Picture yourself as the chair of the pricing committee at a regional lender with a $1 billion auto loan portfolio. Captive lenders are pursuing near-prime lending, and your team is considering cutting interest rates across the near-prime tier to win it back. But charge-offs in that tier are already above target, and no one knows which borrowers are causing the losses or which ones are overpaying and leaving.
Most financial institutions and auto lenders use credit scores, pricing tiers, and rate sheets to set loan prices. This works well for borrowers in the middle of a tier, but problems show up at the edges. Each tier charges everyone its average expected credit risk, so pricing is off for those far from the average. With auto delinquencies still rising in early 2026, costs at both extremes keep going up.
The answer starts with measurement. This explanation covers where loan pricing fails, why margin compression is not the real issue, and how to measure the gap between expected and actual losses. In the end, the committee’s decision comes down to a simple pricing model calculation.
Loan pricing is how a lender sets a loan’s interest charged, fees, and terms. The rate has to cover four costs: funding costs, expected credit risk, servicing costs, and a target profitability margin. Most auto lenders estimate expected loss from a credit score, then sort borrower groups into pricing tiers using risk rating models.
Before funding a loan, three of the four costs are known in advance: the cost of funds comes from the balance sheet, servicing cost from operations, and the target margin from the board’s plan. Expected loss is different because it’s a forecast of default risk and the only borrower-specific factor that can be wrong. Any mistake in this forecast directly affects the rate and overall portfolio profitability.
In simple terms, a rate sheet turns a score into a price. If the score is right for the customer group, the rate covers the loss as planned. If the score is wrong, the rate will be off, and the error often goes unnoticed until there’s a loss or the borrower leaves for competition.
Tiers are used because it’s difficult to price every loan with a continuous score without loan origination software. Lenders split scores into bands, assign a rate to each, and let dealers and branches use that sheet. The rates on those tiers sit far apart. In Q2 2026, the average interest rate was 6.35% for new vehicles and 11.19% for used ones, and the credit spread widens sharply for borrowers below prime. The State of the Auto Finance Market, Q2 2026, provides a breakdown of rates by tier across various credit card loans and auto loan types.
Because the rates are so far apart, moving from one tier to the next can mean hundreds of basis points. If a group is placed on the wrong step, they end up paying too much or too little on every loan-level decision, and this mistake isn’t visible on the rate sheet.
Start by setting up the rate sheet as a model that lists all tiers, cutoff points, and the score linked to each cutoff. Most lenders have a rate sheet, but many don’t maintain a single document that shows which score version corresponds to each cutoff, or when each cutoff was last checked against actual losses. This document should become the standard for future comparisons.
The average rate for each tier is based on the group’s average characteristics, not on each individual. If the score misjudges a group, that group gets the wrong rate. If the score underestimates risk, the rate won’t cover real losses. If it overestimates risk, good borrowers pay too much interest and may leave for a better deal.
The error that people are watching out for is underpricing. When the score rates a group as average for its tier, but that group defaults faster than the tier average, the rate does not cover the difference. By the time the gap shows up in net charge-offs, every loan in that group is already on the books at the old rate. The latest data on cars conforms to this trend: in its Q2 2026 report, the New York Fed stated that the annualized flow of automobile balances into serious delinquency was 3.00%, up from 2.93% the previous year. The Fed’s economists also noted that new auto delinquency cases remain high, even though most other consumer credit products have remained stable over the two-year period.
Overpricing is the error nobody audits, because it never shows up as a charge-off. These borrowers either make their debt payment on time or choose not to take the loan. Instead, overpricing shows up as a lower look-to-book ratio among the applicants lenders want most—those whose credit records seem weaker than their payment history. A dashboard focused on losses won’t catch this lost revenue.
These errors build on each other. When good borrowers are charged too much, they switch to lenders with better pricing strategies and risk-based pricing. The borrowers who stay are, on average, the ones competitors would have charged more. Each quarter, the portfolio shifts more toward the risk the rate sheet underestimates, and raising rates for the whole tier only drives good borrowers away faster.
Old static scorecards make things worse because they’re built on a sample, checked against average tier behavior, and updated on a set schedule instead of when market conditions change. As the portfolio shifts, changes within a tier go unnoticed by pricing. Even making decisions faster doesn’t solve this. As the AI in lending article argues, speed just applies the same tier logic to more loans.
