Fabrice Colas, Ph.D. Risk, data and R&D, from model to decision

Since 2010, I have modelled the risk of adverse events at scale to support treatment choices, limit fraud or focus the allocation of prevention budgets.

This work dealt with genomic data (CNV, methylation, EST, ontologies, HLA), hospital registries (in psychiatry and cardiac surgery), online payments characterised by "their DNA" or, recently, road and rail transport networks.

For example, the R&D programme on road risk produced an ex ante score

, built on the infrastructure and the near environment of roads, which identifies the 5% of the network where 53 to 72% of the following year's crashes will occur, depending on the road user, that is 11 to 47 points better than a score built on historical hot spots alone (ex post).

To bring these models to market, I founded with my own funds the company Ubivar in 2014 (anti-fraud), evaluated at five online merchants, then Oriskami in 2020, whose SAMRoute software was notably evaluated by SNCF Réseau and by ADIF, in Spain.

Based in Rennes, France
English (professional) · French (native) · Spanish (active intermediate) · Dutch (passive intermediate)

hello@fabricecolas.me LinkedIn

Road risk (crash analysis) · Independent R&D programme
2022–2026
The lack of road safety concerns local authorities, ministries, motorway operators and insurers. From 2019 to 2024, the roads of France, the Netherlands and Great Britain counted 1.5 million victims (killed, seriously injured, slightly injured). To avoid as many victims as possible on a constant budget, local authorities, ministries and motorway operators direct their road works, public investment and network monitoring towards the riskiest roads, while insurers price motor risk.
The spatial distribution of crashes on the network is therefore at stake for these various actors. By default, the usual approach consists in studying black spots (hot spots, ex post). Now, by construction, a black spot is a section with at least one past crash, i.e. 3.9% of sections from 2019 to 2024. Yet the crashes of 57.8% of 2024's car victims occur on sections without any crash from 2019 to 2023 (de novo).
This R&D programme aims to develop a risk score (ex ante) for every section, with or without a past crash, from 281 open predictors describing the infrastructure, its use and its near environment.
- The score reveals, ex ante, that 5% of the network concentrates 53–72% of future victims
- Compared with the traditional ex post approach, that is 11 to 47 percentage points more
- Out-of-sample performance is 89–92% (AUC, +15 to 34 points over the PoC)
1. 2026Risk score including the near environment (281 predictors), 2.3 M km of roads, 1.5 M victims (FR, NL, GB, 2019–2024), published as a preprint
2. 2023–2026SAMRoute, road-risk assessment software, launched in September 2024
3. 2022–2023PoC for a startup, 70 M km of roads (global), 2 800 M scores (4 road users)
Rail risk (level crossings) · Independent R&D programme
2025–2026
Level crossings concern rail and road managers (local authorities, the State). In France, from 2020 to 2024, 99 people died in collisions with a train at a level crossing
. In 2024, the French rail safety authority (EPSF) recorded 32 significant accidents across 15 000 level crossings
, with 20 killed
. To reduce this toll, the State and rail managers make level crossings safer or remove them. However, removing a level crossing typically costs €10–20 M. Since 2018, only 5 level crossings have been removed each year on average
. In 2024, the national level-crossing safety programme lists 146 level crossings
.
The choice of level crossings to treat rests mainly on the crash history, road and rail traffic (the traffic moment), safety equipment and road type. In fact, 75% of significant accidents from 2016 to 2024 occur at the 30% of level crossings with the highest traffic moment
. However, the stakes at level crossings are not limited to the lack of safety. Also taken into account are the impact of level-crossing closures on the near environment (for instance, on the arrival of emergency services on site), and the functional dependence of sensitive sites on a mandatory crossing point such as a level crossing.
This R&D programme on level crossings rests on three orders of reading. The first takes up the current criteria, namely the structure (safety in place, geometry) and the crossing traffic (road, rail). This first order falls to road and rail managers. The second consists in listing the sensitive sites near level crossings by cross-referencing open data. The third consists in modelling the interaction of the near environment with level crossings through simulation.
- Simulation of the dependence of sensitive sites on level crossings (inverted origin–destination, IOD)
