What is Trustworthiness (the PC-Index)?
Trustworthiness, measured through the PC-Index, doesn't score a forecast's statistical deviation, it scores what that deviation would have cost in euros. It comes from a backtest: the method chosen for a part is refitted on older history, its forecast is replayed against actual demand through a simplified purchasing simulation, and the result is priced in euros against what a perfect forecast would have achieved. The gap is the PC-Index.
A part whose PC-Index sits near zero needs little attention. A part with a large PC-Index has genuinely cost money in the past, regardless of how small its percentage error looked. The example above shows exactly that: both parts would have had the same MAPE. Their PC-Index differs by orders of magnitude, because for the first part demand simply waits, for the second it's gone for good.
How exactly does the backtest work?
The simulation starts from zero stock, buys exactly the forecast quantity each period, with no lead times, minimum order quantities, or price tiers, and then prices what would have happened: revenue on what shipped, purchase cost on what was bought, holding cost on what sat on the shelf. Unmet demand splits into two parts: genuinely lost sales, and demand that simply waits, exactly as in the opening example. Lost sales are priced at the margin they would have earned.
Two account-level settings shape the result: how much unmet demand is truly lost rather than deferred, and what holding stock costs per year. That's not a minor detail, it's the reason the same history can produce different winning forecasts for two different customers.
Can the same history be scored differently for two customers?
Yes, and it's one of the PC-Index's most important effects.
Example (illustrative): two customers hold identical demand history for the same part.
- Customer A supplies plants that stop the moment a part is missing, so almost all unmet demand is lost for good, and their warehouse is cheap.
- Customer B's customers accept a wait, but their capital is tight and storage is expensive.
Run the same method competition on both accounts, and a method that forecasts slightly high wins for A, because a little extra stock costs less than a lost sale. That same method loses to a leaner one for B, because the surplus stock costs real money while late deliveries cost almost nothing. Neither result is wrong, the best forecast depends on what an error actually costs that specific customer.
What happens to leftover stock at the end of the test window?
The simulated window ends mid-flow, so whatever it's holding or owing at that point gets settled. Leftover stock is credited at its purchase cost and charged holding cost at the rate implied by its own demand, capped so a part with no sales at all is billed for a finite carry rather than an endless one.
Example (illustrative, assuming 20% annual holding cost):
A €20 seal selling 100 a month ends the window with 50 units left, two weeks of cover. The €1,000 is credited, and the holding charge comes to about €4, because the stock clears within two weeks. A €400 casting selling 2 a month ends with 24 units left, a year of cover. The €9,600 is also credited, but the holding charge comes to about €960, because that capital is tied up for a year. Same rule, a 240-times-larger charge, because the casting simply takes a year to sell through. Neither is scrap, both are stock worth exactly what was paid for it, but tied-up capital isn't free.
What is the PC-Index used for?
It's the most important of five measures PartsOS Forecast uses to score every forecast, and the main reason a planner doesn't have to review a thousand parts individually: a part with a low PC-Index was reliably forecastable and needs no review, a part with a high PC-Index has genuinely cost money in the past and belongs on the list. Exactly how that prioritization plays out day to day is the subject of the next article.
Conclusion
A percentage tells you how wrong a forecast was. The PC-Index tells you what being wrong would have cost, for this part, at this customer, under these conditions. That's exactly why not every wrong forecast costs the same, and exactly why it's worth knowing before you spend time reviewing a part.
FAQ
-
Why doesn't every wrong forecast cost the same?
Because the same percentage error can cost almost nothing or a great deal depending on whether unmet demand is lost or simply waits, and how expensive holding stock is for that specific business.
-
What is the PC-Index?
The PC-Index measures how far a part's forecast would have been from a perfect forecast in historical backtesting, expressed in euros rather than percentage points.
-
How is the PC-Index calculated?
Through a backtest: the chosen forecasting method is refitted on older history, its forecast is replayed against actual demand in a simplified purchasing simulation, and priced in euros against a perfect forecast.
-
What happens to leftover stock at the end of the test window?
It's credited at purchase cost and charged holding cost at the rate implied by its own demand. This scores fast- and slow-moving parts fairly, but differently.
-
Is the PC-Index the same for every customer?
No. How much unmet demand is truly lost versus deferred, and what holding stock costs per year, are account-level settings. Two customers with identical demand history for the same part can end up with different winning methods and different PC-Index values.