Sequential Evidence Accumulation¶

Why It Matters¶

In practice, evidence arrives over time, not in a single batch. Waiting for "all the evidence" means making decisions with outdated beliefs. Sequential updating provides a principled way to maintain calibrated beliefs as information trickles in. It also reveals when you've accumulated enough evidence - if the next several pieces barely move your posterior, you're approaching information saturation. The "order doesn't matter" theorem is psychologically liberating: you don't need to process evidence in a specific sequence.

The Key Move¶

Start with today's best probability estimate (prior). When new evidence arrives, update to a posterior. Immediately relabel this posterior as your new prior. When the next evidence arrives, update from this new prior. Repeat. Track your probability trajectory over time. If you notice diminishing updates (each new piece moves you less), you're approaching the limits of what evidence can tell you given your model.

Classic Application¶

Forecasting tournament play (e.g., Good Judgment Project). Question: "Will X happen by date Y?" Start with base rate (prior). As news arrives daily, update sequentially. Each update is small but disciplined. After 50 updates, your probability has traveled from initial 30% to current 75%, reflecting accumulated evidence. Sequential updating prevents: ignoring early evidence, overreacting to recent evidence, forgetting what moved you from your initial position.

Surprising Application¶

Relationship trust dynamics. Trust is a probability about how someone will behave. Each interaction provides evidence. Sequential updating: start with modest prior trust, update after each interaction. Consistent positive behavior slowly accumulates into high posterior trust (which becomes prior for next interaction). One violation provides strong negative evidence, rapidly dropping posterior trust. The sequential framework explains why: trust builds slowly (many small positive updates) but collapses quickly (one large negative update), and why consistent behavior matters more than occasional grand gestures.

Failure Modes¶

Confirmation bias in evidence selection - if you only notice evidence confirming current beliefs, sequential updating entrenches errors. Order independence assumes conditional independence - if evidence pieces interact, order can matter. Forgetting old evidence - humans can't actually remember all previous updates; approximation is necessary. Update exhaustion - updating on every tiny piece of information is cognitively impossible; you need thresholds for "evidence worth updating on."

Go Deeper¶

Lindley, Understanding Uncertainty, Chapter 8 on sequential analysis; Silver, The Signal and the Noise, Chapter 7 on sequential forecasting.