Demand forecasting explained
How Flowlity builds the demand forecast and how the final forecast is decided.
Flowlity forecasts demand as a range of likely outcomes, not a single number. The range is the confidence interval: the wider it is, the less certain the forecast. Flowlity uses this uncertainty to size your safety stock, so a noisy product gets more buffer than a stable one.
How the forecast is built
The forecast comes from two layers:
AI foundation model: trained on all the products of your site, it learns the broad patterns (seasonality, trends, links between products, external signals). It is always active and produces the Flowlity forecast.
Expert models: optional models you add to one product to bring in what you know about it, for example a trend, a seasonal shape, or the launch pattern of similar products. This is the Forecast Studio feature.
The Forecast agent recomputes the forecast every week. You can also recompute it at any time with Compute forecast, which uses credits.
What the model learns from
Cleaned past demand: your demand history after Flowlity has corrected anomalies (outliers, shortages caused by stock-outs, past promotions) and removed price and promotion effects. It is normalized to the last base price, so a past promotion does not inflate the baseline. This is the history the model learns from.
Raw past demand: the history as received from your systems, kept for comparison.
Similar products: the history of a product you link to another one, mainly to manage product replacements. For a new product, the Launch model expert model is the better choice: it learns from the launch of similar products.
Forecast events: promotions and forecast adjustments, managed in Promotions. Only confirmed promotions change the forecast. Planned and in-review promotions are shown for information only.
Product lifecycle: a Start sales or End of life date set in Product strategy. Flowlity also detects launch peaks, so the model does not learn from the first stocking orders, and makes the forecast decline as a product nears its end of life.
On sites with the customer axis, Flowlity forecasts each product on its full history, then splits the forecast between customers based on their share of demand.
How the final forecast is decided
The Final forecast is the forecast Planning uses. For each future period, Flowlity takes:
My forecast, if you entered a value for that period.
Otherwise the baseline: the Flowlity forecast, or the External forecast when the product's Forecast source is set to it.
Then, if your site shows firm or planned demand in Demand, the final forecast is never lower than that future demand.
For products on a reorder point strategy, the final forecast follows future demand.
Example. A product has a Flowlity forecast of 100 units in each of three months:
January
100
none
40
100
February
100
120
60
120
March
100
120
150
150
In March, future demand already exceeds both forecasts, so it sets the final forecast.
Reading the forecast
In the product detail of Demand, the chart shows past demand next to the Final forecast, the Flowlity forecast and its Confidence interval. Turn on Raw past demand (N-1) or Raw past demand (N-2) to check the forecast against the same months of previous years.
To act on what you see, follow How to review a product's forecast, How to apply an expert model and How to recompute a forecast now on the Demand page.
Forecast accuracy
The 🎯 Forecast accuracy view of the Demand list shows how far past forecasts were from actual demand. Use it to find products that need a better input: a corrected history, a similar product, an expert model or a forecast event. See the Metrics glossary for MAE, MAPE and FVA.
Agents involved
Forecast agent: computes the Flowlity forecast every week.
Demand anomaly agent: corrects outliers, shortages, past promotions and launch peaks in your history every week.
Product similarity agent: suggests up to five similar products for each product every week.
Related pages
Last updated