Fraud prevention teams are hired to stop fraud. Their success metrics are typically some combination of fraud rate, chargeback rate, and fraud loss in dollars. What those metrics do not capture is the other side of the ledger: every legitimate transaction that gets declined because the risk score was elevated is also a cost, and at most digital-finance platforms, the cost of false declines exceeds the cost of the fraud they prevent.
This is not a new observation. Industry surveys have consistently estimated that false declines cost global merchants and platforms two to three times more in lost revenue than the fraud they prevent. The ratio varies by product type and customer segment, but the direction is almost universal: over-declining is the larger financial problem, not under-declining.
Why False Declines Are Underreported
The reason fraud teams underweight false decline impact relative to fraud loss is measurement asymmetry. Fraud losses are visible, concrete, and attributed: you know the transaction ID, the amount, the outcome. False decline costs are largely invisible to the fraud team. The declined legitimate customer does not file a report. They leave, often without appealing the decline, and their lost revenue never appears in the fraud team's data.
Customer support data partially captures this: customers who call or write to dispute a decline provide a signal. But research on customer behavior after declines consistently shows that the majority of declined legitimate customers do not contact support. They abandon the transaction and often the platform. The customers who do appeal are the more motivated ones; the less engaged ones simply churn silently.
The result is that fraud teams see their fraud metrics clearly and their false decline metrics poorly. Tightening thresholds improves the visible metric (fraud rate goes down) while degrading the less-visible one (conversion and retention). The incentive structure rewards over-declining because the downside of under-declining is attributed to the fraud team while the downside of over-declining is attributed to the product or the payment experience team.
Measuring the Actual Cost of Your Decline Rate
To make false decline impact visible, you need to estimate the revenue value of the transactions your risk system is blocking. This requires combining your declined transaction data with an estimate of the legitimate transaction rate within the declined population.
The precision of your model at a given score threshold tells you what fraction of the transactions you block are actually fraudulent. If your precision at your current threshold is 40%, you are blocking 60 legitimate transactions for every 40 fraudulent ones. The question is whether the value of preventing those 40 fraudulent transactions exceeds the value of the 60 legitimate transactions you blocked.
For a digital goods transaction with an average value of $45 and a fraud loss rate of 85% on confirmed fraud (the remaining 15% is sometimes recovered through dispute resolution), a fraudulent transaction costs you roughly $38. A declined legitimate transaction costs you the gross margin on that transaction, plus the customer lifetime value impact if the customer churns.
If your margin on a $45 transaction is $15, the immediate revenue cost of a false decline is $15 per declined legitimate transaction. Your threshold is net-positive only if the fraud loss you prevent exceeds the revenue you give up. At a precision of 40%, you prevent $38 x 40 = $1,520 in fraud loss while declining $15 x 60 = $900 in legitimate revenue, which looks positive. But if the declined customers also churn at a meaningful rate, the lifetime value of those 60 customers may substantially exceed the immediate revenue calculation.
Where the Threshold Sits Depends on the Product
There is no single optimal threshold that applies across products or customer segments. The calibration depends on the specific cost structure of your context.
For crypto on-ramps and high-value consumer lending, the fraud loss on a single approved transaction can be large relative to the margin, which shifts the threshold toward higher precision (more willing to block legitimate customers to avoid the occasional large fraudulent transaction). For consumer payments on low-margin digital goods, the math often runs the other way: a single fraud event is not catastrophic, but a sustained decline rate that drives customer churn is expensive over time.
Customer segment matters as much as product type. New customers with no transaction history on your platform look riskier to a behavioral model because the model has no baseline. Blanket caution with new customers based on account age produces a disproportionate false decline rate on exactly the population you most need to convert well: customers in the consideration phase who are evaluating whether to adopt your platform. A bad experience at first transaction is likely to be the last transaction.
High-value established customers should face almost no friction under most threshold configurations, because the model has strong behavioral evidence of legitimacy and the cost of misclassifying them is high. If your threshold is producing declines on established customers with two years of clean transaction history, that is a signal your threshold is too aggressive or your model is degrading relative to your current customer population.
The Appeal Rate as a Calibration Signal
Transaction appeal rates, the fraction of declined transactions where the customer contacts support and successfully appeals the decline, are one of the most direct feedback signals available for threshold calibration. A high appeal rate indicates that your threshold is blocking transactions that have legitimate characteristics; customers who appeal successfully and receive credit are by definition legitimate customers who were incorrectly blocked.
Tracking appeal rate by score band tells you which part of your score distribution is generating avoidable false declines. If the appeal success rate for transactions in the 60-75 score range (moderate-risk) is above 70%, you have a calibration problem in that band: you are blocking legitimate customers at a high rate, and the customers who make the effort to appeal can demonstrate their legitimacy. The question is how many customers in that score band you are losing because they do not appeal.
The conservative assumption is that for every one customer who appeals, two to five legitimate customers simply leave. The exact ratio varies and is hard to measure directly, but it should inform how you weight appeal rate as a calibration signal. A 20% appeal success rate in the 60-75 band suggests the true false positive rate in that band may be significantly higher than your model's precision estimate implies.
Graduated Friction Rather Than Binary Block
Binary block-or-approve is not the only option for transactions in the moderate-risk range. Step-up verification, challenge friction that is proportionate to the transaction value and risk signal, and manual review queuing for high-value borderline transactions are all mechanisms that reduce the effective false decline rate while managing fraud risk.
The cost of a step-up friction event (customer provides additional verification) is much lower than a hard decline for transactions where the customer is motivated to complete the transaction. A customer who is actually trying to make a legitimate high-value purchase will generally accept a modest additional verification step. A fraudster running automated card tests is more likely to abandon at a step-up challenge.
The failure mode to avoid with step-up friction is applying it too broadly to too many transactions, which trains customers to expect friction and increases the baseline abandonment rate. Step-up friction should be reserved for the score range where it provides genuine signal differentiation, not applied uniformly to everything above a moderate-risk threshold. The goal is to resolve uncertainty for transactions where the cost of the uncertainty is high enough to justify the verification overhead.
Knowing When You Are Optimizing for the Wrong Metric
Fraud rate and fraud loss are necessary metrics. They should not be the only metrics. If your quarterly review includes fraud loss but does not include an estimate of false decline revenue impact and customer churn attributable to declines, you are making threshold calibration decisions with an incomplete cost picture.
We are not saying fraud loss does not matter. It does, and letting fraud run unchecked has compounding costs through chargeback rates, card network penalties, and operational overhead in dispute processing. The point is that the precision-recall tradeoff has costs on both sides, and making an informed threshold decision requires estimating both sides with comparable rigor. Measuring one with accounting precision and the other with rough intuition is not a neutral choice; it systematically biases thresholds toward over-declining.
