Forecasting

Forecasting says how much of each part, spare or material will be consumed: the expected consumption per item and in scenarios. From that it computes the reorder point and safety stock returned to the ERP. Today the forecast does not change the plan on its own: its link with planning (scheduled consumption from services, forecast orders, demand that drives the roster) is configured in each project.

Consumption patterns

Each item is consumed in its own way:

Pattern	What it looks like
Stable	Regular consumption, with little variation
Seasonal	Rises and falls with the time of year
Intermittent	Many periods without consumption and occasional peaks
Erratic	Irregular consumption, with no clear pattern

Items are also classified by consumption density: how many months in recent years had consumption. An item that is almost never consumed is not treated the same as one consumed every month.

Model selection

PlanifAI has statistical models (including ones specific to intermittent demand), machine learning models and deep learning models. For each item or group several are trained and the one that is least wrong against the history is chosen (backtest: the past is predicted with what was known at the time and compared with what happened).

It can use external variables that explain consumption, such as the planned usage hours of a fleet.

Scenarios

The forecast is not a single number:

Scenario	What it means
P50	Half of the time, less is consumed
P80	Eight times out of ten, less is consumed
P90	More is consumed only one time in ten

P50 is used to plan for what is expected; P80 or P90 are used to size safety stock, purchasing and capacity according to the risk accepted.

Backtest

The forecast can be compared with the forecast you already use: at several cut-off points in the past, which of the two came closer to what happened.

Until the next delivery

For purchasing, the question is not how much is used each month but whether what you hold lasts until the next order arrives. So the forecast is computed for a window: supplier lead time plus the review period (how often you order).

• The window is forecast as a whole, with its scenarios. Monthly scenarios do not add up: the P90 of four months is not the sum of four monthly P90s.
• Each supplier can have its own lead time and review period, and therefore its own window.

Reorder point and safety stock

From the window forecast and a service level (out of 100 replenishments, how many arrive before stock runs out) come two parameters per item:

Parameter	What it is
Reorder point	When stock falls below this quantity, you order. It is the window demand at the chosen service level
Safety stock	What is held above the expected window demand, to absorb variation

PlanifAI does not place orders: it returns these parameters to the ERP, which keeps ordering with its minimum lots and multiples.

Stock policies

A policy sets the service level per item level (more on regular items, less on occasional ones, on demand for rare ones). Several are compared with their cost: safety stock in euros, expected stock-outs per year and total cost against the parameters in the ERP today. You pick one and it is recorded.

Special cases

Case	How it is handled
New item	It leans on similar items until it has history of its own
Discontinued item	It stops being forecast from its discontinuation date
Supersession	The old item’s history is merged into the new one (configured in the project)
Interchangeable items	The group’s stock counts towards cover (configured in the project)
Maintenance spare parts	Scheduled consumption comes from the services in the plan; unscheduled, from the forecast with usage hours (the link is configured in the project)