Supplier Risk Scoring With Machine Learning

Supplier problems are visible before they happen

When a supplier collapses or quietly stops prioritising your orders, the post-mortem almost always finds warning signs. Deliveries had been slipping for months. Part shipments had become normal. Their account manager left. They asked to change payment terms. Nobody put those facts next to each other.

Supplier risk scoring is the discipline of putting them next to each other on a schedule. Machine learning helps where there are many suppliers and many signals, but a lot of the value comes from simply collecting the data into one view.

Internal signals, from your own systems

These are the most predictive signals for most SMEs because they describe how the supplier treats you specifically.

A request to move from 60-day to 14-day terms, following two months of short shipments, deserves attention. Separately, each looks routine.

External signals worth adding

• Credit scores and limits from a credit reference agency
• Late filing of accounts or changes of directors at Companies House or the local equivalent
• County court judgments or insolvency notices
• Adverse news mentions, filtered for relevance
• Sanctions and restricted party list checks, particularly for Gulf and international trade
• Geographic concentration risk, such as several critical suppliers in one port region

Language models are genuinely useful for the news part: reading articles and classifying whether a mention is a plant closure or a sponsorship announcement. That is a narrow, reviewable task, which is where they belong.

From signals to a score

There are two honest ways to build the score, and the choice depends on your history.

1. A weighted scorecard. Normalise each signal, weight by judgement, and sum. Transparent, quick, and the right starting point when you have not had many supplier failures to learn from.
2. A supervised model. Train on past supplier outcomes, such as a supplier that was dropped, failed or caused a major disruption. This needs a reasonable number of past bad outcomes, which larger procurement teams have and most SMEs do not.

A middle path works well: start with a scorecard, record outcomes consistently, and after a year or two use that history to check and re-weight the scorecard with a model. We explain the general trade-off in machine learning versus a rules engine.

Whichever you choose, show the reasons with the score. A buyer who sees risk 78 will ignore it. A buyer who sees on-time delivery fell from 94% to 71% over three months and the company filed accounts late will pick up the phone.

What to do with a high score

A score is only useful if it triggers something. Agree actions per band before switching anything on.

• Low risk: normal review cycle
• Rising risk: buyer conversation, check for alternative sources, review stock of critical parts
• High risk: qualify a second supplier, increase safety stock on single-sourced items, reduce prepayments
• Critical: escalate to operations and finance, activate contingency plans

Weight the response by how critical the supplier is. A high-risk stationery supplier is an inconvenience. A medium-risk sole supplier of a component in every product you make is a board-level item.

When this is more machinery than you need

If you buy from fifteen suppliers and know every one personally, a quarterly review meeting with a simple spreadsheet will do. Machine learning adds value when you have hundreds of active suppliers, lean stock, or a supply base in several countries where nobody has a feel for every relationship.

It is also pointless if purchase order and receipt data is unreliable. If goods are received in bulk at month-end rather than when they arrive, on-time delivery figures are fiction. Fix receiving discipline before building scores on it.

A worked example for a mid-sized manufacturer

Picture an engineering firm with 220 active suppliers, of which around 30 supply parts that go into most finished products. Purchasing is three people. Nobody has time to look at delivery trends for 220 companies, so attention goes to whichever supplier caused trouble last week.

A weekly scorecard ranks all 220 but only pushes the critical 30 into the buyers' Monday review, plus any other supplier whose score jumped sharply. In a typical first run, a few names on the list surprise nobody, and one or two are single-source suppliers whose delivery performance has been sliding quietly for a quarter. Those are the conversations worth having now, while there is still time to qualify an alternative.

How we build it

At SpiderHunts we usually start with an internal-signals-only scorecard built from ERP data, because it can be ready in a few weeks and it surfaces problems immediately. External data feeds and a model come second, once the purchasing team is using the first version.

The score updates weekly and lands where buyers already work, whether that is the ERP, a shared dashboard or a Monday email. Related process work is covered in procurement and supplier automation, and the modelling sits within our machine learning practice.

Frequently asked questions