From Mirror Statisticsto Border Decisions
Operationalising trade-gap asymmetries as real-time customs risk priors in low-capacity revenue administrations
Mirror statistics have diagnosed trade mis-invoicing for six decades without ever routing a single container. This paper specifies the conversion — four rules that turn a famously noisy research statistic into an operational border signal.
The gap this closes
Trade is recorded twice — once by the exporter, once by the importer — so systematic asymmetries between the two records constitute evidence of misreporting. The technique dates to Bhagwati in the 1960s and was given econometric form by Fisman and Wei's finding that the China–Hong Kong reporting gap rises with the tariff rate.
Yet customs risk engines — ASYCUDA foremost — score consignments against internal history only. What the exporting country has already published plays no role in deciding which container is examined this morning. Mirror statistics diagnose; they do not route.
Download the paper PDFHeadline result
On synthetic data calibrated to Papua New Guinea's import profile, the engine recovers planted under-declaration rates within 1.1 percentage points in every case — through fifteen per cent multiplicative noise and a full CIF/FOB round trip.
Corridors planted at 32, 22, 18 and 15 per cent recovered as 32.5, 22.0, 19.1 and 15.3 per cent, the pipeline running blind.
corridor planted at 4%
The second result matters more. The first shows the engine finds what is there; the second shows it declines to find what is not — the property on which officer trust, trader facilitation and diplomatic viability all depend.
Mirror statistics have diagnosed for six decades. They can route now.
Four noise-suppression rules
Raw gaps are noisy for known reasons: imports are recorded CIF and exports FOB; goods ship in one month and arrive in another; entrepôt trade misattributes origin. A statistic ignoring these generates false accusations at a rate that destroys officer trust and diplomatic goodwill.
Lawful differences are not fraud
CIF converts to FOB-equivalents using Comtrade Plus paired fields, otherwise a BACI-calibrated deflator. Timing absorbed by centred rolling sums. The fallback factor is a declared, inspectable parameter — not a buried constant.
Persistence beats size
A single month's gap is noise. A heading enters the prior only if the share of months breaching a ten per cent floor meets a three-quarters persistence threshold. Recurrence is the signature of a business model rather than an accident.
Gaps route, they do not accuse
Persistent headings contribute a value-weighted median gap to a corridor prior capped at 0.60. The prior modifies future scrutiny probability; it attaches no allegation to any past declaration and generates no retrospective assessment.
Transshipment triangles are composite corridors
Goods made in A and consolidated in hub S produce a spurious positive gap on A and a negative one on S. Composite corridors — for PNG, China–Singapore–PNG — eliminate sign errors rather than perfect magnitudes.
From prior to verdict
Five bounded factors, fixed weights published by the administration rather than learned. Every transform is monotone and saturating, so no input dominates beyond its weight.
Why not a learned model
Labelled ground truth in a low-capacity administration is scarce, delayed and selection-biased — audits happen where suspicion already fell — so learned coefficients inherit the administration's blind spots, including corridors whose entire declared history is contaminated. The mirror prior exists precisely because external data breaks that circularity. The clearance decision stays arithmetic.
Recovery experiment
The correctness claim is mechanical, so it admits direct validation on synthetic data with planted truth — a test no real data can provide, since real truth is unobserved.
| Corridor | Planted | Recovered | Persistent | Interpretation |
|---|---|---|---|---|
| CHN | 32% | 32.5% | 7 of 7 | Within 0.5 pp |
| SGP | 22% | 22.0% | 7 of 7 | Recovered exactly |
| MYS | 18% | 19.1% | 7 of 7 | Within 1.1 pp |
| IDN | 15% | 15.3% | 7 of 7 | Within 0.3 pp |
| AUSCOMPLIANT CONTROL | 4% | 0.0% | 0 of 7 | Correctly suppressed |
Table 1. Twelve months, seven HS-6 headings per corridor, 15% multiplicative noise, 4% reporting noise, CIF/FOB factor 0.90, materiality floor 10%, persistence 75%, fixed seed.
Prior stays at zero
Halving the floor admits the compliant corridor's largest heading in two months of twelve — still below the persistence threshold.
Robust to transit windows
Extending lag from two to three months shifts recovered priors by less than half a percentage point.
The soft underbelly
A five-point error propagates one-for-one into every prior — why Rule 01 prefers published paired fields over the fallback.
Explainability and deployment
A red channel imposes demurrage indistinguishable from a penalty, initiates valuation disputes, and opens audit proceedings — each contestable under review regimes customs statutes establish. In contest, 'why was this consignment selected' is not a courtesy but a ground.
Published weights, counterfactual verdicts. Every non-green declaration generates a document stating factor values, weights, arithmetic, and the movements that would have produced a green channel — identical for officer and appellant.
Disclosure does not invite gaming. A monotone score over behaviourally-grounded factors is gameable only by reducing the scored behaviours — the policy goal wearing a different hat.
The collective statistic, bounded. The prior is capped, carries one-fifth weight, and cannot alone cross the amber threshold from a compliant baseline. The innovation is that the conditioning statistic is measured rather than intuited.
The diagnostic precedes the relationship. Because both sides of the mirror are public, a national leakage picture is tabled before any agreement exists — inverting the usual order in which access precedes evidence.
None of the binding institutional constraints is load-bearing on day one.
Selected references: Bhagwati (1964, 1967) · Fisman & Wei (2004) · Javorcik & Narciso (2008) · Kellenberg & Levinson (2019) · Chalendard et al. (2020, 2023) · Carrère & Grigoriou (2015) · Gaulier & Zignago (2010) · Kim et al. (2020) · Citron (2008) · Wachter et al. (2018) · WCO SAFE (2018) · UNCTAD ASYCUDA (2023).
Reproducibility. The reference implementation and synthetic validation are released with the paper. No live administrative data was used. Views are the author's and do not represent any revenue administration. © 2026 Maitras.ai.