History repeats until it doesn't.
Statistical and machine learning forecasts learn the pattern of past demand. When a price, a tariff or a supplier changes, the pattern breaks and the forecast breaks with it.
Demand forecasting software that knows what moves sales.
RootCause.ai builds your forecast from the drivers behind demand, discovered in your own data. Simulate a change to a price, a promotion or a delivery date and see what it does to every week's forecast before you commit.
Demand Forecast
- Net units vs baseline
- +8.4%+8.4%
- 90% interval
- +5.1%+5.1%+11.2%+11.2%
- Peak
- 1,6931,693
Daily units
Question: What will weekly units do over the next 12 weeks, and what's driving them?
Answer: Five drivers move this forecast. The plan has the price cut and the promotion in; switch any of them and I'll re-run it.
Together they move net units +8.4% over the next 12 weeks, with a 90% interval of +5.1% to +11.2%.
Driver flow
Daily units, 10 weeks of history and a 12-week forecast with its 90% interval. Drivers changed: Price −10%; Promotion, 2 weeks. Net units +8.4% against the baseline forecast over the next 12 weeks, 90% interval +5.1% to +11.2%. Illustrative data.
Demand-driver analysis, causal forecasting and scenario modeling, on the RootCause.ai decision intelligence platform.
Why does your demand forecast keep missing?
A forecast built on sales history assumes the coming months will look like the past. Then a price changes, a promotion runs or a tariff lands, and the forecast misses. If you raise prices and the forecast ignores it, it will be too high every week from then on. Planners call that forecast bias, and it stays until you find the cause.
Nobody can say which driver moved.
When the forecast misses, the team gets a number and no reason. Price, promotions, weather and supply all changed that quarter, and the model can't separate their effects, so the same miss comes back next quarter.
What is causal forecasting?
Causal forecasting predicts demand from the drivers that cause it, rather than from sales history alone. RootCause finds those drivers in your own data. You get the size of each driver's effect and the evidence behind it. When the forecast changes, you can see which driver moved it.
What are demand drivers?
Demand drivers are the factors that change how much customers buy. Internal drivers are the ones you set, such as price and promotions. External drivers are the ones you can only plan around, such as weather and tariffs.
You control
Internal drivers
- Price
- Promotions
- Distribution
- Assortment
- Marketing spend
- Stock availability
You don't
External drivers
- Weather
- Macroeconomics
- Commodity prices
- Events
- Tariffs
What to look for in demand forecasting software
If the forecast has to explain itself and answer what-if questions, check for these six things in a demo.
Shows the drivers
Which drivers move demand, with the direction, size and evidence for each.
Lets you change one
Set a price, promotion, tariff or delivery date and see every week's forecast move, with an interval.
Counts the whole effect
Net units after the dip that follows a promotion, not only the peak week.
Uses outside data
Weather, commodity prices and macroeconomic data joined to your own on product, place and time.
Reports honest accuracy
WAPE and bias at the level you plan at, as well as MAPE.
Fits your stack
Runs alongside your planning system, inside your own environment.
From raw data to a forecast you can test.
Connect your data
Sales, pricing, promotions, inventory and external data connect into one data ontology, without a cleaning project first.
Find the drivers
SPARC-fast runs causal discovery across every variable, so the drivers come from your data rather than a workshop. Each one carries a direction, a size and the evidence behind it.
Simulate the change
Change a driver on the causal digital twin and get the new forecast for every week, with a 90% interval around it.
The forecasting questions planners ask first.
Each question runs as a simulation on your forecast. To test a promotion before you run it, set the discount, timing and products, then read the change in units, revenue and margin with a 90% interval, after the volume pulled forward from later weeks.
How many extra units does a 15% discount sell once pull-forward is counted?
See the lift during the promotion and the dip after it, so the plan reflects net units and margin rather than the peak week.
How much safety stock does a better forecast free up?
