix from session logs, checking sample size per row.
  3. Compute the fundamental matrix, expected steps to absorption and absorption probabilities.
  4. Calculate removal effects for each state to gauge its contribution to success.
  5. Test whether a first-order model is adequate versus higher-order (e.g. likelihood ratio or BIC).
  6. Simulate design changes by modifying transition probabilities and predict their effect on conversion.

Sources

  • Kemeny & Snell, Finite Markov Chains (1960)
  • Anderl, Becker, von Wangenheim & Schumann, Mapping the customer journey: Lessons learned from graph-based online attribution modeling (2016), International Journal of Research in Marketing

Also known as

  • Absorbing Markov chains
  • Markov attribution
  • Transition matrix journey modelling

Categories

  • Information Theory, Physics & Mathematics