paysim-catboost-demo-v1

Synthetic demonstration artifact. This card describes the model currently served by POST /api/score on this site. It is not a claim of production performance for any operator, sector, or customer dataset.

Identity

Training data

Public synthetic PaySim mobile-money simulator events. Labels and distributions are simulated. Nothing here is production wallet, bank, insurer, clinical, or logistics traffic.

Feature contract (request schema)

Pre-transaction fields accepted by the live endpoint:

• step — simulator hour index
• type — one of CASH_IN, CASH_OUT, DEBIT, PAYMENT, TRANSFER (one-hot encoded inside the artifact)
• amount
• oldbalanceOrg
• oldbalanceDest — accepted by the API for workflow context; the frozen tree features listed in the artifact are amount, oldbalanceOrg, and type one-hots

Artifact feature_names: amount, oldbalanceOrg, type_CASH_IN, type_CASH_OUT, type_DEBIT, type_PAYMENT, type_TRANSFER.

Training windows and parameters (as embedded)

Metadata embedded in the deployed JSON states:

• train steps 1–594; hold-out steps 595–743
• recipe note: human/catboost_bakeoff with auto_class_weights=Balanced
• encoding: one-hot type (Worker-safe; no CTR tables)
• CatBoost-style parameters embedded: iterations 400, depth 6, learning_rate 0.08, l2_leaf_reg 3.0, Logloss, random_seed 42
• hold-out average precision recorded in the artifact: 0.9562 (synthetic hold-out only)

Decision policy

The demo returns ALLOW or REVIEW under a demo threshold. Production policy would rank under a fixed review budget. The UI never presents FRAUD as an automatic factual verdict.

Calibration

Not established for real-world probability. The UI labels the output a synthetic-model risk score, not a calibrated probability.

Reason codes

Endpoint reason_codes are heuristic / model-agnostic explanations derived from input patterns (for example drain of origin balance, zero destination balance, cash-out/transfer channel). They are not native CatBoost SHAP values.

Relationship to Research figures

The Research page includes a CatBoost-versus-XGBoost comparison and a native CatBoost SHAP analysis from the local human-paper experiment pack. Those figures share the same synthetic domain and a similar three-feature idea, but this card does not claim byte-for-byte identity between every exploratory notebook run and this digest unless independently re-verified. Treat Research as method evidence; treat this digest as the live artifact identity.

Known limitations / prohibited uses

• Not for production fraud decisions, credit, employment, clinical diagnosis, or autonomous adverse action
• Does not cover SIM-swap, device, agent collusion, account-takeover, or social-engineering signals beyond the five synthetic fields
• Do not post customer, patient, or other sensitive data to the public demo endpoint
• Public CORS allows * for this synthetic demo only; customer deployments must use approved origins and authentication

Owner

MIHZI Ltd · Kigali · info@mihzi.com

Open the live demo → Research note →