Logistics Intelligence Series | Module 3 of 6

Driver turnover:
spotting the group about to tip - before the notices arrive

Replacing a truck driver costs around €9,500. And the market is empty. This model identifies from shift plans, telematics and overtime data in which driver groups departures are building up - before the resignation letters land on your desk.

Model scenario - fictional data

This module is part of a six-part worked example about a fictional mid-sized logistics company (150 trucks, 43 regular customers). All figures, datasets, model metrics and euro amounts are fully simulated - they illustrate the methodology and the order of magnitude of the achievable effect. They are not results of real clients and not a performance promise. What values are achievable on your real data is established only in a pilot project.

01 The most expensive problem in logistics

Driver shortage is not new - but the hidden patterns behind it are

In our model calculation, turnover in road freight stands at 20% per year (an assumption - no published figure exists for Germany; USA: LTL ~12%, truckload above 90%). For a fleet of 150 trucks and 210 drivers that means: 42 departures per year that you must replace. The cost per departure - recruiting, onboarding, vehicle idle time, colleagues' overtime - is around €9,500 (published US range ~$6,000-12,000, UGPTI/FreightWaves).

But: resignations rarely come from nowhere. There are patterns in the data that are visible weeks in advance - if you look. Not in a dispatcher's gut feeling, but in the numbers your system already captures.

42 departures × €9,500 = €399,000 cost base per year

Seven retained drivers save a net €53,900 - and stabilise your route planning.

HR Data
Shift plans
Telematic patterns
Churn model
Group risk
Intervention

02 Data foundation - what HR and dispatch know together

The decisive link: personnel data + operational data + behavioural patterns

We combine three data sources that exist in every logistics company but are never analysed together: HR master data, shift-planning exports and telematics behaviour data. The simulation covers 210 drivers over 24 months.

▸ Example values - method illustration, not a performed analysis
Dataset: 5.040 Monatsdatensätze, 210 Fahrer
Fluktuation: 20.0% pro Jahr (42 Abgänge)
GroupDriversAvg. overtime (h/month)Avg. sick daysAvg. weekend shiftsShift consistency (0-1)
Long-haul south2821.41.33.60.54
Long-haul north3119.81.02.90.71
Local transport8916.20.82.10.83
Swap body6214.90.71.80.88

03 Exploratory analysis - what leavers have in common

Patterns that no performance review makes visible

Leavers vs. stayers - behavioural difference (3 months before resignation)
↳ The Early Warning Signal

3 months before resignation drivers show a clear pattern: +85% more sick days, +67% more harsh-braking events and a 28% decline in willingness to work overtime. These are not coincidences - they are the data-based "internal departure" from the company.

Turnover by length of service
↳ The Critical Phase

Drivers with less than 1.5 years' tenure have a turnover rate of 42% per year. After 4 years it drops to 15%. The first 18 months are the danger zone - and exactly where targeted intervention has the highest ROI.

Turnover risk by shift type × commute time

04 Feature engineering - reading driver behaviour

16 features from three data sources that together form a picture

▸ Example values - method illustration, not a performed analysis
Feature-Matrix: 16 Features × 5.040 Monatsdatensätze

The key: trend features. Not the absolute value matters ("3 sick days") but the change ("1 more sick day than the 3-month average"). A driver who suddenly takes more sick leave than usual sends a stronger signal than one who has always taken a lot.

05 Model - random forest + survival analysis

Two perspectives: who leaves? And when?

We train a Random Forest Classifier for the 90-day prediction ("does this driver resign in the next 3 months?") and complement it with a Kaplan-Meier survival analysis for the question "how likely is it that this driver type is still here after X months?" Reporting happens exclusively at group level.

65 %
Accuracy
60 %
Recall

(order of magnitude of the published US study)

↳ Interpretation for HR

Recall 60% means: the model identifies 3 out of 5 upcoming resignations in advance. Accuracy 65% is in the order of magnitude of the published US long-haul study (Correll et al. 2024, Expert Systems with Applications: 60-70% accuracy, 50-60% recall, strongest predictor is shift consistency). Even with these deliberately sober figures, the calculation holds.

