Method

What is .Déjà?

.Déjà's business assessment framework is designed with input from investors and business operators, drawing on their combined experience to assess tech companies at different points in history. Comparable historical company cases remain scarce. Our prediction research compares a simple valuation baseline, XGBoost, and TabPFN-v2, a pretrained foundation model for tabular data. TabPFN is a transformer pretrained on synthetic tables; it uses labeled company examples as context without updating its pretrained weights in our experiments. Its approach lets us investigate deep learning with a small company dataset. The website’s alpha-0.1 currently presents TabPFN-v2; the research lab retains the other models as comparators. As AI becomes a new layer of infrastructure, some of the patterns seen in dot-com companies during the internet bubble may reappear in today’s AI-native companies.

What does alpha-0.1 show?

The timeline selects the outcome year: past and present years show recorded valuations, while future years show model estimates. The table reports probabilities for two nested valuation thresholds, so they overlap and do not sum to 100%; calibration is unvalidated, and they are not first-crossing probabilities. The direct model uses the latest company snapshot before the target year with the matching horizon. Public forecasts extend up to three years; local research retains horizons through ten years. Charts show rolling t+1 estimates, with genuine t+3 forecasts extending the final two years where available; recorded valuations stay separate and missing years remain blank. A local TabPFN-v2 experiment explores flat, optimistic, and pessimistic paths for seven factor scores and lagged valuation. These are assumptions, not likelihoods or causal effects. The benchmark is frozen as of October 22, 2026. New-company forecasts use separately dated snapshots; live estimates are unvalidated, and historical estimates are retrospective, not forecasts made at the time.

How do agents participate?

For each company and target date, human analysts work with AI agents to reconstruct its historical situation from dated evidence. Some sources are retrospective, so this reconstruction does not establish that every source was available at the time. The website’s company discovery flow also uses these agents: when a company is not yet in the catalogue, they help identify it from a name or website before evidence gathering and assessment. Analysts define the framework and can review assessments; agents collect evidence, check dates, and produce initial factor scores.

What is the goal?

The goal is to present an evidence-linked profile of a company at a specific point in history and investigate how its recorded history relates to later outcomes. Language models assist with evidence and scoring. Predictive models are evaluated separately using the collected data. Current results concern future valuation tiers; they do not establish a company's ultimate success or investment returns.

Interested in the specific assessment methods or research dataset? Contact the developer.