Run TIC Risk Open on API data

TIC Risk Open is TIC’s bankruptcy model for Swedish limited companies, published as open weights with a command, tic-risk, that scores a company on your own machine. The model reads a company’s history as a set of files. The LENS API gives you those files for any aktiebolag, read from the same data TIC scores with, so the score you compute is TIC’s score. You can also add your own data before you score.

1. Get the command and the model.
2. Download a company’s input by its ID or by its organisationsnummer.
3. Check it and score it.

Access

• A plan with max or enterprise.
• The API key needs the risk-model-input scope. The scope is not part of any plan: contact TIC to have it added to a key. Without it the endpoint returns 403 with the code insufficient_permissions.
• Each download counts as one request against the key’s monthly limit (see Rate limits).
• Aktiebolag only, the companies the model was trained on. Another legal form returns 422 with the code legal_form_not_scored.

1. Get the command and the model

Download the signed Windows build or the static Linux build from the tic-risk releases, then download the model (about 900 MB, every file checked against its SHA-256 checksum):

tic-risk download

The command runs offline on the CPU. Only download uses the network.

2. Download a company’s input

By the company’s internal ID (Download the company’s risk model input):

curl -H "x-api-key: YOUR_API_KEY" \
  -o company.zip \
  "https://lens-api.tic.io/companies/12345/risk/model-input"
unzip company.zip -d company

Or by its organisationsnummer, with or without the dash (by registration number):

curl -H "x-api-key: YOUR_API_KEY" \
  -o company.zip \
  "https://lens-api.tic.io/companies/by-registration-number/556677-8899/risk/model-input"

On Windows, use Expand-Archive company.zip -DestinationPath company in PowerShell.

The archive holds the files at its root, in the format the command reads (tic-risk-input/1):

The data is not anonymised. IDs are TIC’s internal IDs, the same as everywhere in the LENS API, so focal is the company’s id and a related company’s ID can be looked up with Get core company details. Persons are not in the files: a board change says only which role started or ended.

3. Check and score

tic-risk check --input company
tic-risk score --input company --as-of 2026-11-10

check shows what the model reads on each date and anything it leaves out. score with one --as-of date, normally the asOf you downloaded for, prints the 12-month probability of bankruptcy with its risk class (1 = highest risk, 5 = lowest), a score from 0 to 100 and a 90 % interval, the risk history for the last 36 months, the other four predictions and the events that drive the score. It takes a few seconds.

Without --as-of the command scores every date the input was prepared for (as_of in meta.json, 37 dates with the default monthsBack) and adds a chart and a table of every prediction over time, which takes about half a minute. Add --json for one JSON line per date and prediction:

tic-risk --json score --input company --as-of 2026-11-10

check can warn about events for companies that are not in lives.csv. These are related companies of another legal form, such as an owning kommanditbolag. The model leaves them out, as TIC’s own scoring does.

The same answer as TIC

The response carries the model TIC scores with in two headers:

When X-Tic-Risk-Vocab-Hash equals vocab_hash in the downloaded model’s checkpoint/manifest.json, your model is TIC’s. The probabilities, intervals, classes and scores are then TIC’s. The model’s raw output can differ by about one in a million, because TIC scores on a GPU and the command on the CPU.

Earlier dates

The model only reads what was known on the date it scores, so one download can be scored as of earlier dates:

tic-risk score --input company --as-of 2025-05-10 --as-of 2025-11-10 --as-of 2026-05-10

The files hold the related companies the model reads on the dates they were prepared for: asOf (default today) and the same day of each of the monthsBack months before it (default 36, the months of the risk history). For another date the command warns that data on related companies may be missing; download with a larger monthsBack (at most 60) or an earlier asOf to cover it.

Adding your own data

The input is plain text, so you can add what you know about the company before you score it: append lines to events.jsonl with the company’s ID as c and the same event types and codes as the existing lines. Run tic-risk check afterwards: it lists event types, attributes and codes the model does not know, and those are left out. The code lists are in the tic-risk repository under crates/risk-tokens/texts/.

Errors