How does a multi-feature Bayesian prescreen work, and how is it different from authentication?

Short answer: It turns each observation into a likelihood ratio, adds their logarithms to the prior on the log-odds scale, and converts the total back into an appearance-consistency probability. It uses appearance only and illustrative parameters, so it is not authentication and not a conclusion on authenticity. Its job is to order your next step: keep talking, ask for photos, or send it for authentication.

The site tool gives a visual-consistency probability for reference only. It is not an official authentication certificate and is not a verdict on authenticity; for authentication, use the brand or a professional authentication service.

The calculation, step by step

Step one, the prior: choose a starting point based on the sales channel, for example 0.60 for a general resale marketplace, which is an illustrative value. Step two, observations: pick one category for each of the six features, choosing unobservable when a detail is unclear. Step three, likelihood ratios: each category maps to a number; above 1 pushes the estimate up, below 1 pulls it down.

Step four, combine: convert the prior to log-odds, add the logarithm of each likelihood ratio, and convert back to a probability. To limit overconfidence, the result is kept between 3% and 97%, and a range is reported using the lower and upper bounds of each parameter.

An illustrative example

A prior of 0.60 corresponds to odds of 1.5. Suppose the stitching matches the reference, with a likelihood ratio of 2.0, the hardware match carries 2.5, and the other four features are unobservable. The odds become 1.5 × 2.0 × 2.5 = 7.5, or a probability of about 88%. Every number here is illustrative.

Now add a clearly low price, with a likelihood ratio of 0.3. The odds drop to 2.25 and the probability to about 69%. One strong piece of evidence can move the result a long way.

The limits of the naive assumption

Naive means the model assumes each feature is independent of the others. In reality they are not: the same workshop may get both the stitching and the hardware right. When several features agree, the result therefore leans optimistic. The statistical literature notes that naive Bayes often ranks cases well even when its probability values are too extreme.

That is why we keep the result within 3% to 97% and report a range alongside it. The range reflects only the parameter bounds; it is not a statistical confidence interval.

How this differs from authentication

Authentication means a professional examines the physical item, drawing on feel, construction and materials and sometimes instruments, and the issuer stands behind the conclusion. A prescreen uses only your observations of appearance and illustrative parameters to produce a reference probability. They serve different purposes: the prescreen decides the next step, authentication gives the conclusion.

Follow-up questions

Why does the result never reach 100% even when all six features match?

The ceiling is set to 97% on purpose, because the naive assumption makes agreeing features look more convincing than they are.

Where do the likelihood ratios come from?

They are currently illustrative and need calibration against inspected items with known outcomes. Until then, read the output as a ranking aid.

Can I fill in only two or three features?

Yes. Blank features count as unobservable and do not move the result; more observations make it more informative.

The prescreen we built follows exactly these steps and returns a probability, a range and the single photo most worth requesting. It is a reference probability based on appearance, not an official authentication certificate and not a conclusion on authenticity; for authentication, please use the brand or a professional authentication service.

About this tool

Sources

  1. Domingos, P. & Pazzani, M., 1997, On the Optimality of the Simple Bayesian Classifier under Zero-One Loss, Machine Learning 29(2-3)
  2. Hand, D. J. & Yu, K., 2001, Idiot's Bayes: Not So Stupid After All? International Statistical Review 69(3)
  3. Gigerenzer, G. & Hoffrage, U., 1995, How to Improve Bayesian Reasoning Without Instruction: Frequency Formats, Psychological Review 102(4)