We publish findings in dollars. You can reproduce every one.
Every case study below runs on data you can get yourself, and we print the commands in full. No anonymised client logos, no percentages without a denominator. Where a model is weak we say so and tell you what to do instead. Treat that as the product, not as modesty.
Four categories, four different answers.
We picked them to span the range: a thin catalog, a market you already understand, a dataset that caps its own precision, and the seller's side of the table.
Battery management ICs
A thin catalog, the size of a real category. Same spec, different badge, 3.9x the price. Price a part yourself in a calculator that runs the trained model.
36% addressable premium · full write-up here
Read the case study → 78,000 listings · CraigslistUsed vehicles
Prices span orders of magnitude in a market you already have intuitions about. Check whether the model reasons about price the way you do before you trust it where you cannot check.
R² 0.78 · median error 14%
Read it on GitHub → 36,668 bolts · DLA PUB LOGAerospace fasteners
Here the data itself caps how precise any number can be, and we publish the model that proves it. Knowing which estimates to open with is worth more than a tighter average.
R² 0.32 · median error 80%
Read it on GitHub → 401,125 records · Blue BookHeavy equipment resale
The only study that flips the chair. Same engine, seller's side: what a used machine will fetch at auction, which specs carry the value, and where money is being left on the table.
R² 0.74 · seller side
Read it on GitHub →We hold ourselves to three rules.
Public data only
We only use datasets you can get yourself. Where licensing forbids redistribution, we ship the fetch script and a scrubbed sample instead of the full pull.
We publish the weaknesses
Each study says where the model is unreliable, by price band and by feature, in the same voice as the findings. You cannot trust a benchmark you cannot audit.
We lead with the decision
We headline a decision in dollars every time. We report the model statistics underneath it, and we never let them stand in for a finding.
Your own data beats all of this.
We run these on public catalogs so you can reproduce them. Your own paid-price history is richer than anything a catalog will give you, and it never leaves your machine.
pip install p2predict[mcp]