Playbookovi/ Zadnje ažuriranje: 2026-10-29

Playbook: the forecast said all is well — was it ever right?

The situation

A forecasting tool says *80 % chance this invoice is paid late*. Should you call the customer today? That depends on one thing nobody usually shows you: when this tool said 80 % in the past, how often was it right?

Otiox freezes every prediction at the moment it is made and checks it against what then happened. So each detector carries its own track record, and you can see which ones to act on and which ones are still learning.

Before you start

Accuracy is measured only on predictions whose outcome is known. A prediction still open does not count yet.
Nothing here changes a forecast. The engine's confidence is shown as it was given.

Step 1 — See how good the forecasts are

Open Command → Accuracy & calibration.

Accuracy & calibration: predictions with known outcomes and the Brier score

What you see: in the demo 43 predictions with a known outcome in the last 180 days and a Brier score of 0.2564, marked *Moderate*. The Brier score is the average squared error: 0 is perfect, 0.25 is a coin toss. Next to it: how many of the 51 detectors have enough decided cases to be calibrated (none yet in the demo; a detector whose outcomes are all one kind cannot be calibrated).

Step 2 — Read each detector's record

Scroll to By detector.

By detector: what each detector said, and how often it happened

What you see, per detector: its Brier score, calibration error, how often it came true, and *said / happened*:

|---|---|---|---|---|

| Invoice payment risk | 80 % | 100 % | 0.04 | 9 (all happened) |

| Stockout risk | 90 % | 83 % | 0.14 | 6 (5 happened) |

| Lot expiring | 85 % | 0 % | 0.73 | 6 (none happened) |

How to read it: invoice payment risk is trustworthy, even a little cautious. Stockout risk is close to its word. Lot expiring is badly overconfident: six times it warned, six times the lot was used or sold in time. Treat its warnings as reminders until its record improves.

Averted counts cases where someone acted and the event did not happen. They are not counted against the detector.

Step 3 — Use the forecasts that earn trust

Open Command → Forecasts.

Forecasts: what runs out first, from recent rates and incoming purchase orders

What you see: what runs out first, from recent consumption and the purchase orders already on the way, each with a range rather than a single date. These forecasts are frozen too, so in a month you can check them the same way.

Why it matters

A confidence number without a track record is decoration. With one, it becomes a decision rule: act on the detectors that have earned it, review the others.

What you have at the end

Every prediction checked against what happened.

A track record per detector, in plain numbers.
Forecasts with ranges, frozen for later checks.

Screenshots come from the Otiox demo workspace. Choose Switch to demo in the sidebar to follow along; changes there never touch live data. Command reads your records and never changes them.