Systematic effects of evaluation protocols in WiFi fingerprint localisation: Uncertainty and dependence on survey geometry

The positioning error of a WiFi fingerprinting system is commonly determined by dividing a database at random into training and testing partitions. This article treats that division as a metrological question: it determines the conditions of measurement. A survey records several measurements at each of a few reference points, so dividing the measurements does not divide the positions : UJIIndoorLoc draws 19 937 fingerprints from only 933 reference points, and a random division places every test position in training. The result is obtained under a repeatability condition in the sense of the VIM: re-identification of surveyed positions, not localisation of new ones. Four protocols separating re-identification, spatial generalisation and deployment are applied across three public databases, to a region-conditioned support vector regression localiser and three further learners. Upon UJIIndoorLoc, the displacement between a random division and a device-and-time holdout is a protocol-induced systematic effect of ( − 4.60 ± 1.26 ) m against 10.85 m under the holdout’s conditions, 3.6 times the expanded uncertainty of 1.29 m evaluated per the GUM. It exceeds it under every holdout clustering, although the lower end of its interval does not under clustering by handset. Thinning the reference points against a size-matched control calibrates the effect against survey geometry, the inflation rising from 1.32 to 2.01 as spacing widens from 3 to 17 m whilst the control stays flat; the calibration is within-survey. Tampere, with one capture per position, shows no detectable effect, as predicted. Evaluations should report the protocol and survey geometry alongside the positioning error. The harness is released.