- Simulation of the near environment of level crossings under a higher rail frequency
1. 2026Impact of a higher rail frequency on emergency services (ongoing study)
2. 2025–2026Software measuring the dependence on level crossings (IOD), evaluated by the level-crossing teams of SNCF Réseau (FR) and ADIF (ES)
Online payment fraud · R&D work (Travix, then Ubivar)
2013–2018
In e-commerce, before an order is sent to the customer, the online payment is commonly analysed for its fraud risk. If fraud or a payment default occurs afterwards, the cost price of the order is a dead loss for the merchant, as are the related handling fees. And if the merchant suffers repeated fraud, the major card networks such as Visa and Mastercard can place the merchant under monitoring, or even withdraw the merchant's right to accept the networks' cards.
The serious adverse event of e-commerce is therefore the payment default. To protect themselves, merchants use anti-fraud systems. Traditionally, these systems comprise lists (white, grey, black), override rules and scoring rules combined with thresholds for automatic rejection and manual review. However, such a system has thousands of parameters that drift over time. As analysts often lack time and analytical perspective, configuring and maintaining these traditional systems is difficult.
The R&D work on anti-fraud has three phases. The first simply analyses the anti-fraud system, finds rule or threshold changes and deploys them to production, on the anti-fraud system in place, with constant resources and within a few weeks. The second aims to load 1.7 million card payments into a back-testing engine, to evaluate changes before they go into production. The third aims to develop autopilot anti-fraud software, replacing traditional anti-fraud.
- Risk Management Intelligence (2017, 242 pages, ISBN 9781979086950)
- At Ubivar, 94 online merchants targeted, 11 commercial proposals and 5 clients
- At Travix, with the fraud risk under control (0.063% → 0.1%), the share of automatically rejected bookings falls from 18.5% to 2.4% (−87%), i.e. €64–70 M a year (7 to 8% of sales recovered out of the €900 M of card payments). As for the manual-review rate, it falls over the same period from 17.6% to 13.7% (−22%, i.e. −90/day)
1. 2014–2018Ubivar, SaaS anti-fraud software for online merchants
2. 2013–2014Travix (Amsterdam, NL) and Vayama (California, US), building and running a back-testing engine, 32 rule and threshold changes put into production
Health risk · Academic R&D (NL, US, FR)
2010–2021
A health system allocates a given budget so as to produce as much health as possible for its reference population. Within this frame, insurers, the State, private funds and care providers steer this allocation towards care, investment, medical devices and research. The role of health researchers is to identify the causes and risk factors of diseases, to design the means to prevent, screen for and treat them, to test each means against the reference treatment, then to pass the result on to practitioners, authorities and manufacturers.
High-volume problems present few adverse events, frequent enough for a single model to describe each event. In health, by contrast, adverse events abound, such as death, post-operative complication, relapse in addiction care and screening error. Each adverse event has its own time horizon and, often, a set of risk factors that act together. Research teams therefore bring together practitioners, epidemiologists, statisticians and sometimes engineers, in order to define the event and its horizon, choose the study design, quantify the uncertainty and bound what the data allow one to conclude.
Health work addresses varied adverse events, such as the onset of a disease, death, post-operative complication or relapse. The work consists in merging heterogeneous data (omics, clinical, follow-up, demographic), choosing the measures, correcting sampling bias (adjustment, matching), identifying the risk factors and, ultimately, reducing the uncertainty linked to treatment choices.
- Translational Psychiatry (2020, IF=6.222, DOI)
- European Journal of Cardio-Thoracic Surgery (2020, IF=4.191, DOI)
- The Journal of Cardiovascular Surgery (2020, IF=1.888, DOI)
- BMC Genomics (2012, IF=4.397, DOI)
- Movement Disorders (2011, IF=4.558 in 2012, DOI)
1. 2021Planning of a clinical trial in addiction care, CHRU de Brest (FR)
2. 2018–2019Analyst methodologist, cardiac surgery, CHU de Rennes (FR)
3. 2010–2013Postdoc in human genetics at UCLA (US), UMC Utrecht and LUMC (NL)

Road risk (crash analysis) · Independent R&D programme

Rail risk (level crossings) · Independent R&D programme

Online payment fraud · R&D work (Travix, then Ubivar)

Health risk · Academic R&D (NL, US, FR)