Mortgage and Structured Finance Markets Archives

Models & Markets Update: July 2026

Key Takeaways

Prepayment models remain reliable, but the fast bias showing up in in-the-money cohorts is worth watching if rates keep climbing and more loans move into the money

Non-QM credit risk is becoming a bigger focus right as CM 7.1 goes live — the model is arriving just as delinquency trends turn less benign

The market has flipped from pricing rate cuts to pricing a hike this year — a real shift in the rate outlook, not noise

Elevated rates and mortgage costs look like a multi-year reality rather than a temporary plateau, with labor data adding a note of underlying softness beneath a falling headline unemployment rate

This month’s call covered prepayment model back-testing, a deep dive on conventional model performance, the growing Non-QM universe, the production release of RiskSpan’s Non-QM Credit Model (CM 7.1), and the latest macro picture.

You can read the recap below or click here for the entire recording.

Prepayment Model Back-Testing

The prepayment model continues to closely track realized speeds across Fannie, Freddie, and Ginnie collateral. Discount coupons (WAC 5.5 and below) remain structurally low and range-bound at roughly 4–8 CPR, consistent with these loans sitting out of the money in a turnover-driven regime. Premium coupons (WAC 6 and above) eased slightly in June, as a 13 bp increase in the driving mortgage rate outweighed a 5% rise in day count — visible in a modest month-over-month flattening of the conventional S-curve. FHA/VA aging-ramp peaks have also come down from prior levels, with VA loans continuing to ramp faster and higher than FHA.

Conventional S-Curve Flattening MoM

FHA/VA Aging-Ramp Peak Declined

Conventional Deep Dive

A closer look at FH30 and FN30 performance found the model running fast versus actuals in June — a bias of roughly 1 CPR — with month-over-month direction captured well but magnitude off: the model overreacts in out-of-the-money cohorts and underreacts in in-the-money cohorts, where its error is markedly larger (~2 CPR vs. ~0.8 CPR).

Curtailment and buyout activity move in opposite directions over a loan’s life. Curtailments are front-loaded, peaking in the newest vintages and fading with age, while buyouts build slowly and grow as loans season and credit stress accumulates. The effect is large: comparing 2023 to 2025 vintages in higher-coupon cohorts, Freddie Mac loans show roughly 7x the buyout activity, and Fannie Mae loans show roughly 3x.

Non-QM Universe and Credit Model 7.1

Non-QM issuance is tracking toward roughly $100B for 2026 — Q2 came in around $27B, down 16% quarter-over-quarter but up 58% year-over-year — with originations running at about $170–180B. Delinquency data gave a mildly encouraging signal, with the 2023 vintage appearing to plateau around 24 months of loan age, though risk stays concentrated in bank-statement and self-employed borrowers and lower FICO bands. Zooming out, delinquency has been climbing across consumer credit broadly (subprime auto, personal loans, mortgage, Non-QM) since 2021.

Non-QM RMBS Issuance and 2026 Trajectory

Consumer Credit Delinquency by Product Type

Against that backdrop, RiskSpan’s Non-QM Credit Model — CM 7.1 — has moved into production and is now available to all clients. It runs four independent transition models by documentation type (Bank Statement, DSCR, Full Doc, Other), feeding into a single liquidation-timeline and severity model recalibrated to current Non-QM performance.

CM 7.1 Model Structure

Macroeconomic Update

The biggest change this month is at the Fed. A couple of months ago the market was pricing in up to two rate cuts this year; CME FedWatch data now show meaningful odds of a hike instead, with the Fed funds rate seen moving from 350–375 bps toward 375–400 bps by year-end.

Fed Funds Rate to Remain High in 2026

Longer rates have followed suit. The 10-year Treasury, just above 4.5%, is expected to climb another 20–25 bps toward ~4.9% before easing — but to hold above 4% for a couple of years. Mortgage rates, currently around 6.5–6.6%, are expected to reach as high as ~6.9% by spring 2027 and stay above 6.5% for the next two to three years.

The labor market sent a mixed signal. June payrolls rose just 57K, roughly half of what was expected, and April/May were revised down a combined 75,000 jobs — pointing to real underlying softness. Headline unemployment actually ticked down to 4.2%, but that came from falling labor force participation (61.5%, the lowest since 2021) rather than hiring; the household survey showed 507,000 fewer people employed. On a brighter note, CPI eased to 3.5% year-over-year and wage growth matched it exactly — the first time in 12–18 months that real wage growth hasn’t been negative.

