Data ethics policy
The rules we hold ourselves to, and the checks that catch us when we do not.
Last updated 8 August 2026.
Every figure is computed, not written
No number appears on this site or in a deliverable unless code produced it. That is enforced rather than intended: a script deletes every result file, re-runs all eight published analyses and diffs the output against what is published. If a headline figure moves, the check fails.
A second check, because the first was not enough
Reproduction passed twenty out of twenty while four published sentences were still factually wrong — a count of three where the data said seven, a share attributed to the wrong number of markets, and two more like it. Every one was a number typed inside a formatted string, where nothing was verifying it.
A second check now reads the prose itself and flags any figure that was typed rather than calculated. It found all four. We mention this because a policy that only lists the checks that worked is not worth reading.
Limitations are part of the result
The tool that writes our results refuses to save one that carries no findings or no limitations. Across the published portfolio that is 51 findings and 50 limitations. A piece of analysis that cannot say what it failed to establish is not finished.
We publish the work that went against us
An exception-reporting method we built lost to a naive rule that read a third as many rows. A repricing case has a confidence interval that spans zero, so we cannot tell you the sign. A demand model beats its baseline by less than anyone would like. All three are published, because a portfolio with no losses in it has been curated.
No invented clients, numbers or people
This site once carried an invented benchmark study, invented engagements, invented testimonials and invented colleagues. They were removed — about 41% of the site's written copy — rather than quietly softened. There are no client names here because the one engagement behind this practice is under NDA: no name, no logo and no figures from it appear anywhere, and saying so plainly is cheaper than the alternative.
Public data, used as licensed
The published work uses twelve public datasets. Most are from the UCI Machine Learning Repository under CC BY 4.0; the rest are government and institutional registries — American, Mexican and Canadian — published under their own open licences. Every one of them permits commercial reuse, and every one is cited with its source and its licence on the project page that uses it. We check the licence text rather than the badge, and several otherwise ideal datasets were rejected for this work because they are licensed for academic use only, or because a third party had re-uploaded someone else's data under a licence they had no right to grant.
None of it is personal data about identifiable living people.
Client data stays where it is
Our default is to work inside your environment rather than take copies out. We do not train anything on your data for use elsewhere, and we do not publish your figures. What we take away is method.
Tell us when we are wrong
If a figure here does not reproduce, or a sentence overstates what the data supports, email azlan@predictifyanalytics.com. Corrections are published as visibly as the original claim.