What We Do
Know what next quarter looks like,and how much to trust it.
The problem with the workbook is not that it is wrong. It is that nobody can prove it is right. It is manual, it has grown byzantine, and there is no way to tell whether a cell got missed three tabs back. We build a forecast that gets tested the only way a forecast can be: freeze it, wait, and check what actually happened.
The problem
How would you knowif a cell got missed?
Most operating models are a workbook that one person maintains. It has grown for years, it holds real institutional knowledge, and it is genuinely useful. It is also impossible to audit. Tabs reference tabs, a few assumptions are hardcoded three levels down, and nobody has traced the whole thing end to end since the person who built it left.
So when the number comes out, the honest position is that you trust the person, not the model. That works until the plan built on it turns out to be wrong, and then there is no way to say which part was wrong or by how much.
We are not trying to take the workbook away. The CFO model usually encodes things a statistical model will never learn. We run ours next to it and reconcile the difference, so when they disagree you know where and can decide which one you believe.
How we test it
Freeze it, wait, then check.
A forecast that has seen the answer is not a forecast. We cut the history at a date, train on what came before, and score against months the model was never shown. Then we publish the result whether it flatters us or not.
Every location, every category, every day
One model trained across the whole estate rather than one per site. A location that opened last quarter borrows from the shape of sixty others, which is the only way a new site gets a usable forecast at all.
Staffing, promos and weather held constant
We can re-score history as if every site had been well staffed, with promotions stripped out and weather at seasonal normal. That answers a question the raw numbers cannot: what would we have sold if we had not been short handed.
Every version kept
Each month's forecast is stored as it was published. When someone asks what we said in January, the answer is a lookup, not a re-run that quietly benefits from knowing how the year went.
Unit economics
A forecast of revenueis only half the answer.
Lifetime value and payback tend to be built on placeholders that were reasonable once and never revisited. A flat percentage for card fees. Shipping cost taken from what the customer paid rather than what the carrier charged. Cost of goods from a spreadsheet that stopped being updated.
Each of those is a small assumption, and each one points the same direction: the business looks more profitable than it is. We replace them with transaction-level actuals reconciled to the general ledger, so the number ties to what the accountant sees.
What changed
What it changed.
Forecast 98.6% accurate
We locked the forecast and left it alone. Six months later the company total landed within 1.4% of actual and forty-two of forty-eight locations came in within ten percent. Staffing and budget now start from the forecast instead of arguing with it.
A year out, as accurate as a month out
Forecast error stayed flat from one month ahead to twelve, so a full year of hiring and lease decisions can be planned against it rather than revisited every quarter.
Payback finance will sign off on
Their lifetime-value math ran on a card fee somebody had estimated once and nobody had checked since. It is wired to what the processors charge per order now, so marketing sets acquisition budgets against a payback number finance agrees with.
Margin they can price against
Shipping was booked as whatever the customer paid, and most orders ship free, so it read as close to nothing. With carrier spend flowing in per order, they can price a bundle and know what is left after it ships.
What else
This one rarely shows up alone.
Want to know how good your forecast actually is?
Give us history through last year and we will show you what it would have said about this one.