How Narada forecasts demand for each part

Every week, Narada predicts how many of each part will sell at each branch over the next 13 weeks. It tries several ways of predicting each part that has been selling, and keeps the one that has proved best on that part’s own history.

A forecast only drives your buying once it has earned it. Until then, the part buys on its recent average.

Step	What Narada does	Why
1. Read	Weekly sales for each part at each branch, up to two years back	More history shows the part’s real rhythm
2. Clean	Lifts weeks when the part was out of stock	Empty shelves hide real demand
3. Compete	Several methods predict the part. The plain average is the bar to beat	A clever method must prove itself
4. Range	Adds a low-to-high range around the forecast	No forecast is exact
5. Earn trust	Checks the forecast against what really sold	Only a proven forecast drives buying

What does Narada learn from?

Your sales history, week by week, for each part at each branch. It reads up to two years. It also notices yearly patterns in that history, such as a part that always sells more in summer.

If you tell it Months your sales peak on Your Numbers, it gives those months to the forecast as a hint. The hint reaches the method that learns across your whole book. It does not add stock on its own.

How does Narada pick a method for a part?

Several methods compete for every part. Each one predicts the most recent weeks without being shown them. Then Narada compares the predictions with what really sold.

The bar is the plain average: the average of the last 13 weeks, carried forward. A method has to beat it clearly to win. If nothing beats it, the plain average is the forecast. That is a good result, not a failure. Many slow parts are best served by it.

Parts that sell now and then get methods built for that pattern. Steady sellers also get a method that learns from your whole book at once, so a part can borrow what similar parts teach.

What about weeks when the part was out of stock?

A week with an empty shelf shows low sales, but customers still wanted the part. Learning from that week would teach Narada that demand fell.

So when a part was out of stock for much of a week, Narada lifts that week’s sales toward its usual level first. The lift is capped, so one bad week cannot inflate a forecast. Those weeks are also left out when Narada grades how good a forecast was.

This needs your stock files. Without stock history, Narada cannot tell an empty shelf from a quiet week.

What if my history is short?

With less than about six months of sales in your files, Narada uses only the simpler methods. The method that learns across your whole book switches on once there is enough history. With less than about two months, Narada does not forecast yet. Parts then buy on their recent average.

The Data page tells you when your history is too short for the full forecast, and how much more it needs.

Is there a safety check on the forecast?

Yes. If a method predicts a week far above anything the part has ever sold, Narada does not trust it. That part falls back to the plain average for that run.

Narada also watches the whole book. If forecasts start running clearly too high or too low against real sales, it relearns straight away instead of waiting.

How often does it relearn?

Narada relearns every method from scratch once a month. Every week in between, it updates the forecasts with the newest sales, using what it learned that month.

What is the range?

Next to each forecast, Narada gives a low and a high. Real sales should land inside that range most weeks.

When Narada builds a range, it first tests it on recent weeks it held back. If too many of those weeks fell outside, it widens the range. Later, it grades ranges against real sales.

The range matters for buying. A part with a wide range needs more spare stock than one with a narrow range. See What sets your safety stock.

When does the forecast drive buying?

The forecast earns its say in two steps:

Test	If it passes	If it fails
1. It beat the plain average on that part’s own recent weeks	The forecast sets how many Narada expects to sell	The part’s recent average sales are used instead
2. Its range has held up: ranges from the same method, on parts like it, have caught real sales often enough	The range sizes the spare stock	Spare stock comes from how the part’s sales have swung

That is on purpose. A forecast that has not proved itself should not spend your money.

Each line’s working on What to Order names the method it used: the forecast, or the recent average.

What can’t Narada see?

Promotions, unless your sales already show their effect. Plan a promotion’s extra yourself.
A new customer or a lost one, until their orders show in sales.
A part that has never sold here. Narada has nothing to learn from yet.

Check it yourself

Open a part’s page from What to Order. The chart shows real sales, the forecast and its range, and the page shows how the forecast feeds the order.