tic-oss/tic-risk-v2

TIC Risk Open

TIC Risk Open is a 232M-parameter model that predicts the risk of bankruptcy for Swedish limited companies and can also be used to understand how to train a similar model for other countries.

TIC Risk Open was trained on public records for every limited company in Sweden, with no personal data, and reads each company's history as a sequence of events: annual reports and their figures, late-filing fees, tax registrations, enforcement debts, payment orders, board changes (without names) and related companies. For each company it predicts:

Prediction	Name	Horizon
Bankruptcy	bankruptcy_12m	12 months
Bankruptcy	bankruptcy_24m	24 months
New or increased enforcement debt	enforcement_debt_12m	12 months
Tax registration revoked for misconduct	fskatt_revoked_12m	12 months
Overdue annual report	report_overdue_12m	12 months

Each prediction is a probability with a 90% prediction interval. The 12-month bankruptcy prediction also comes with a risk class (1 = highest risk, 5 = lowest) and a score from 0 to 100 (100 = lowest risk).

The paper describes how the model was trained and how it compares with TabICL.

How to use

The tic-risk CLI runs the model offline on the CPU. Get the signed Windows or static Linux binary from GitHub, which comes with four example companies, then download the model and score one of them:

tic-risk download
tic-risk score --input examples/company-b

tic-risk download checks every file against its SHA-256 checksum. If you prefer, you can download the model with hf download tic-oss/tic-risk-v2 --local-dir tic-risk-model instead.

To score your own companies, write their events in the input format described in the CLI's README and run tic-risk check --input <dir> first to validate the files.

Model details

• Input: each event becomes one token. Codes start from a text embedding of the code's description, amounts are scaled logarithmically, and each token carries its date. The model only sees events that were known on the date it scores.
• Architecture: 18 layers with a hidden size of 768. Three of every four layers are gated delta-net (recurrent) layers and every fourth is full attention.
• Related companies: owner companies and audit firms are always included, plus up to 24 other related companies.
• Economy: 26 national series, including unemployment, GDP, prices, house prices and interest rates.
• Output: a probability for each of the next 24 months. The probabilities are calibrated on companies held out from training, and the prediction intervals come from the errors on the same companies.

Training data

Public records for every limited company in Sweden, from the company registry, the tax agency, the enforcement authority and the statistics office. The model uses no personal data. When a person holds a role in a company, the model only sees that the role exists.

Evaluation

Evaluated on a test set of 10% of the companies, which were excluded from all training, scored at 12 monthly dates:

Prediction	Positive rate	AUC	PR-AUC	Caught in riskiest 10%	Calibration error
Bankruptcy, 12 months	0.91%	0.924	0.254	76%	0.0009
Bankruptcy, 24 months	1.85%	0.890	0.201	64%	0.0007
Enforcement debt, 12 months	5.22%	0.891	0.508	64%	0.0051
Tax registration revoked, 12 months	0.65%	0.951	0.195	87%	0.0017
Overdue annual report, 12 months	3.83%	0.894	0.396	64%	0.0030

"Caught in riskiest 10%" is the share of companies with the event that the model ranked among its riskiest 10%.

Compared with TabICL on 605 features built from the same data, on the same test set:

Prediction	AUC	TabICL AUC	PR-AUC	TabICL PR-AUC
Bankruptcy, 12 months	0.924	0.913	0.254	0.256
Bankruptcy, 24 months	0.890	0.853	0.201	0.149
Enforcement debt, 12 months	0.891	0.884	0.508	0.520
Tax registration revoked, 12 months	0.951	0.952	0.195	0.243
Overdue annual report, 12 months	0.8944	0.8936	0.396	0.400

TIC Risk has the higher AUC on four of the five predictions. TabICL has the higher PR-AUC on four, partly because calibration gives groups of companies the same probability, which lowers PR-AUC more than AUC.

Limitations

• Swedish limited companies only. Other company types and other countries are not covered, but the way the model was trained can be used as a reference for building one.
• Calibrated probabilities come in steps, so similar companies can get the same probability.

Extending the model

The model is meant as a starting point. You can train a LoRA adapter on your own companies and outcomes and compare it with the base model, or train a model for another country the same way from its company events.

Files

checkpoint/          model weights (safetensors), manifest and vocabulary
model/               vocabulary, code embeddings, normalisation and economic data
calibration/         calibration, one file per prediction
bands/               prediction intervals, one file per prediction

License

Dual-licensed under MIT or Apache-2.0, at your option (LICENSE). The metadata lists Apache-2.0 because a model card can only name one license.