Plain, honest documentation of our methodology — including what the model can't do. We publish the machinery so you can calibrate how much to trust it.
Is this a genuine simulation, or an AI telling plausible stories? This paper shows the actual machinery — how outcomes are computed, not narrated, how 2.5M personas are built from real data, and how we measure uncertainty — and is equally clear about what the model can’t do.
The “math decides, the AI only voices it” principle, with a worked churn example
What a persona really is — and the honest truth about “2.5 million”
Confidence ranges vs. predictive accuracy — the distinction we insist on
Our held-out validation protocol — and why we won’t quote an accuracy number yet
The validation protocol — published before we quote any accuracy number, because the protocol is what makes a number trustworthy. We lock a prediction before launch, score it on outcomes the model never saw, and commit to publishing the misses, not just the wins.
The held-out protocol: lock the prediction → wait → score on unseen launches
Sentiment, conversion, per-segment, per-country error, and calibration — how each is scored
Guardrails against fooling ourselves: no peeking, no tuning on the test set, failures in-scope
Cold vs. calibrated accuracy — and an honest status on what we can publish today
A deep-dive on sub-country resolution. Most tools model “the UAE.” We model the difference between a luxury buyer in Palm Jumeirah and a value-driven shopper in Deira — and this paper shows how much that resolution changes the answer.
How zones are defined: ~50 priority countries broken into hundreds of local areas
What changes between zones — income, culture, brand affinity, price elasticity
A worked (illustrative) comparison: the same offer, two neighborhoods, two outcomes
When neighborhood resolution matters most (and when country-level is enough)