Over the past six quarters, collect predicted and actual losses by score band and show both in percentage points. Positive gaps reveal future net charge-offs that the rate doesn’t cover, while negative gaps show net interest income lost to competitors as borrowers leave. Neither appears separately in loss reports. Present both on a single page for the pricing committee.
Not at the system level. Credit union and bank net interest margins (NIMs) were wider in 2026 than a year earlier, while the flow of auto balances into serious delinquency rose. Now, the pressure on lenders is focused on certain groups and channels because prices and actual losses have moved apart.
Margins are actually wider, not tighter. Federally insured credit unions reported a net interest margin of 3.44% of average assets year-to-date through Q1 2026, up from 3.24% a year earlier. In Q2 2026, FDIC-insured banks had a net interest margin of 3.32%, 6 basis points higher than the same quarter last year. Neither figure describes a system being squeezed by funding costs or interest rate shifts.
The pressure is real, but it happens lower down in the system. During the same 12 months in which the flow of auto balances into serious delinquency rose, system margins widened, so the averages could absorb losses in specific groups with income earned elsewhere. While the top-line credit spread looks fine, a closer look by tier and channel shows that some groups may now have rates that don’t cover their losses.
If you see this as just general margin compression, you miss the real issue. A blanket rate cut to protect volume gives a discount to every borrower in the tier, including those already underpriced, so margin given up on good borrowers ends up funding bigger losses on risky ones, including those already underpriced. Since margin compression occurs at the segment level, the solution should be targeted at that level as well.
Before reviewing the rate sheet, break down the margin by channel and tier. For both direct and indirect loans, and for new and used vehicles, check the spread at the pricing decision point. If you find a specific group carrying the loss, reprice only that group. Don’t adjust other factors, since any cut there just reduces margin from healthier segments.
Compare your score’s prediction with the actual behavior of each population in percentage points; a positive gap means the rate does not cover the risk, whereas a negative gap indicates the population is priced too high. Rank the gaps by the product of their size and population, then apply your current pricing controls to the largest gap.
The measurement begins with the score the lender already runs. The dotData Signal Intelligence Platform converts that base score into a probability and then treats the difference between the predicted and actual results as the thing to explain. dotData Core, the platform’s engine, examines the lender’s associated tables—namely, applications, tradelines, inquiries, and performance—and ranks potential Driver Signals based on how well they account for errors in the existing score. The result is a ranked list of mispriced subpopulations.
Each finding is provided as a readable SQL rule, including information about the source tables, the time window, and the population concerned. The gaps are given in percentage points in both directions, so both the underpriced and the overpriced populations appear on the same list. The rules are incorporated into the pricing adjustments that a lender is already accustomed to running. dotData does not set the rate; the pricing engine and the credit committee still determine it.
As an illustrative example, in a synthetic auto-lending sample of 52,847 applications with a 10.87% charge-off rate, one Driver Signal identified applicants with 13 or more years between their earliest and latest tradeline. This group accounted for 15.8% of the total applications. Their actual charge-off rate was 5.96%, compared to the 8.68% predicted by the baseline score—a difference of −2.72 percentage points. The score imposed too heavy a penalty on those borrowers, so this is first and foremost a pricing issue rather than a credit management issue.
Illustrative example: Go back to the lender from the opening. Its auto book is $1 billion, and $200 million of it sits in the near-prime tier. The proposal on the table is to cut that tier’s rate by half a percentage point (50 basis points) to match what captive lenders are offering.
Day 1: price the proposal. Half a point off $200 million of loans means $1.0 million less interest income every year. Every borrower in the tier gets the discount — including borrowers who were already paying too little for the default risk they carry.
Days 2–4: look inside the tier. Before the vote, the risk team compares the losses predicted by the score with the losses that actually occurred. dotData Core ranks the groups where those two numbers differ most, and two groups stand out.
Group A ($40 million) is priced too high. The score expected 4.0% of these balances to be charged off each year, but only 2.5% were. These borrowers are paying for risk they don’t carry, making them the most likely to refinance with a cheaper lender.
Group B ($30 million) is priced too low. The score expected 4.0%, and 6.0% was charged off. That 2-point shortfall costs $600,000 a year that the rate never covered, and the tier-wide cut would add another $150,000 to it.
