Backtesting & Fine-Tuning

What it's for

Backtesting lets you validate Vision Score against your own historical portfolio — real applicants whose eventual loan performance you already know — before relying on the score for live underwriting decisions. This answers the question every risk team asks before adopting a new signal: does this actually separate my good borrowers from my bad ones?

Fine-Tuning goes a step further: once you've backtested, you can request a model trained specifically on your portfolio's data and loan performance patterns, rather than relying on the generic model.

What you can see

Batch History — every backtest you've run, with:

  • Batch ID — unique identifier for the run.
  • Date — when it was submitted.
  • Status — e.g. completed, failed, or "Uploaded data deleted" (see note below).
  • Images — number of selfies included in the batch.
  • Model — which model version was used (Generic, or a fine-tuned variant once available).
  • Gini — the resulting Gini coefficient, a standard measure of how well the score separates good outcomes from bad ones. Higher is better; a well-performing risk score typically lands meaningfully above zero.

Note on data retention: batches show "Uploaded data deleted" once TrustVision has automatically removed the uploaded selfie images and performance labels after processing. The Gini score and results remain available even after the underlying images are deleted. See Data & Privacy for how data deletion works and how to request it manually.

Fine-Tuning request form — to request a custom model, you provide:

  • Performance Labels — select which loan performance definition matches how you track bad debt (e.g. FPD30, TPD30, MOB06/09/12_DPD30). This tells TrustVision what "bad" means for your book, so the fine-tuned model optimizes for your actual definition of risk.
  • You can improve model accuracy further by including rejected-applicant data, product line segmentation (e.g. Cash Loan, 2-Wheeler, Consumer Durables, Personal Loan), and additional demographic fields, if available.

Running a backtest

  1. Run an initial backtest on a representative sample of your existing portfolio (up to 100,000 images per batch) via Sandbox → Run a Portfolio Backtest.
  2. Review the Gini score to gauge how much lift Vision Score adds on top of — or instead of — your existing credit assessment.
  3. Submit a Fine-Tuning request with your loan performance labels once the initial result is promising, so TrustVision can train a model specific to your portfolio.
  4. Switch to the fine-tuned model once available — it appears as a Model option for future backtests and for production use.

A note on privacy

Only upload data you're authorized to share for this purpose, and treat selfie images and any linked performance/demographic data as sensitive personal data throughout. TrustVision automatically deletes uploaded images and labels after processing (score results are unaffected); you can also request full deletion at any time — see Data & Privacy.

  • To try the product without a full portfolio backtest, see Sandbox.
  • To monitor how Vision Score performs on live traffic once you've gone live, see Vision Score Dashboard.
  • To move from a validated backtest to a live plan, see Billing & Plans.