Sample-data demonstration
This shows our AI testing process on public sample data. You can't upload your own data here.
From YourCloudGroup Model Manager.

Question: Which applicants are a higher credit risk
Sample data: Statlog German Credit (UCI)
Attribution: Hofmann, H. (1994). Statlog (German Credit Data) [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5NC77. CC BY 4.0. Modified: split into training and held-out rows, predictions added.
Predicts: credit risk (Classification)

Result
Its predictive value is MODERATE, measured on held-out rows that took no part in choosing it.
Certification: CERTIFIED — AUDITABLE & REPEATABLE PREDICTIVE MODEL
Value verified: yes. The winning model demonstrated useful predictive value over the baseline on holdout rows that took no part in choosing it, with limitations.
Winner: Wide NN, 30 epochs, best of 12 models built
Trained on: 900 rows

Certificate figures: the holdout half of the blind rows, 50 rows that took no part in choosing the winner
Balanced-accuracy lift: 47.8% (90% range 24.5-69.9%)
Balanced accuracy: 73.9% vs 50.0% for the baseline
Baseline: always predict the most common training class ('good')

Compare figures: all 100 blind rows, including the half the winner was chosen on
Correct: 74 of 100
Confusion (rows = actual, columns = predicted): actual 0 (bad): 0 (bad) 20, 1 (good) 15; actual 1 (good): 0 (bad) 11, 1 (good) 54
The product marks these Compare figures as diagnostic: "These results do not establish predictive performance" (outcome timing was not declared for this run).

The predictions file
100 held-out rows, one per line: heldout_row, checking_account_status, duration_months, credit_history, purpose, credit_amount, savings_account, employment_since, installment_rate_pct_income, other_debtors_guarantors, residence_since, property, age_years, other_installment_plans, housing, existing_credits_at_bank, job, people_liable_for_maintenance, telephone, actual_credit_risk, predicted_credit_risk, confidence.
heldout_row numbers the held-out rows 1 to N in the order the product tested them. It is not a row number from the original dataset: the split shuffled the rows.

Run: model fe238326-b603-4442-9713-a9e590f1ab89, tournament german_credit_20261028-212543, completed 2026-10-28T21:26:36Z
Dataset file sha256: 83668f2e8c5d0f27d3f1a72e3ae1d97bd60f692c4ecfff409514ccb919f5d2b9