Every point of unexplained error is paid for in stock. Tie forecast error to carrying cost and service level, then see which drivers to fix first.
How much does a cold snap move unit demand in each region?
Weather sits in the model as a driver with a measured effect, so seasonal businesses plan from the forecast instead of last winter.
If a supplier slips three weeks, which SKUs run short first?
Simulate a late shipment or a capacity cut and see which products, regions and weeks feel it before the purchase orders go out.
Statistical, machine learning and causal forecasting compared
Each method has its place. The difference shows up when you want to change something.
| Criterion | Statistical | Machine learning | Causal (RootCause.ai)causal discovery and simulation |
|---|---|---|---|
| Learns from | The past pattern of demand | Past demand plus related features | The drivers that change demand, found in your data |
| Best for | Stable items with long, steady history | Many items with rich history and features | Decisions that change price, promotions, supply or timing |
| When conditions change | Follows the old pattern until refit | Needs retraining on new data | Updates through the driver that changed |
| What if we cut price 5%? | Only for drivers added by hand | Within the range it was trained on | Yes, with a 90% interval |
| Explains the number | Trend and seasonality | Which features the model used | The chain of drivers behind each change |
Start with one KPI in one market.
RootCause runs alongside your planning system and inside your own environment, so nothing your planners rely on has to change.
Pick the forecast
One KPI in one market, the forecast that costs you most when it misses.
Map the drivers
We connect your data and discover what moves that KPI.
Validate
The model is tested against held-out periods before anyone relies on it.
Simulate
Run the changes you're weighing and compare them with your current forecast.
Frequently asked questions
What is demand forecasting software?
Demand forecasting software estimates future demand for each product, place and period from your data. RootCause builds that forecast from the drivers behind demand, learned from your own sales, pricing, promotion and external data, so you can see why it moves and test a change before you make it.
What is the difference between demand forecasting and demand planning?
Demand forecasting estimates how much customers will buy. Demand planning decides how the business responds: what to make, stock and ship, and when, usually through an S&OP process. RootCause works on the forecasting side, showing what drives demand and testing decisions against it, alongside the planning system your team already uses.
How do you calculate forecast accuracy?
Forecast accuracy is usually reported as 1 minus an error measure. MAPE averages the absolute percentage error of each item. WAPE divides total absolute error by total actual demand, so high-volume items count for more. Measure at the level you plan at, usually SKU by week, because aggregate accuracy hides the misses that break shipments.
What is WAPE vs MAPE?
MAPE gives every item the same weight, so it breaks when actuals are zero and lets low-volume items dominate the score. WAPE weights the error by volume: total absolute error divided by total actual demand. Use WAPE when the question is how many units you got wrong, and MAPE when you want to compare items of different sizes.
What is forecast bias and how do you calculate it?
Forecast bias is a consistent lean in one direction. Calculate it as the sum of forecast minus actual, divided by the sum of actuals. A positive result means you over-forecast, a negative one means you under-forecast. Random error cancels out over time. Bias doesn't, and it usually points to a driver the model is missing.
What counts as good forecast accuracy?
It depends on the level and the product. Aggregate monthly forecasts are far more accurate than SKU-by-week forecasts, and stable products beat new or promoted ones. Compare against your own baseline and a naive forecast, such as the same week last year, rather than an industry average.
Does RootCause replace our planning system?
No. RootCause runs alongside the planning system and S&OP process you already use. It adds the drivers behind the forecast and lets you test decisions against it, through the product, an API, an SDK or MCP, inside your own environment.
What data do we need to start?
Sales history at the level you plan at, plus prices, promotions and inventory if you have them. Spreadsheets are fine. RootCause joins them into one data ontology and can add weather, commodity and macroeconomic data on the same product, place and time keys.
How long does it take to get a working model?
About a week on your own data for a first forecast you can test, starting with one KPI in one market.
Want to see the future in business simulations in action?
Book a 30-minute call with the team to find out more.