Loading FinSeek
Initializing AI models...
Loading FinSeek
Initializing AI models...
Understanding the AI and statistical models powering FinSeek
FinSeek uses an ensemble approach with a 2-of-3 voting system. A transaction is flagged as fraud only when at least 2 models agree, dramatically reducing false positives.
Final_Flag = (Model1_Vote + Model2_Vote + Model3_Vote) ≥ 2Performance: Achieves <5 false positives and >130 true positives on test data
Optimized for minimal false positives - only flags when very confident
Where:
Base Features:
Engineered Features:
Custom Threshold: Tuned to minimize false positives while maintaining high recall
Ensemble of decision trees that vote to detect anomalous patterns
Where:
Random Forest builds multiple decision trees on different samples of your data. Each tree learns different patterns. By voting together, they catch fraud that might slip past a single model.
Example: If 7 out of 10 trees say a $10,000 3am transfer is fraud, the ensemble flags it as fraud with 70% confidence.
Custom Threshold: Optimized for the 2-of-3 ensemble voting system
Optimized to catch maximum fraud - aggressive detection
Where:
Low Threshold: Set aggressively to maximize recall (catch more fraud)
The final fraud decision requires at least 2 out of 3 models to agree. This consensus approach dramatically reduces false positives while maintaining high detection rates.
High-Precision LightGBM
Votes YES only when very confident
Random Forest
Catches patterns via tree voting
High-Recall LightGBM
Aggressive fraud detection
Example: If High-Recall LightGBM and Random Forest both flag a transaction, but High-Precision LightGBM says it's clean, the transaction is still flagged as fraud (2 out of 3).
Each transaction type has a learned coefficient that adjusts the fraud score. Positive values increase fraud likelihood, negative values decrease it.
Wire transfers and account-to-account movements
+5.0585
coefficient
ATM withdrawals and cash extraction
+3.0961
coefficient
Standard debit card purchases
-0.5993
coefficient
Bill payments and recurring charges
-9.5075
coefficient
Deposits and incoming transfers
0.0000
coefficient
How to read this: These coefficients are learned from historical fraud patterns. For example, TRANSFER transactions add +5.0585 to the fraud score, making them more likely to be flagged, while PAYMENT transactions subtract -9.5075, indicating they're typically legitimate.
Where:
R² measures how well the line fits historical data (0-1, higher = better fit)
Actual Implementation: Fits a straight line through historical quarterly fraud amounts. Extrapolates this trend forward with seasonal and growth adjustments (2.5% QoQ growth, Q4 +15% seasonal factor).
Where:
Low α (0.1-0.3): More weight on past → Stable, less reactive
High α (0.7-0.9): More weight on recent → Responsive to changes
Default: α = 0.2 provides steady predictions with some responsiveness
Actual Implementation: Applies exponential smoothing to historical fraud amounts, giving more weight to recent quarters while still considering historical patterns.
Short MA (3 periods): Captures recent changes
Long MA (6 periods): Shows overall trend
Positive Trend: Short > Long → Fraud increasing
Negative Trend: Short < Long → Fraud decreasing
Actual Implementation: Calculates both 3-quarter and 6-quarter moving averages. Projects forward by adding the detected trend to the recent average.
Why the multiplier? The budget includes not just the detected fraud amount, but also resources for investigation, prevention systems, customer support, and a buffer for potential undetected fraud. The 1.35× recommended multiplier balances comprehensive coverage with cost efficiency.