RISK & DECISION · SC-19
Fraud Detection & Explainability System
A cost-sensitive fraud scoring system with threshold optimization, explanations and investigation prioritization.
Starting point
Accuracy alone cannot choose a lending, fraud or review threshold. The business needs expected loss, false-positive cost and capacity constraints.
01 · BUSINESS PROBLEM
Risk decisions fail when the score is separated from the cost of being wrong.
Accuracy alone cannot choose a lending, fraud or review threshold. The business needs expected loss, false-positive cost and capacity constraints.
Wrong threshold
A technically strong model can create a poor operating decision.
Hidden trade-off
False positives and false negatives have different economics.
No explanation
Material decisions need inspectable drivers.
02 · DECISION LOGIC
From probability to an explicit risk decision.
The operating threshold is chosen using cost, risk appetite and review capacity.
Decision sequence
Each stage answers a different operating question
Produce a calibrated risk score.
Expose material drivers.
Price the cost of each error type.
Approve, review, decline or investigate.
Decision rule — The best threshold is a business decision supported by model evidence.
03 · WHAT CHANGED
Risk model, economics and decision logic in one path.
Prediction is connected to expected cost, explanation and review workflow.
Calibrate the probability estimate.
Measure performance beyond accuracy.
Optimize the operating threshold under explicit costs.
Expose explanations and review queues.
04 · ARCHITECTURE
A modular path from input to decision.
Inputs → preparation → core logic → validation → decision output
SC-19 · SYSTEM ARCHITECTURE
Inputs → preparation → core logic → validation → decision output
Public portfolio implementation
Inputs
Source signals
Capture the operating inputs required by the system. [Python]
Preparation layer
Normalize context and create a stable analytical contract. [XGBoost]
Core system
Core engine
Run the main analytical or automation logic. [SHAP]
Decision logic
Apply the rule, model or orchestration logic that changes the decision. [Scikit-learn]
Validation
Validation
Test outputs against explicit quality criteria. [Imbalanced-learn]
Controls
Keep approvals, thresholds or constraints visible. [FastAPI]
Decision output
Decision output
Expose the result in a form the user can act on. [PostgreSQL]
Monitoring
Record outcomes, exceptions and evidence for iteration. [Plotly]
Integration boundaries
Python
Defined responsibility inside the system; replaceable if another tool fits the requirement better.
XGBoost
Defined responsibility inside the system; replaceable if another tool fits the requirement better.
SHAP
Defined responsibility inside the system; replaceable if another tool fits the requirement better.
05 · EVIDENCE & ECONOMICS
Measure what changes the decision.
Public implementation, inspectable technical proof and decision-focused validation.
Risk evaluation
PR AUC
0.58
Representative public example.
Fraud recall
82%
Representative public example.
False positives
3.4%
Representative public example.
Decision coverage
Thresholds tested
15
Representative public example.
Validation cases
50k
Representative public example.
Reference economics
10k
Reference scenario
1 pp
Illustrative improvement
100
Decision value
06 · TECHNICAL PROOF
Review the code behind the project.
Tools used
BUSINESS CONCLUSION
Risk modelling creates value when it improves the decision threshold.
The commercial outcome comes from better allocation of risk and review capacity, not from a higher headline accuracy.