Feature importance - what the model reveals about your drivers

The change in sick days is the strongest single predictor - stronger than tenure or salary. A driver who suddenly takes more sick leave is in all likelihood already job-hunting. But: it is the combination of sick-day trend + declining overtime + increasing harsh-braking events that makes the model so accurate.

06 Risk dashboard - risk at group level instead of individual scores

What the monthly report for fleet management and HR looks like: four groups, no names

The model works at group level - no individual scores go to managers. It reports anonymous counts and group risks: of 210 drivers, 25 fall into the highest risk group, 38 into the medium one, 147 into the low one. Which group is under pressure is shown by the drivers shift consistency, weekend load and overtime trend.

147
Low risk
38
Medium risk
25
Highest risk group
210
Drivers total
Long-haul south - 28 drivers
⚠ Group risk
high
Driver: Shift consistency low (frequent plan changes) · Weekend load well above fleet average · Overtime trend declining
Measure: Stabilise shift plans, redistribute weekend load
Long-haul north - 31 drivers
Group risk
medium
Driver: Shift consistency medium · Multi-day routes unevenly distributed · Overtime trend slightly declining
Measure: Rotate multi-day routes, extend planning lead time
Local transport - 89 drivers
Group risk
low
Driver: Shift consistency high · Weekend load at fleet average · Overtime trend stable
Measure: Monitor, quarterly team-level review
Swap body - 62 drivers
Group risk
low
Stability: Fixed routes, high shift consistency · Low weekend load · Stable overtime patterns
Measure: None - use the pattern as a model for other groups
↳ Recommendation: Long-Haul South

The long-haul south group shows the most critical pattern: low shift consistency, high weekend load and a tipping overtime trend. The measure targets the group, not individuals: stabilise shift plans, redistribute weekend load, rotate multi-day routes. Expected departures in this group at 20% turnover: around 6 per year at €9,500 each. Cost of replanning: one dispatch workshop.

07 Survival analysis - when it becomes critical

Kaplan-Meier curves show which groups you lose first

Kaplan-Meier survival curve - retention probability by shift type
↳ What the Curve Reveals

After 18 months only 62% of long-haul drivers are still with the company - compared to 78% for local drivers. The steepest drop occurs between months 6 and 14. That is the period in which it is decided whether a driver stays. Targeted driver-retention measures are the biggest lever in this window.

08 Business impact - the retention calculation

What it costs to lose drivers - and what it yields to keep them

Cost breakdown: what losing a driver really costs
€53.900
Net savings / year
7 drivers
Resignations prevented
ItemValue
Departures per year (20% of 210 drivers)42 drivers
Of which detected in advance by the model (recall 60%)25 drivers
Retained through intervention (30%)7 drivers
Cost per driver loss€9.500
Gross saving (7 × €9,500)€66.500
Intervention costs (7 × €1,800)- €12.600
Net saving per year€53.900
↳ The Real Value

The €53,900 covers only direct costs. Not included: the stability of your route planning (fewer re-routes, fewer customer failures), the better team morale (fewer overtime stand-ins) and the competitive advantage in the labour market ("They care"). In an industry where every second haulier is looking for drivers, that is priceless.

09 Data protection, AI Act & implementation in your fleet

A sensitive topic - done right

Driver turnover is a people issue, not a pure data problem. That is why the model works on aggregates: evaluation happens at group level, no person-level rankings, and the works council is involved from the start. The model does not replace a conversation - it shows where conversations are needed:

① Connect data sources

Merge HR master data + shift planning + telematics in anonymised form. Ensure GDPR-compliant processing. Involve the works council.

② Monthly group report

Aggregated evaluation for fleet management + head of HR: group risks and anonymous counts. No individual scores, no person-level rankings - neither for managers nor for dispatchers.

③ Intervention toolkit

Pre-defined measures per risk group at team level: stabilise shift plans, redistribute weekend load, adjust route design, training, incentive scheme. Make measurable what works.

What getting started looks like in your fleet is shown in our AI consulting for logistics companies.
All 6 modules: AI in logistics