# Track record and methodology

> How FinSentS checks its forecasts against what happened, and where to read how they are made.

## Track record

**Track record** scores every forecast family against what actually happened. Each family has a row:

| Column | What it says |
|---|---|
| **Forecasts**, **Series**, **Outcomes** | How many forecasts were scored, on how many series, against how many observed outcomes |
| **Median error** | The typical distance between the forecast's median and the outcome |
| **Skill** | 1 − error / error of the last known value: above 0, the forecast beats simply repeating the last value |
| **CRPS** | A score of the whole forecast, not just its median; compare it only between series of the same scale |
| **Brier (up)** | How good the forecast's probability of ending *above* the last known value was; lower is better |
| **90 % interval held** | How often the outcome fell inside the 90 % interval — near 90 % is right |
| **PIT** | Where outcomes fell among the forecast's quantiles |

**Reading the PIT**: tall outer bars mean the intervals were too narrow, tall middle ones too wide, a tilt a biased median.

The served forecasts are stored every night and never rewritten, then scored when the outcome is known. Only the current engine's forecasts are scored.

## Methodology

**Methodology** is the reference for how the figures are made and what they cannot do: the data, the sentiment, buzz and volatility forecasts, the nowcasts, the calibrated intervals, the AI briefs and the limits. Read it there: [Methodology](https://www.finsents.com/methodology). This help explains how to use the screens; the methodology explains the numbers.