In short: Pre-bid default prediction scores a lead's default risk from the limited "ping" data available at auction time, so lenders can bid what each lead is actually worth. The result is pricing risk instead of rejecting it: lower bids on risky leads, higher bids on good ones, and fewer losses from the winner's curse.
In modern lead marketplaces, lenders operate under a significant information asymmetry. Within milliseconds, they must decide whether to bid on a lead based on a highly redacted data set. While the lead seller sees the full application, the buyer sees only a subset, creating a systematic disadvantage where risk is often mispriced.
AI-powered pre-bid default prediction is designed to flip this dynamic. By modeling default risk directly from the available "ping" attributes, lenders can bid precisely what each lead is worth—no more, no less. This transition transforms lead acquisition from a game of chance into a precision science, ensuring sustainable margins in even the most competitive auctions.
The True Cost of Mispriced Risk
When lenders overpay for risky leads, the consequences ripple through the entire portfolio. This often manifests as "Adverse Selection"—the classic Winner's Curse, where the lender who wins the auction is frequently the one who most significantly overestimated the lead's value.
Overpaying leads to higher acquisition costs and increased default rates, both of which erode unit economics. Conversely, lenders who underbid out of caution lose valuable volume to more informed competitors. The goal is to find the "indifference price"—the exact bid where the expected profit from a funded loan perfectly balances the risk of default. That calculation is only as good as its labels; Predicting Funded Loans, Not Just Approvals covers how to choose and mature them.
Moving Beyond Rigid Credit Tiers
Traditional lead buying has long relied on blunt filters, such as "only buy leads with a 680+ FICO." This approach is flawed for two reasons. First, it ignores the "information loss" inherent in a single score; a 680 FICO borrower who has recently paid off significant debt is fundamentally different from one who has just opened several new lines of credit. Second, it creates "Competitive Compression," where every lender is fighting for the same narrow segments, driving up costs for everyone.
AI models solve this by utilizing the full feature space available at bid time. This includes application data like loan purpose and income, behavioral signals such as device type and session timing, and even geographic factors tied to regional economic indicators. Gradient-boosted models are well suited to this kind of mixed, tabular data; a large benchmark study found that tree-based models still outperform deep learning on typical tabular datasets.
Architecture for Sub-100ms Decisions
Effective pre-bid models must balance predictive power with extreme low latency. Because lead auctions typically require a response in under 100 milliseconds, there is no time for complex, multi-model ensembles.
The optimal architecture consists of a high-speed feature engineering layer and a optimized gradient boosting model, such as XGBoost or LightGBM. A calibration layer ensures that the predicted probabilities accurately reflect true default rates, while a final translation step converts that probability into an optimal bid price. This last step is crucial: a lead with a 15% predicted default rate isn't "bad" if the loan amount and interest rate provide enough margin to absorb the risk.
Large language models stay out of this sub-100ms path, but they help around it. After a lead is won, they can read uploaded bank statements and pay stubs to extract verified income and obligations, which become better labels for the next version of the pre-bid model. They should never produce the risk score itself: lenders must be able to state the specific principal reasons for an adverse action under Regulation B, whatever kind of model made the decision. The CFPB issued guidance making that point about complex algorithms in 2022 and 2023, then withdrew it in May 2025; the regulation itself did not change. See LLM Model Risk and Compliance in Lending.
Integration with Dynamic Pricing
By integrating default prediction with dynamic pricing systems, lenders can move away from fixed bids and toward a value-based strategy:
Expected Value = (1 − P(default)) × Expected Profit − P(default) × Expected Loss
Maximum Bid = Expected Value × (1 − Target Margin)
This formula allows the system to adjust automatically as market conditions shift. If default rates in a specific segment rise, bids decrease instantly without manual intervention. Conversely, if a borrower profile shows improved performance, the system increases bids to capture more of that high-value volume.
Pricing, Not Rejecting
The typical outcome of moving to pre-bid default scoring is not a higher rejection rate. Lenders keep their funded volume while early defaults fall, because they stop paying top-tier prices for bottom-tier risk. The key insight is that they aren't rejecting more leads; they are simply pricing them correctly. Leads that were previously bought at a "flat" tier-based price were now purchased at risk-adjusted levels that protected the lender's margin.
This level of precision also enables more sophisticated portfolio management. Lenders can set dynamic concentration limits to avoid over-exposure to specific risk factors and adjust bidding intensity based on real-time funding capacity. As we've discussed in our overview of AI for lending, the ultimate winners in this space are those who treat every lead as a portfolio optimization problem rather than a one-off transaction.
Frequently Asked Questions
What is pre-bid default prediction?
It is a model that estimates a lead's probability of default using only the data available in the auction ping, before the lender bids, so the bid can reflect risk-adjusted value.
How should a lender set a bid from a default probability?
Compute expected value as the profit if the loan performs, weighted by the probability it does, minus the expected loss weighted by the default probability. Then bid below that by your target margin.
Which models work best for default prediction in real time?
Calibrated gradient-boosting models such as XGBoost or LightGBM are the usual choice: accurate on tabular data, fast enough for sub-100ms responses, and explainable with tools like SHAP.
Ready to price leads based on true default risk? Contact Plato AI to implement pre-bid scoring that optimizes every acquisition decision in real-time.