How fast facts go stale
fid scores a stored fact by how long ago it was measured, using a half-life per type of information. Tested against four public datasets where the truth is known, the hand-picked half-lives were off by up to —×, and most facts decay toward a baseline rather than to zero.
Guessed half-lives against measured ones
Half-life here is the time until half of the stored facts are wrong. A company's CEO lasts —, not the 1 year guessed. A moving ship's position lasts —, not 4 hours. Log scale.
Every fact, scored on data the fit never saw
Each model is fitted on half of the companies, airports or ships and scored on the other half. Brier score is the mean squared error of a probability; lower is better, and the best in each row is bold. "Constant" ignores age entirely.
| Brier score | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| Fact | Data | Observations | Guess | Measured | Baseline | Guess | Decay to 0 | To baseline | Constant |
Facts decay toward a baseline, not to zero
A week-old "the airport is VFR" is still right — of the time, because clear weather is the usual weather. fid's original formula sends every stale fact to 0. Adding a baseline b, the chance a stale fact is right anyway, cuts the largest gap from — to —. Where nothing is right by chance, as with a former CEO, the fit finds b = —.
Airport flight category unchanged
VFR, MVFR, IFR or LIFR, 10 US airports, a full year of reports
Airport still IFR or worse
Only reports of low ceiling or visibility
Same person still CEO
Wikidata, yearly snapshots 2004–2016
Half-life depends on what the thing is doing
Ship still within 1 km of its reported position
Cargo ships and tankers in US waters on six days spread through 2023, from AIS. Same ships, same days: a moving ship's position is half wrong after —, while — of moored or anchored ships are still within 1 km after 12 hours. Square-root time scale.
Some facts come back
Airport temperature still within ±2 °F
A reading 24 hours old is right — of the time; one 12 hours old, —. The daily cycle brings it back, and no decay model can express that.
Flight legs are not independent steps
— US flights in 2024. fid multiplies a fixed noise per processing step, which assumes the steps are independent. For an aircraft's legs through a day, they are not: what happened on the last leg matters far more than the average.
On time, by what the previous leg did
Arrived within 15 minutes of schedule, not cancelled or diverted
On time, by scheduled departure
Delays build through the day
On time, by month
The same schedule is a weaker source in some months than others
On time, by airline
One score per source hides a spread from — to —. Carrier codes as reported to the DOT.