Synthetic Data for Business Intelligence Starts With a Data Model
A food brand signs with a national retailer. The analyst who will own the weekly sales report logs in to the retailer's supplier portal and finds nothing to report on. The first scan file won't exist until the product is on a shelf and someone buys it. At a manufacturer, the channel finance manager has the opposite problem. Her claim file and payout ledger hold years of rebate history, an auditor is ready to review them, and neither file can leave the company, because together they show every distributor's volume and tier.
Both are waiting on data, and both assume the reporting has to wait too. It doesn't have to, and that is the case for synthetic data for business intelligence. Synthetic data does not replace real data for decisions. It replaces the wait for real data, and the silence of data under NDA. You build the apparatus on synthetic data: the model, the reports, the rules. You run the decision on real data.
The examples in this post come from three companies that do not exist. The first is Cinderhaven Provisions, a fictional specialty food company whose figures are a synthetic dataset I built so that work like this can be shown in full. Its launch example shows what the analyst's company would want to know early: a projected $499,200 in first-year revenue requires $156,352 in working capital in Month 1, before the first payment arrives. A report built after the scan data shows up is too late to show that.
The practices behind this are already established
Building on a data model and testing on synthetic data are not new ideas. Four established practices support them, each from a field that has relied on it for years.
Settle definitions once, in a model. A data model is a written description of the entities a business keeps records about, the fields each entity owns, and the keys that join them. Data warehouse practice has long asked for definitions to be agreed there. The Kimball Group's guidance on shared dimensions asks for a single definition, agreed with the people who govern the business's data, and credits that agreement with consistent analysis and lower development cost later.
Rehearse on synthetic data before the real data exists. Britain's Office for National Statistics has to have its processing ready before a census is taken. A synthetic census dataset was among its Data Science Campus's contributions to the 2019 census rehearsal, and pipeline testing during the wait for real data is a use the Campus names for the technique. Each census is a single event with no do-overs, so the processing has to work before the real forms arrive.
Build outside the confidential data, then run inside it. The US Census Bureau holds survey records linked to administrative data that it cannot release. Its answer was a synthetic file: modeled values placed in the structure of the real records. Researchers write and test their analysis on that file. Their code then goes to the Bureau, which executes it against the confidential records and releases only the output it has cleared. The Bureau does not treat the synthetic result as final. Its paper presents the run on real records as the answer to remaining doubts about accuracy.
Test a method against an answer you already know. Statisticians evaluate their own methods with simulation studies. A 2019 tutorial in Statistics in Medicine names the central advantage: the analyst generated the data and therefore holds the true answer, so a method's bias can be measured. One volume of that journal alone supplied its authors with 100 papers containing a simulation study. Software engineers apply the same principle to defects. They plant faults in a program deliberately and grade a test suite by how many it catches, a technique called mutation testing.
A reporting layer has the same problems these practices solve
None of those fields calls its work business intelligence. But a reporting layer faces the same problems, and the same practices apply to it.
By business intelligence I mean the reports, the rules and the few numbers a leadership team decides from. At its simplest it is three numbers on a Monday morning. Whatever its size, it has three parts, and each can be wrong without producing an error message.
The definitions. A report can only say what "fill rate" or "penetration" was once someone has decided what the word means. Penetration alone can mean three different metrics on the same dashboard.
The pipeline. Raw rows are joined, filtered and totaled before they become a report. A join on the wrong field still produces a number.
The rules and estimates. A threshold turns a number into a verdict, and some figures on a report, such as the lift from a promotion, are estimates and not counts. Most operating questions can be answered by a rule written down once, but a written rule can still be a wrong rule.
Real data is a poor place to test any of the three. The definitions get settled under deadline. The pipeline's first run is on numbers people are already deciding from. And there is no known right answer to check an estimate against.
Each practice above covers one of these. The data model settles the definitions. A rehearsal on synthetic data tests the pipeline. A known answer tests the estimates and rules. And the Census Bureau's arrangement covers the case where the data cannot leave the company. After a look at the alternatives, the sections below take them in that order.
Waiting for the real data has a cost
A report that is wrong still gets used, and its cost shows up in the decisions made from it. So the question is where its mistakes get found. There are four options, and three of them have a specific weakness.
| Option | What it offers | What it can't do |
|---|---|---|
| Wait, then build | Real numbers from the start | The report arrives after the decisions it was meant to inform, and it is built under deadline |
| Build and test on live data | Real numbers, tested in use | Mistakes are found by people acting on them, and an estimate can't be checked because the true answer is unknown |
| Anonymize a real file | Real structure with names removed | It needs a real file and its owner's consent, and it still holds an unknown number of errors |
| Synthetic data on a data model | Available before the data exists, with every error and true value known | It can't tell you your own numbers |
Synthetic data on a data model is the only one of the four that works before the data exists, checks a method against a known answer, and involves no confidential records. It does not replace testing on real data. It moves most of the mistakes to a stage where finding them costs nothing, and leaves the real data for the one check only it can do, the reconciliation.