Day 5: decide. The committee drops the tier-wide cut and adjusts only the two groups. Group A’s rate drops by a full point (100 basis points), costing $400,000 per year. Group B’s rate goes up by 2 points (200 basis points), bringing in $600,000 a year and covering its shortfall.
| Plan | Change | Effect on annual income |
|---|---|---|
| Tier-wide cut | 0.5 points off all $200 million | −$1.0 million |
| Targeted fix | 1 point off Group A’s $40 million | −$400,000 |
| 2 points added to Group B’s $30 million | +$600,000 | |
| Net | +$200,000 |
Same tier, same week: the targeted fix comes out $1.2 million a year ahead.
Outside the example, the rule is the same. Use the ranking to decide which groups to examine, and as a rule of thumb, multiply the gap size by the group size to decide whether to act. A small gap across a large group can be worth more than a big gap in a small one.
Five steps turn this into a quarterly routine:
Each rate on the sheet is a statement about loss. If the score is accurate, the statement holds true, and the margin works as planned. But if the score is wrong, the error shows up in two ways: losses the rate didn’t cover, and good borrowers who paid too much before switching to a lender with a fairer rate.
Tier imperfections: Score-based pricing tiers risk mispricing borrowers who live at the edges of their tier. Borrowers who are underpriced are more likely to default, increasing the lender’s exposure, while overpriced borrowers are good borrowers who will either refinance for better rates or never close in the first place.
Segment-specific pressure: Broad net interest rate margins are healthy, indicating that pressure on portfolios is being driven by specific mispriced groups rather than through system-wide compression.
Gap measurement: Comparing predicted losses against actual losses can uncover variances that may be difficult to spot using traditional means and that tier average loss reports will simply not capture.
Target fixes vs. blanket changes: Adjusting rates for mispriced segments can preserve margins and generate a significantly healthier financial outcome than applying broad, tier-wide rate cuts.
Quarterly reviews: Auditing score predictions against actual performance allows lenders to monitor rate sheets and adjust them to accurately reflect borrowers’ true ability to pay, protecting long-term profitability and asset value.
Loan pricing is the process of setting a loan’s rate and terms. In the case of auto lending, the rate must account for the cost of funds, expected credit losses, servicing costs, and the target margin, with the expected loss component generally determined by linking credit scores to pricing tiers. The various components combine to determine the final rate through a pricing waterfall.
Most auto lenders assign pricing tiers to different credit scores. The various levels are quite distant: in Q2 2026, the average interest rate was 6.35% for new vehicles and 11.19% for used ones, and the spread widens sharply for borrowers below prime. With indirect loans, there is an additional stage: dealers can increase the lender’s buy rate before the contract is signed, thereby introducing a pricing level not set out in the rate sheet.
Loan pricing models convert the expected loss, funding cost, servicing cost, and target margin into a rate for each borrower or group. The expected loss is the vulnerable point, since the other factors like base rate and capital costs are known in advance; therefore, this analysis of auto loan pricing model optimization focuses on identifying the groups the model misprices.
Risk-based pricing sets the rate according to credit risk, usually through score-based tiers. It also triggers a disclosure rule. According to Regulation V, a creditor must provide notice of risk-based pricing when, based on a consumer report, it grants credit on terms that are materially less favorable than the best loan terms available to a substantial proportion of its customers. The purpose of this notice is to inform the borrower that the terms have been set based on the credit report.
Credit unions use the same system: score, tier, and rate. Their margins have room: federally insured credit unions reported a net interest margin of 3.44% of average assets year-to-date through Q1 2026. With this margin, targeted rate adjustments are a more cost-effective way to compete for volume than a general rate cut against captive lenders and banks.
Find out where your rate sheet is off. dotData helps auto lenders and credit unions identify the groups their scores misprice and provides each one as a rule the pricing committee can review. Contact the dotData lending team.
Auto loan balances reached $1.69 trillion in Q1 2026, and lenders originated $182.1 billion in…
Getting a real read on auto credit means pulling data from at least six places…
Key Takeaways The Net Charge-Off (NCO) landscape is evolving: Fitch's subprime auto ABS 60-day delinquency…
Key Takeaways Monitoring is where the loss is decided now: The score at origination was…
Key Takeaways The speed of approval does not equate to accuracy: In contrast to a…
Key Takeaways "Best" is misleading at best: An AI platform can top every review site…