A forecast with a range beats a confident single number
Demand, cash, capacity, churn. Once your history sits in one place, it can tell you something useful about next quarter. We build models that show a range, state their assumptions in writing, and say plainly when they should not be trusted.
In short: We forecast demand, cash, capacity or churn from the history you already hold, once that history is in one place and the definitions are settled. Every forecast ships with a range rather than a point, its assumptions written down, and a stated set of conditions under which it stops being reliable.
Four things worth forecasting
Not everything benefits from being predicted. A forecast pays for itself when it changes a commitment you have to make before you can see the outcome: stock you order, people you hire, a facility you sign for. If the decision can safely wait until the number is known, the honest advice is to wait. These four come up most often because in each case the money is spent well ahead of the evidence.
Demand
How much of what, and when. It earns its keep when you buy stock ahead, book production slots, or staff to a volume that has not arrived yet.
Cash
Money in and out by week, modeled from when customers actually pay rather than from the terms printed on the invoice. Those two are rarely the same.
Capacity
Whether the people, vehicles or machines you have will cover what is coming. This is the forecast that changes hiring decisions, which makes it the one worth getting right.
Churn
Which accounts are starting to resemble the ones that left. Worth building only if somebody has agreed in advance to pick up the phone about the list.
Definitions first, model second
A forecast inherits every argument buried in the metric it is built on. If sales counts a deal at signature and finance counts it at delivery, then a revenue forecast does not settle that disagreement. It gives both sides a fresh number to disagree about, with a decimal point and an air of authority attached. We have watched this happen. The model was fine. The meeting was not.
So the first conversation is about wording, not math. What counts as an active customer. Whether a refund reverses the original month or lands in the current one. Which entities are inside the number and which are not. Sometimes that discussion takes longer than the modeling does, and it is the reason we usually want your source systems joined and agreed before we quote for anything predictive.
If a definition genuinely cannot be settled, we will forecast both versions and label them. What we will not do is pick one quietly and let you find out later.
Why every output has a range
A forecast quoted as one figure has hidden its uncertainty rather than removed it. Someone then plans against that figure as though it were a measurement, and when reality lands eight percent away the whole exercise is judged a failure, even though eight percent may have been an excellent result for that data.
We give you a central estimate and a band around it, with the band widening as the horizon extends and as recent data gets noisier. Alongside it sits a short written list of what the model is assuming: that the pricing structure holds, that the two large accounts behave as they have, that last year’s disrupted quarter was an exception rather than the new normal. Those assumptions are the part a human should argue with. The band tells you how much room to leave. Together they support a decision in a way a single number never does, because a range tells you whether the choice is close.
How we pick a model
We start deliberately simple. Last year’s figure plus growth, or a seasonal average, gives us a baseline that costs almost nothing to build. Then anything more elaborate has to beat it on data the model has never seen. Classical time-series methods handle steady seasonal patterns well. Gradient-boosted trees do better when the outcome depends on a pile of other variables, like price, weather, promotions and day of week. Deeper machine learning approaches earn their place occasionally, usually where there is a lot of history and genuine complexity in it.
Sometimes the baseline wins. When that happens we tell you, hand over the simple rule, and stop billing. It is a short conversation and not a comfortable one, but a client running a spreadsheet formula they understand is in a better position than one paying to maintain a model that adds nothing.
Living with a forecast after launch
Models decay. Your customer mix shifts, a competitor changes their pricing, a product line gets retired, and the patterns the model learned slowly stop describing the business. None of that announces itself. So we backtest against held-out periods before launch, report error in your own units rather than in a statistic nobody can picture, and then keep measuring predictions against outcomes every month afterwards. When accuracy drifts past an agreed threshold, someone gets told and the model gets retrained or retired. A model quietly degrading in the background is worse than no model at all, because people are still trusting it. Most clients receive the accuracy check as part of their scheduled reporting rather than as a separate thing to remember.
What we won’t do
We won’t forecast a metric two teams define differently
A forecast built on a contested number inherits the argument and adds false precision to it. We will hold the work until the definition is written down and agreed, even when that puts weeks between the kickoff and anything you can look at.
We won’t model a pattern your history is too short to contain
You cannot learn annual seasonality from eleven months of data, whatever the model architecture. Where the history is thin we will give you a simple rule, a review date, and an honest explanation rather than something that looks sophisticated and is guessing.
We won’t quote an accuracy figure before we have tested it
Any percentage promised at the proposal stage is invented, because nobody has seen your data yet. You get a number from us after backtesting on your own history, and it is whatever it turns out to be.
What would you do differently if you could see next quarter?
Tell us the decision the forecast is meant to support and roughly how much history you hold. We will tell you whether your data can carry it, and what to do instead if it cannot.
It depends on the pattern you want the model to see. Weekly cycles show up in a few months of data. Annual seasonality usually needs two or three years before a model can separate it from an underlying trend. With less than that we will normally recommend a simple baseline and a review date instead.
How accurate will it be?
We cannot answer that before testing on your data, and any firm number quoted up front is guesswork. We backtest against periods the model never saw and report the error in your own units, so you can judge whether the accuracy is good enough for the decision you are actually making.
What happens when something unusual happens?
Models learn from the past, so a genuine shock sits outside what they know. Two things help: the range widens automatically as recent data gets noisier, and the written assumptions let you see quickly which ones no longer hold. A person still has to make the call, and should.
Do we need a dashboard first?
Not necessarily, but you do need your data in one place with agreed definitions, which is the same groundwork either way. Plenty of clients receive forecasts as a scheduled report rather than as a screen they have to remember to open.
