When the Auto-Bid Channel Replaces the Baseline: Baseline Flexibility, Experiment Calibration, and the Return Meridian Reports for a Performance-Chased Channel
Year: 2026
Type: preprint
Abstract: Platform auto-bidding sets a channel's budget from recent conversions, so the channel's spend tracks demand. On a third-party synthetic benchmark with endogenous spend (Heusch, arXiv:2608.21130), Meridian 2.0.0 was fitted across the benchmark's budgeting mechanisms at five seeds, varying the number of time knots in the model's baseline, the strength of performance chasing, and whether the channel's prior was calibrated with a go-dark experiment. With one knot, which is what the library's minimum configuration fits to a single-market model, the performance-chased channel absorbed most of the sales level: the fitted baseline averaged 327 against true weekly non-marketing sales of 2,298, and the channel's average return read 37.50 against a true 4.29, with the 90 percent interval excluding the truth on four of five seeds. The overstatement fell monotonically as knots were added. At 52 knots or under automatic knot selection the posterior for the channel's return did not differ from Meridian's default prior (median 1.22), which is also where PLA's and Meta's returns sat on every row without the mechanism. Across auto-bidding strength the overstatement appeared abruptly, with each instance settling either at the prior or at about nine times truth. Calibration with the experiment brought the average return to 3.83 to 4.16, within ten percent of each experiment's own measurement, while the marginal return on which reallocation acts stayed at 1.79 to 2.35 against a true 4.15, because the response curve's half-saturation point remained at its default prior. No configuration tested recovered both. A standalone package regenerates every fit exactly; its long-chain refits resolve the seed-level instability in 45 of 49 fragile fits to a single answer.
Author: Birba, Trevor
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7221521777