Bayesian Updating in Financial Probability Analysis

Most bad forecasting is not bad arithmetic. It is starting from the wrong base rate, or over-reacting to evidence that was not very diagnostic.

The structure of the idea

Bayesian updating is a disciplined procedure for revising a belief when new information arrives. It has three ingredients: a prior (what you believed beforehand), the evidence, and the likelihood — how much more probable that evidence is if your hypothesis is true than if it is false.

The output is a posterior belief. Crucially, the update is proportional: weak evidence should move a belief slightly, strong evidence a lot. Most forecasting errors come from ignoring that proportionality.

The base rate is the part people skip

The prior usually should be the base rate — how often this kind of thing happens in general. Recessions in any given year, companies missing guidance, ceasefires holding: each has a historical frequency, and that frequency is where a well-formed estimate starts.

Vivid, specific, emotionally engaging evidence tends to crowd out the base rate. This is base rate neglect, and it is why dramatic outcomes get systematically overpriced: a compelling story about how something could happen displaces the far duller question of how often such things actually do.

How diagnostic is the evidence, really?

The strength of an update depends on how much the evidence distinguishes between worlds. Evidence that is roughly as likely whether or not your hypothesis is true is not diagnostic, and should barely move your belief.

Much financial news falls into this category. A commentator expressing concern about a recession is nearly as likely in a year without one as in a year with one, so it carries almost no information. A yield curve inverting, or credit spreads widening sharply, is meaningfully more common ahead of downturns — genuinely diagnostic, and worth a real update.

Prediction markets as continuous Bayesian machines

A prediction market is close to a live implementation of this process. The current price is the collective prior; news arrives; participants who consider it diagnostic trade; the price settles at a new posterior.

This is why the size of a move matters. A small shift on major news suggests the market considered it already priced or not very diagnostic. A large shift on apparently minor news suggests it was far more informative than it looked — often the more interesting signal of the two.

Applying it without the arithmetic

The full formula is rarely necessary. The discipline is: start from a base rate rather than a headline, ask honestly how much more likely this evidence would be if your hypothesis were true, and update proportionally.

The most common failure is not miscalculation. It is starting from a vivid story instead of a frequency, and updating far too aggressively on evidence that barely distinguishes between outcomes.