HomeExpertiseBackgroundCase StudiesBook a consultation

Mobility · Car-sharing · Cyprus

An explainable behavioural risk-scoring & rating system

For a car-sharing platform on a Cyprus island fleet, I designed a system that moved risk management from reactive incident handling to early, interpretable prediction of risky driving — visible to both operations and drivers.

Context

On an island, the stability of a car-sharing service depends heavily on how well behavioural risk is managed. The fleet is limited and every vehicle is a valuable operational asset — even a small rise in incident rates quickly turns into higher repair costs and reduced availability for customers.

Users were a mix of local residents, expatriates and tourists with very different driving experience and habits. Tourism seasonality added pressure: during peak periods large numbers of new users joined, increasing the likelihood of risky driving and operational incidents. Under these conditions, investigating problems after they happen does little to prevent the next one.

The core challenge

Incident management was reactive: when a problem occurred, teams investigated it manually, reconstructing events from fragmented data across several systems. As the platform grew, each new incident demanded more manual analysis while the behavioural patterns behind them stayed invisible until damage had already occurred. A relatively small group of users accounted for a disproportionate share of incidents — but they were usually identified only after several violations. The challenge: surface risk accumulation earlier, before it became an operational problem.

The solution — two layers

1. Predictive scoring. Trip events were translated into a cumulative behavioural risk score. Rather than abstract user profiles, the model used observable signals from real usage — speeding, aggressive driving, nighttime conditions, correct use of lights, parking behaviour and rental completion. Context mattered: nighttime driving wasn't automatically negative, but it raised the relative weight of other signals because conditions are statistically riskier at night. Crucially, the model was kept interpretable — every risk signal traceable to observable events, not opaque algorithmic decisions.

2. Rating system. Internal scoring was translated into a rating visible to operations and users. Each rental contributed to a rating history, and every change could be linked to specific trip events. Operations could spot recurring patterns without manual reconstruction; drivers got understandable feedback and could see how their behaviour affected their reliability — turning the system into behavioural feedback, not just enforcement.

Operational impact

The platform shifted from reacting to incidents to highlighting the patterns that historically preceded them — intervening earlier and reducing repeat violations among high-risk users. Severe damage cases became rarer, disputes easier to resolve with documented evidence, and support workload fell. Most importantly, the platform could keep growing its user base without operational workload rising at the same pace.

Quantified outcomes (NDA-safe ranges)

Incident rate per rental−20–35%
Repeat incidents in high-risk cohorts−30–45%
Medium & major damage cases−25–40%
Speed-related violations−15–30%
Support tickets per 1,000 rentals−20–30%
Dispute rate−25–40%
Average case resolution time−30–50%
Vehicle downtime due to repairs−15–25%
Retention of reliable users+10–20%

Key design decisions

  • Interpretability over maximum complexity. A fully opaque ML model might predict marginally better, but operations needed decisions they could explain and defend to users.
  • Patterns, not isolated events. One speeding event doesn't make a risky user; recurring combinations do. The system interpreted behaviour over time.
  • Analytics separated from UX. Internally rich; externally a clear rating with a short explanation — preserving both usefulness and trust.
  • Grounded in real operations. Historical incidents and support cases defined which signals actually mattered for the platform's risk profile.

Strategic insight

The project showed how behavioural data becomes a practical decision architecture: interpretable signals, operational processes and user feedback combined into one coherent product logic. The same principles extend beyond mobility — to marketplaces, fintech and other digital ecosystems where user behaviour directly drives operational risk.

← Back to all case studies