Definitions: the data model settles what the numbers mean
Writing a data model forces the definitions into the open. The business has to state whether a case is a product in its own right or a quantity of another product, which record holds the official case weight, and what must match before two rows count as the same shipment. For a reporting layer, the "consistent analysis" Kimball describes means that two reports using the same word produce the same number.
The Product Master Data Model shows the decisions for one Cinderhaven product, a jar of Calabrian chili marinara, followed from brand to pallet. Six jars make an inner pack. Twenty-four make a case, which carries its own barcode number, separate from the jar's. Forty-eight cases make a pallet, which is tracked as a shipment and has no product number. One retailer sets the item up by the jar. A club retailer and a distributor set it up by the case. A spreadsheet with one row per product cannot hold both answers.
Two other posts go deeper on this. One explains how the product master model works and what it prevents. The other, the Product Data Health Audit, puts a dollar figure on the fields the model would have assigned an owner.
When a definition is left unsettled, the reporting layer shows it as two reports that disagree while both are correct. The Cinderhaven dataset holds a pair like that. By the brand's internal records, 99.2% of ordered units shipped. Counted the way a retailer counts, as shipments that arrive on time and complete, the figure runs as low as 84.5%. Both numbers are right. They answer different questions, and no document said which one "fill rate" meant. The difference is not academic. A team reading 99.2% believes fulfillment is fine, while the gap between the two figures carries about $57,000 a year in chargebacks and modeled lost sales in the Cinderhaven dataset. This is also the first of the four data hygiene tests: whether fields mean what they say.
The model is also what makes synthetic data worth generating. A generator that fills a model writes rows that obey its keys: every shipment points at a real case, every case at a real product, every deduction at a real invoice. Random rows in a spreadsheet have nothing to join to, so a report built on them shows nothing about whether its joins or totals are right.
Reports: rehearse the pipeline before the data arrives
A reporting layer can be rehearsed the way the ONS rehearsed a census. The Cinderhaven warehouse holds 36 months of synthetic orders, shipments, deductions and payments, and 313 automated tests run on it. The tests confirm that keys are unique, required fields are filled, and every transaction points at a product and a customer that exist. When a test fails there, a broken report has been found before anyone made a decision from it.
A rehearsed report can also answer questions early. Cost of Saying Yes is the launch example from the opening, modeled month by month. The Recall Blast Radius does the same for a product recall, which a company has to plan for without having been through one. In one scenario a single ingredient lot goes into one production batch. In another, a shared lot is used across several production runs, and the recall grows to 23 lots and 5,785 cases in the channel. The difference is whether the lot records link to each other, and the data model decides that.
Estimates and rules: check them against a known answer
Many figures on a report are estimates. The lift from a promotion, the return on a campaign and the amount overpaid on a rebate program are all calculated by a method, and on real data there is no way to see how far the method is from the truth. A simulation study tests an estimate. Planted faults test whether an exception report or an audit catches what it should. Three examples on this site do one or the other.
Lift Math tests an estimate. The lift number on a promotion report tells a team whether a promotion paid for itself. In Lift Math every Cinderhaven promotion has a true lift that was generated, then kept hidden from the two methods that estimate it. Scored against that truth, the simple method's typical error is 34%, and the more rigorous method's is 41%. A promotion report built on real data cannot be scored this way, because real promotion data has no true lift to compare against. Both methods also err in the same direction. They credit promotions with more lift than the promotions produced. A team relying on either report would see its promotions as more profitable than they were and keep funding some that lost money. Even on the more forgiving method, 2,137 of 5,484 promotions lost money.
Campaign Math tests a ranking. It runs on Quellwood Provisions, an invented direct-to-consumer food brand with one synthetic quarter of web analytics and nine kinds of tracking defect planted at stated rates. Nine of its 12 paid campaigns change rank by two or more places depending on which attribution rule the report uses, and $5,064 of $72,621 in paid spend cannot be traced to a campaign. A marketing dashboard shows one ranking. The synthetic quarter shows how much that ranking depends on a rule and on data defects. One campaign returns 0.69 times its spend under one rule and 1.88 times under another, with the same spend and the same sales. A team that sees only one of those figures would either cut the campaign or give it more budget, and nothing on the dashboard would indicate that the other reading exists.
