Predictify Analytics is a decision intelligence consultancy. We build the forecasts, models and reporting that survive a board meeting.
Independent by design — no reseller commissions, no technology partnerships that bias what we recommend. The work is published: twenty-five projects with the results, the method and the limitations in the open.
What the raw table costs you · UCI Online Retail II · CC BY 4.0
What is actually wrong with this dataset, what would a data contract have caught, and what does the damage cost in money?
Finance, ops and the board are quoting different revenue figures off the same table, and nobody can say which reading is correct. That is where the work starts — not with a model.
Two spreadsheet tabs, a transaction line per row, no product master and no documented grain. Exactly the shape most operational exports arrive in.
Python and pandas over the raw export. No warehouse to buy first, no platform commitment, and every step reproducible on your own machine.
Thirty-one rules covering grain, keys, types, ranges, referential integrity and sign conventions — each with a severity, written before the data is scored so the thresholds cannot be tuned to flatter the result.
The contract runs across the whole table rather than a sample, and each violation is attributed back to the rule that caught it.
Both spreadsheet tabs carry the same nine days. Anyone who concatenates them inherits the double count, and it survives naive de-duplication because the line values differ.
Four legitimate readings of the same table sit $2.37m apart. The engagement ends with one definition chosen, written down and enforced in CI — so it stays one number after we leave.
The method, the results and what we could not establish — all of it in the open, on public data. You can check every figure on this slideshow against its source.
The analysis does not change with the tool. See the same work rendered in each of them.
Every figure below comes out of code that is published and re-executed on demand. A verification script regenerates each result from the raw data and fails if a single headline number moves.
Results that came out badly are published too: a forecasting model that barely beats a moving average, an exception rule that lost to a naive one, a repricing range that spans zero. All twenty-two projects.
We are independent — no reseller commissions, no vendor partnerships. The platform gets chosen after the decision is understood, not before the kick-off deck is written.
See the project behind each serviceThe same six practices, by sector
Demand, revenue, risk and capacity models that hold up at horizon. Built against a stated accuracy bar, backtested on data your team can re-run.
Operating model, architecture roadmap and the governance to keep it honest. Costed, sequenced and written so it survives a change of sponsor.
Pipelines, warehouse and semantic layer built for change, not for a demo. Tested, documented and owned by a named person on your side.
KPI trees, board-grade reporting and the rituals that make numbers get used. Fewer charts, clearer thresholds, an owner for every metric.
Models in production with monitoring, retraining and a clear rollback path.
Elasticity, promotion effectiveness and margin recovery you can bank. Tested in market, measured against the promotion you would otherwise have run.
The gap is rarely the model. It is that nobody wrote down which decision the model was for, who owns it, or what changes when the number moves.
Before any modelling we write down the decision, its owner, its cadence and its weight in the P&L. If those four boxes cannot be filled in, the model does not get funded.
Every model ships against a threshold agreed before the work starts, measured on hold-out data your analysts can reproduce without us in the room. If it misses, we say so.
Retraining schedule, monitoring, rollback path and a named internal owner before we leave. A model nobody can retire safely is a liability, not an asset.
No twelve-month programme sold on a slide. Each phase produces something you own outright and can take to your board whether or not you hire us for the next one.
A decision audit. Which decisions actually move the P&L, what each one is worth, and what would have to be true to make them better. Two weeks, fixed fee, written findings.
Target-state model and architecture. Written, costed and sequenced, with the options we rejected and why. Stopping here is a legitimate outcome — the design is yours either way.
Production pipelines and models, shipped in fortnightly increments against the accuracy bar agreed in Design. If a model cannot clear that bar, you hear it in the increment where it happens.
Adoption, training and handover, with a measurement plan that survives our departure. We write the value statement with your finance team, then we leave.
These are published analyses on public data rather than client work, which is why every figure can be checked rather than taken on trust. Client engagements are separate, and their figures stay with the client.
Proven distance from the best plan possible. A charger siting plan for two countries, with a proof of how close to optimal it is.
Gross margin a markdown needs to pay for itself. Discount elasticity on 4.5 million store-days, and the profit arithmetic that follows from it.
Of invoice payment dates hit exactly, at best. How precisely the date an invoice will be paid can actually be predicted, on a real purchase-to-pay ledger.
A subscription education technology platform. We built the data platform underneath it and four pieces of work on top.
An AI dashboard over their own data, so the questions the business asks get answered from the platform rather than from an export.
Recommendation models for educational content — what a given learner is served next, and why.
A monthly subscription forecast running three years out, which is a planning horizon rather than a reporting one.
Automated data cleaning and quality checks, so the numbers the other three depend on hold up without anyone remembering to look.
The client is under NDA, so there is no name, no logo and no figures here — and there will not be until they say otherwise. The twenty-two projects on this site are the same method on public data, where you can re-run every number yourself.
See the method on public dataWritten by the people doing the work, mostly about the parts that go wrong. No gated PDFs.
Book a 20-minute scoping call. You leave with a one-page view of the opportunity — no charge, no deck.