TIC Risk v2
An open AI model with 232 million parameters that assesses credit and fraud risk for limited companies. It is trained on public records, with no personal data.
Download the modelHow accurate it is
Tested on companies that were excluded from training, 76% of the bankruptcies within a year were among the 10% of companies the model rated riskiest. The model is compared with TabICL, a pretrained model for tables, on the same data.
- AUC 0.924 for bankruptcy within 12 months (TabICL: 0.913)
- AUC 0.890 for bankruptcy within 24 months (TabICL: 0.853)
- Calibrated probabilities with a 90% prediction interval
- Risk class 1–5 and a score from 0 to 100 for bankruptcy within a year
Run it yourself
The tic-risk tool downloads the model and runs it offline on your own computer, with no database, network or API. It is available for Windows and Linux and comes with four sample companies. To use the model in production you need the companies' current events, from your own sources or from a credit bureau.
A base to build on
The same approach works wherever credit bureaus or other organisations have access to company events. The model can also be strengthened with your own data, for example with a LoRA adapter trained on your own outcomes.
Open to everyone
The model is openly available on Hugging Face, and the paper describes how it was trained and tested.