What next quarter looks like
Demand forecasting for small business, built from your own history: a baseline with its uncertainty drawn as a range, a method you can check, and a verdict on whether your data supports a forecast at all.
Our AI assistant asks about your project and can build you a free plan.
The problem
In most small businesses, the forecast is last year plus a gut feeling. The feeling is usually well informed. The trouble is that nobody can check it or argue with it, and it gets worse right when conditions change, which is when you need it most.
The opposite failure is a forecast that's far too sure of itself: a single line into the future, with no range and no word on what it assumes. People plan around it as if it were a fact, and when it misses, they conclude that forecasting doesn't work for them.
How we approach it
We start by trying to talk you out of it. The first question is whether your history can support a forecast at all. Sometimes it can't: there isn't enough of it, the business changed completely eighteen months ago, or demand depends on three large customers whose decisions aren't in your data. In those cases a forecast is an expensive way to dress up a guess, and you'll hear that from us even if it isn't the answer you hoped for.
If the data does support it, we build a baseline you can check. We fit the model on part of your history, test it on the part it never saw, and report how it did on that part. How well a model fits data it has already seen tells you little: one can explain the past perfectly and predict nothing, which is a common and expensive mistake.
Where machine learning fits
Sometimes a statistical model is right: seasonality, a trend, a handful of drivers. Sometimes a machine-learning model does better because the relationships are messier. We find out which one wins by testing both on your data. For most small businesses the simpler model holds its own and is far easier to keep running, and we'll tell you when that's the case.
Every forecast we deliver comes with a range, because a single number hides how uncertain the future is.
What we won't do
We won't hand you a number without a range, or a model we haven't tested on data it never saw. If the data can't support a forecast, we won't make one just because it was asked for.
Before we forecast anything, we ask three things. How uncertain is the data going in? Given that, how far out can the result be trusted? And what would it be worth to reduce that uncertainty? The third answer is often the most valuable project.
A worked example
Take a manufacturer forecasting monthly demand for its top product line.
The measured history is a solid line. From today on, the forecast is a dashed line inside a band that widens the further out it goes, as it should. Next month is a reasonable bet. Six months out is a range, and showing it as a single number would claim more certainty than the data has.
One point is marked: the month where the forecast crosses the level at which current capacity stops being enough. That's where the decision sits: hire, subcontract, or start quoting longer lead times. Knowing it's nine weeks away gives you time to choose, and "sometime soon" does not.
Fixed scope
Forecast Starter
What you get
- A baseline forecast built from your own history, drawn with uncertainty bands around it
- A verdict on whether your data supports forecasting at all. Sometimes it doesn't, and we'll tell you so
- The method, written down so you can check it
- A reusable model, so the next forecast does not start from nothing
What we need from you
- History, ideally two or more years of whatever you want to predict
- What you already know that the data does not: a price change, a lost customer, a new location
- The decision the forecast is meant to inform
What is out of scope
- Real-time systems
- More than two things forecast at once
- Deploying the model into your production systems
- How long it takes
- Two to three weeks. Start dates depend on current availability.
- What happens after
- Thirty days of fixes are included, then the optional care plan. For forecasts that usually means a monthly re-run and a short note on what changed.
- Price
- We give you a fixed price before we start, usually within two business days.
Need something bigger?
A forecast needs updating. The useful version is a monthly re-run with a short note on what changed and why, and a care plan covers that without a bigger project. If you need the model deployed into your production systems, that's engineering work, and we scope it separately.
Every project is handled directly by a member of our team.
If you let us publish what we build together, you get a discount. The work has to be something we can publish, and you approve every word before it goes live. Our team applies the discount when we quote.
Questions people ask
How much history do I need?
Two years is comfortable. A year works if the pattern is stable. With less than that, the best we can usually give you is a range based on judgement, without a model, and we'll tell you so.
What if my business changed recently?
Then the older data may mislead the model. Sometimes the fix is to use only what came after the change; sometimes it's to model the change directly. Either way, we decide it with you and write it down where you can see it.
Can you forecast something we have never sold?
Not from your own history, because there's nothing in it to learn from. That's a different exercise, built on comparable products and assumptions, and it should be labeled that way so nobody mistakes it for a forecast.
How accurate will it be?
We can't know before seeing the data, and anyone who tells you sooner is guessing. What we can promise is that you'll be told how the model performed on data it had never seen, measured in a way that makes sense for your decision.
Is this machine learning?
Sometimes. We test both statistical and machine-learning approaches and use whichever performs better on your data. The simpler one often wins, and it's also cheaper to keep running.
What decision is your forecast meant to inform?
Tell us the decision you're trying to make. If we're not the right people, we'll say so.
Our AI assistant asks about your project and can build you a free plan.