Payout Math tests an audit. It runs on Galewick Climate Systems, an invented HVAC manufacturer with one synthetic program year of bonus claims and rebate statements, including 129 planted payout errors. The validator sees only what a program's own systems would hold, plus the published rules. It finds all 129 and flags nothing else: $17,597 of $521,804 in approved payouts that should not have been approved.
When the data can't leave: build outside, run inside
For a company, the Census Bureau's arrangement means a report or an audit can be built by someone who never sees the confidential file. The work is developed and proven on a synthetic ledger with the same entities and keys as the real one. The company's own staff then run it on the real files, and no data is sent out. As at the Bureau, the result that counts comes from the real file.
The synthetic data has to resemble the real thing
There is a fair objection to all of this. The Census file is modeled on real records. The datasets on this site are generated from written rules, with no real records behind them, so they are only useful if they look like what a real company would hold. The person who writes the generator could also keep adjusting it until the tool looks good. Three safeguards address both problems.
Settings are fixed before generation. For Quellwood and Galewick, each generator setting is drawn from a range that carries a written reason, before any rows exist.
A separate check compares the result with published figures. The check can report pass or fail, and it cannot change the data. In Galewick's program year, 8.8% of paid bonus claims carry a defect, inside the 6-15% range a rebate-audit vendor reports from its own audits. A second check failed: the share of dollars overpaid came out below an older published figure. I left the planted error rates unchanged and reported the gap, because adjusting the data until it passes would make the check meaningless.
A second review checks realism against published numbers. For Quellwood and Galewick, I had an AI assistant, Claude, review each finished dataset against published industry figures and judge whether it looked like a real company's data. Each dataset failed its first review. Each generator was corrected and the data regenerated, with the planted error rates held fixed. This is a second check on my work and not an outside audit.
These safeguards do not hold equally for all three datasets. Cinderhaven was built first, before the rules were written down. Its settings were chosen by hand and audited afterward, and in June 2026 I regenerated it because the earlier version filled orders at a rate no retailer would accept from a supplier. Its headline figures are now locked, and a build fails if one drifts. The post on why these examples are synthetic sets out the generation rules in full, with a table showing which rule holds for which dataset.
One more limit applies. I wrote both the generators and the tools that audit their output, so the two could share a mistake. Payout Math reduces that risk by checking results against values worked out by hand from the rulebook. Lift Math reduces it by finalizing the estimating methods before they could read the true values. Neither equals an independent audit. The real protection comes when the method is run on real data and reconciled against a ledger I did not write.
Three things it cannot do
It cannot tell you your own number. The $156,352 figure belongs to the modeled launch's price, payment terms and sales rate. Change any of them and it changes. Your number can only come from your own data.
It cannot surface a defect that was never planted. Quellwood's quarter holds nine kinds of defect because nine were planted. A real export can fail in a tenth way, and a tool that found all nine has not been tested against the tenth.
It cannot substitute for a reconciliation with the real ledger. Tests on synthetic data check that the pipeline is consistent with itself. They cannot check that a report's revenue matches the general ledger, because the ledger exists only on the real side. The first run on real data has to be a reconciliation.
The day real data arrives, only the numbers change
Go back to the analyst from the opening. On the day the first scan file arrives, it loads into the table built for it. The sales report runs, because it has already run every week on synthetic stores. The definitions were settled in the model. The rules have their thresholds, which were set before any results existed to influence them. The one new task is the reconciliation, and she can give it her full attention because the rest is already built.
The channel finance manager's audit was built and verified on a synthetic ledger, then run by her own staff on the real files. In both cases the decision is made on real data.
That is why this is worth doing. Mistakes in definitions, joins and rules are found while they cost nothing to fix, before anyone has acted on them. Estimates are checked against a known answer, which real data can never provide. Confidential files stay where they are. And on the first day of real data, the team's time goes to reconciling the numbers, because the reports are already built.
Keeping the reporting layer trustworthy after that day is a matter of process hygiene, and it starts from a layer whose definitions, reports and rules were already right when the data arrived.
Check the method yourself
The synthetic files, the lists of planted errors and the rules behind Campaign Math and Payout Math are published, so their figures can be recomputed. If you would like the same method applied to a report you are waiting to build, or a ledger that cannot be sent, write to me and tell me which. I will reply with the entities and keys the model would need. No call is required, and no data has to leave your company.
Pages this post draws on
Next step
Get the number for your own data. Seven business days.
Send the export you trust least; get back your three most expensive data problems, in dollars, with a fix plan. Or start with the written intake: a few minutes by form, no call.
Private, expiring upload, never email. Mutual NDA before anything moves. Files destroyed within 30 days of delivery, with a certificate. Methods published, tools open source.