Export Assurance — Risk-Targeted Machine Learning

GoSource analysed export assurance data and built a machine learning model to test more targeted and evidence-based compliance interventions.

Challenge

Australian agricultural exports are worth about $80 billion a year. The farmers behind them spend a remarkable amount of their time being checked.

Document reviews. Inspections. Audits. Compliance checks. Enforcement actions. Sampling. Many farmers reported compliance activity two days out of every three.

And only one audit in five produced a corrective action. You can read that as a compliant industry. You can also read it as four visits in five going somewhere they did not need to go. The Department had no exporter-level risk profile to choose between those readings, and no view of any individual exporter’s track record.

Solution

We joined Export Assurance, part of the Taking Farmers to Market programme, as strategic partner. The work ran in two parts.

First, find the pain. We mapped the export lifecycle end to end, from a producer working out whether a market is worth entering through to the audit at the other end. That got laid against every commodity category the country exports: plant and meat, dairy and eggs, fish, genetic material, and live animals both livestock and not.

Then we measured it, which is the part that usually gets skipped. How much export moves, what it costs, how much regulation sits on it, what that costs, how often each process runs, and how much it hurts. Severity and frequency both, because a rare agony and a constant irritation are different problems.

Second, test whether a model could aim better than a schedule. We built one on risk profile data to predict non-compliance: who, what, where. It ran against historical audit and inspection schedules and against synthetic data, then went in front of the business users who would have to act on what it said. That last step is not a formality. A model nobody trusts changes nothing.

Outcomes

A pain analysis covering the whole $80 billion market, and a finding the Department could act on: assurance effort was not landing where the risk was.

The model beat the existing scheduling at predicting non-compliance, and pointed at entities nobody had been looking at. Its output overlays onto planned assurance activity, so targeting can be evidence-based rather than rotational.

What the Department really got was a decision it could defend. Enough evidence to move to alpha knowing why, with candidate projects ready to put in front of its own subject matter experts.

Technologies

Machine learning for predictive compliance modelling. Severity and frequency scoring over the analytics. Synthetic data for testing. Risk profiling across commodities, establishments, entities, activities and markets. Export lifecycle and service mapping. User research, and showcases with the people who would use the result.