Markov chains vs. Shapley value: which credit-splitting math should you trust?

Markov chains vs. Shapley value: which credit-splitting math should you trust?

When practitioners move past last-click (crediting the final touch before conversion) to an algorithmic model, two methods dominate. Which fits your data?

The two approaches
Markov-chain attribution models the customer journey as a graph of states and asks a counterfactual question: if we remove channel X entirely, how much does conversion probability drop? That drop — the "removal effect" — is the channel's credit.
Shapley value, borrowed from cooperative game theory, instead averages a channel's marginal contribution across every possible ordering of channels in the journey.

What the difference means in practice
Markov is path-aware: it respects sequence and is cheap to compute on millions of journeys, which is why most open-source libraries (ChannelAttribution in R) default to it. But it assumes the "memoryless" property — the next step depends only on the current state, not the full history. Long, looping journeys violate this.
Shapley is order-agnostic and axiomatically fair (it's the only allocation satisfying efficiency, symmetry, and the null-player rule), but cost explodes combinatorially past 8-10 channels, forcing sampling approximations that reintroduce variance.

The nuance
Neither measures causation. Both redistribute observed correlations more honestly than last-click — that is all. A channel can earn high removal-effect credit simply because it co-occurs with intent, not because it created it.

Bottom line for practitioners: Use Markov when journeys are short, high-volume, and sequence matters. Use Shapley when you have few channels and need a defensible, auditable split for budget politics. Validate either against a holdout experiment before trusting it.