Railway supplier risk · data guide

What our supplier data says, in plain words

A guided tour of the 10,000-supplier dataset for people who are not risk experts. Every chart is computed from supplier_risk_dataset.csv. Hover any chart for exact values. Metrics marked ★ matter most to a procurement engineer.

Step 1

The basics: one row = one supplier

Each of the 10,000 rows is a company that sells parts or services to the railway operator. Each row has three kinds of information:

18 risk scores
financial, quality, ESG, cyber… → 1 total risk score
a weighted average + Business numbers
spend, delivery, order behaviour

Reading a score: higher = worse

Every score runs from 0 (no risk) to 100 (maximum risk), and each score gets a label from the same four bands. Careful: this holds even for "ESG Performance" and "Credit Rating". A high number there is bad, which is the opposite of how those words normally sound.

Low0 – 24

Medium25 – 49

High50 – 74

Critical75 – 100

Data quality: the file is clean, with no missing values, no duplicates, and every label matching its score. You can trust the numbers as given. Keep in mind that the data is synthetic: supplier names are made up and there is no country, product category or history.

Step 2

The big picture

How risky are suppliers overall?

Number of suppliers per total-risk score (bins of 5 points)

How much money sits in each risk level?

Purchase-order volume (€) per total-risk level

What this means: most suppliers sit in the middle (Medium/High), because an average of 18 scores rarely hits the extremes. Still, more than half of all spend goes to High or Critical suppliers. That is the core of the problem.

Step 3

How the total risk score is built

We reverse-engineered the formula, and it is exact for all 10,000 suppliers: Total = Σ (weight × score), rounded. The weights add up to 100%. Here is how much each of the 18 scores counts. The colours group related scores into 6 "pillars".

For a procurement engineer: the three heaviest factors, Operational Resilience, Financial Health and Quality (10% each), are exactly the ones that decide whether parts arrive on time and work. A supplier going bankrupt or shipping defective brake parts hurts far more than bad press (Media Sentiment counts only 3%).

Step 4

One supplier, explained point by point

Because the formula is exact, we can show exactly where a supplier's total comes from. Each bar shows how many points a factor adds (weight × score). The bars add up to the total.

Supplier:

Why it matters: you never have to trust a "black box" score. You can say "this supplier is 87 because financial health and credit are at 100", and act on that specific cause, for example by asking for financial guarantees or qualifying a second source.

Step 5

Hidden risks: when the average lies

An average can hide a single disaster. Supplier has a total of only , but one of its scores is at the maximum. Bars past the dashed line are Critical.

suppliers look Low/Medium overall but have at least one Critical score, together holding of spend. Someone who only reads the total score would never notice them.

For a procurement engineer: always check the worst single score, not just the total. A supplier with Quality = 100 or Financial Health = 100 can stop your production line even if its average looks fine.

Step 6

Where the money is: spend × risk

Risk alone doesn't tell you what to do first. A risky supplier you buy €20k from is a footnote; a risky supplier you buy €50M from is a crisis. Each dot is a supplier (sample of ~2,300 out of 10,000).

Other suppliers Priority group (big spend + High risk + hard to replace) Hidden risk

Spend is concentrated in a few suppliers

Share of total spend held by the top X% of suppliers

Key metric: Exposure ★

Exposure = spend × (total risk ÷ 100). It is a rough "euros at risk" number. A €50M supplier with risk 87 has an exposure of €43.5M.

Because spend and risk are unrelated in this data (correlation ≈ 0), exposure adds new information. Big suppliers are not automatically safer or riskier.

Total exposure across all suppliers: .

Takeaway: the top 10% of suppliers hold of all spend. Focus monitoring there first.

Step 7

Delivery performance: what risk looks like on the shop floor

Three operational numbers describe how a supplier behaves day to day. Each chart shows the average per total-risk level.

Delivery Rating ★

% delivered on time and correct. Higher = better. Axis starts at 70%.

Order Confirmation Rate ★

% of orders confirmed in time. Higher = better. Axis starts at 70%.

Order Change Rate ★

% of orders changed after placing. Lower = better.

What this means: Critical suppliers deliver on time only ~80% of the time, against ~100% for Low-risk ones. That is one late delivery in five. The scores that predict late deliveries best are Operational Resilience and Quality, so these are the early-warning signals to watch.

Step 8

How hard is a supplier to replace?

Two internal labels describe this: Sourcing Risk Level ★ (how hard it is to find an alternative) and Supply Chain Complexity ★ (how complicated the logistics and sub-suppliers are). Each cell shows the number of suppliers; darker = more.

For a procurement engineer: a High-risk supplier is a real problem only when you can't switch. "High sourcing risk + Complex" (top-right cell) means single-source, long qualification, and certified rail parts like bogies, brakes or signalling. These suppliers need a backup plan before something goes wrong.

Step 9

The priority list: act on these first

The priority group combines three conditions:

big spend (top 25%, ≥ );
High or Critical total risk;
hard to replace (Sourcing Risk = High).

suppliers meet all three, holding . Here are the top 10 by exposure:

Hero example for the demo: SUPP_09733. €50M spend, risk 87, 14 of 18 scores Critical, delivers worse than even its scores predict. It is #1 in exposure.

Step 10

Glossary: every term in plain words

Outlined cards (★) are the ones a procurement engineer should look at first.

Business and operational fields

The 18 risk scores (0 = no risk, 100 = max risk)

Metrics we derived (not in the original file)

Step 11

What the data does not contain

Missing	Why a buyer cares	What we do in the demo
Country / location	Floods, strikes and wars are regional	Simulated, linked to the geo scores, labelled SIMULATED
Part category (brakes, bogies, signalling…)	Buyers think and search in categories	Simulated from complexity and sourcing level
History over time	No trends, no "what changed this week"	Simulated monthly history anchored to today's values
Bill of materials (which train uses which part)	No "which trains are affected"	Small simulated fleet BOM
Alternative suppliers	Can't propose a switch	Same simulated category, lower risk, ranked
News, certificates, audits	Evidence behind a score	Live web search (Tavily) and document reading by agents

Things NOT to claim: that big suppliers are riskier (they aren't, correlation ≈ 0); that ML "predicts" the total score (it is a fixed formula); that clusters were "discovered" (the data has no natural clusters).