Research · open methodology
The product rests on a research program — pacing, treasury, FX, AI governance. Papers are released with the underlying methodology and, where licensing allows, the dataset. Where we cite a number publicly, we cite the paper it came from.
Six papers · 2025–26
ADM-WP-2026-04
K. Novak, E. Lambert, R. Iyer · Journal of Marketing Analytics · vol. 14 · June 2026
We formalise pacing as a constrained stochastic control problem with non-stationary cost dynamics. The proposed estimator outperforms moving-average baselines on 11 of 12 real-world panels; mean absolute percentage error of end-of-period landing improves by 38%.
ADM-WP-2026-03
S. Aldridge, T. Wei · ADmetric working paper · May 2026
Across 2,400 advertiser-months we measure the cost of platform-native FX conversion vs. corridor-managed alternatives. The median advertiser absorbs 113 basis points of avoidable cost; the 90th percentile absorbs 207 bps. We present a decision rule for when corridor management pays for itself.
ADM-WP-2026-02
P. Halpern, M. Demir · Quantitative Finance Letters · forthcoming · April 2026
We show that 28-day forecast MAPE is causally linked to upstream reconciliation accuracy. A one-percentage-point lift in reconciliation accuracy reduces forecast MAPE by 0.42 percentage points (95% CI 0.31–0.53). Implications for audit committees and finance controls.
ADM-WP-2026-01
J. Okonkwo, A. Reinhardt · ACM CHI Workshop on AI Governance · March 2026
We instrument 38 production policy engines and measure the trade-off between guardrail latency and operator override rate. Below 120 ms P95, override rate collapses by 71%; above 300 ms, operators bypass guardrails within two weeks. Design implications for any money-moving automation.
ADM-WP-2025-06
L. Marin, F. Kovac, S. Aldridge · ADmetric working paper · November 2025
Sampling 612 mid-market firms we find a median 4.2 weeks of paid media spend held in zero-yield operating accounts. We model the opportunity cost at €148M aggregate per annum and propose a laddered sweep policy that preserves daily operational headroom.
ADM-WP-2025-05
R. Iyer, E. Lambert · Marketing Science · under review · September 2025
Using inverse-propensity weighting across 1.4M reallocation decisions, we estimate the incremental return of automated vs. manual reallocation. The automated policy yields +6.1% incremental ROAS (95% CI 4.8–7.4) with no measurable increase in brand-safety incidents.
Open methodology
Every paper ships with a reproducibility note, a statement of conflicts, and (where licensing allows) the dataset and code. Submissions and replication requests welcome.