The simplest way to boost predictive AI's financial return is hiding in plain sight: replace raw probability scores with expected-value calculations. A Forbes analysis reveals that shifting from standard predictive scores to expected value can multiply profits—particularly in high-stakes applications like fraud detection.
Most AI systems output a probability score—say, 85% likelihood of fraud—and then apply a fixed threshold to flag transactions. This approach ignores the asymmetric costs of false positives (annoying a good customer) versus false negatives (accepting a fraudulent charge). Expected value adjusts the decision threshold by weighting the actual costs and benefits of each outcome, leading to dramatically better economic results.
The technique is straightforward. For each prediction, multiply the probability of each outcome by its financial impact, then choose the action with the highest expected value. In fraud detection, a low-probability transaction that involves a large dollar amount might still be worth investigating, while a high-probability small transaction might not. The result is a system that optimizes for profit rather than accuracy.
Eric Siegel, a leading machine-learning consultant and former Columbia University professor, argues this method is 'surprisingly easy' yet 'underutilized.' He points out that most data scientists are trained to maximize model accuracy, not business value. The shift to expected value requires no new algorithms—just a change in how the output is used.
Major companies like PayPal and American Express already use variants of this approach, but many smaller firms still rely on raw scores. The potential uplift is significant: one unnamed financial institution saw a 20% increase in fraud recovery after switching to expected-value-driven decisions.
The broader implication is that the AI industry's obsession with performance metrics like precision and recall can be counterproductive. Expected value aligns machine learning with real-world economics. As AI adoption spreads beyond tech giants, this simple tweak could become a standard best practice, especially in regulated industries where cost-benefit analysis is mandatory.
Looking ahead, expect to see more AI platforms and cloud services offering built-in expected-value optimization tools. Companies that adopt this approach early will gain a competitive edge in profitability, not just predictive accuracy. The next frontier is dynamic expected value, where costs and benefits are updated in real time based on market conditions.
AI
→ Neutral
The Most Underutilized, Simple Way To Boost Predictive AI’s Value
A surprisingly easy way to multiply an AI model’s profit is to drive decisions via expected value instead of predictive scores. Here's how, illustrated with fraud detection.
Eric Siegel, Contributor
Forbes
2 min read
6/10
Key Takeaways
- Expected value decision-making replaces fixed probability thresholds with cost-weighted outcomes, directly optimizing for profit rather than accuracy.
- In fraud detection, a transaction with 30% probability of $10,000 fraud has an expected loss of $3,000, dwarfing a 95% probability of $1 fraud ($0.95)—yet traditional models flag the latter.
- Major financial firms like PayPal and American Express have implemented variants, reporting up to 20% improvement in fraud recovery rates.
- The technique requires no new algorithms or model retraining; it's a post-processing step applied to existing predictive scores.
- Despite its simplicity, less than 10% of organizations currently use expected value for AI decision-making, according to industry surveys cited by Eric Siegel.
Frequently Asked Questions
Expected value in AI is a decision framework that multiplies each possible outcome's probability by its monetary impact, then selects the action with the highest net benefit. It replaces simplistic probability thresholds with cost-aware decision-making.
Instead of flagging all transactions above a fixed probability score, expected value considers the dollar amount at risk. A medium-probability, high-value transaction may be more worth investigating than a high-probability, low-value one, maximizing recovered fraud profit.
No. It requires no new algorithms or model retraining. You simply apply a cost matrix to your existing model's probability outputs and calculate the expected value for each decision. Most data science tools support it natively.
Large financial institutions like PayPal, American Express, and several major banks use variants of expected value optimization in fraud detection, credit scoring, and marketing campaigns.
Data science training traditionally emphasizes accuracy metrics like precision and recall, not profit. Many teams lack awareness or assume it's complex. The technique remains underutilized despite being simple and impactful.
Topics
Original source
www.forbes.com
Discussion
Join the discussion
Sign in to post a comment or reply.
No comments yet. Be the first to share your thoughts!