Inflation and Unemployment

We continue to add additional analytics reports on the RiskSpan Platform. Please visit www.riskspan.com to request access.

As always, please feel free to contact us to discuss or learn more.

I Was the Cron Job: What a 1990s Back Office Taught Me About AI Token Costs

My first Wall Street job as a junior programmer was to process pool tapes from the GSEs and Ginnie Mae. Every month, the tapes arrived and an operator loaded them. Then I ran through a series of scripts — the same series, in the same order — to produce prepayment reports. The reports were printed, photocopied, stapled. Every desk got its copy before the open. For this, the firm paid me slightly more than the prevailing minimum wage.

This went on for a couple of years. Eventually two things happened: I wanted to get paid more, and management decided I could do more useful things — like help build prepayment and OAS models. My boss’s solution was not to hire another me. It was to automate the whole process with proper error handling and hand it off to the operator who was already running other overnight jobs.

Looking back, the strange part isn’t that the job got automated. It’s that for two years, nobody thought it was odd to pay a human to be a scheduler. The job was never “Suhrud runs the reports.” The job was “the reports get run.” The moment someone articulated it that way, the answer was obvious.

I’ve been thinking about this a lot lately, because the industry is busy making the same mistake with a much more expensive employee.

The rerun

Today we wire LLMs into our systems through connectors and MCP servers, and then we ask them to perform recurring tasks. Pull the latest remittance data. Run the QC checks. Format the surveillance report. Post it to the channel. Every day. Sometimes every hour.

Here’s what actually happens on each run: the model re-reads its instructions, re-loads its tool definitions, rediscovers the schema, re-reasons through steps it reasoned through yesterday, and re-derives the entire task from first principles. It is me in that back office — except the photocopier bills by the token, and the junior programmer occasionally hallucinates a column mapping.

That is the junior programmer problem: an LLM is a bright, expensive generalist being asked to do clerical work on repeat. It can help figure out the process, draft the checklist, and handle exceptions. But if the job is the same every morning, the goal should be to convert the junior programmer’s reasoning into reviewed code and let the system run it. The junior programmer belongs at the design table and the exception desk — not standing at the photocopier forever.

Two problems compound here. One is cost. The other, worse one, is that the junior programmer is non-deterministic.

The economics

Take a realistic daily agentic job: a report run against a data warehouse via MCP connectors, with QC checks and a written summary.

Each turn carries a base context of roughly 20K tokens — system prompt, tool definitions, instructions. The agent makes maybe 15 tool calls (list the tables, query, check, query again), each returning a few thousand tokens of results. Because every turn re-sends the growing conversation, cumulative input compounds to roughly 800K–850K input tokens per run, plus a few thousand tokens of output.

At current mid-tier model pricing (about $3 per million input tokens, $15 per million output), that’s roughly $2.50 to $2.75 per run. Daily across 22 business days: ~$55 to $60 a month. For one report. Run thirty recurring jobs like that — not an unusual number for a surveillance or analytics desk — and you’re around $1,700 a month for work that is identical every day.

Now the honest version, because a sophisticated reader will raise it immediately: prompt caching cuts repeated input cost by about 90%, and batch APIs knock another 50% off. A well-engineered agent run might land closer to $0.30–0.50, or roughly $10 a month per job. That’s real, and if you’re going to keep agents in the loop, you should absolutely use both.

But compare the alternative. A scheduled script on Cloudflare Workers runs free or nearly so. AWS Lambda executes this class of job for well under a dollar a month. And the one-time cost of having the LLM write that script? Maybe $1–5 of tokens. Even against the fully optimized agent, the deployed script is one hundred to a thousand times cheaper than the agent — and that gap is before you price the risk.

The overconfidence risk (AI hallucination) is worse than the bill

A deployed script is deterministic and auditable. Same inputs, same outputs. It’s versioned, diffable, and reviewable. When it breaks, it breaks loudly and identically.

An LLM performing the task fresh each morning is a new draw from a distribution. It will usually do the right thing. Occasionally it will skip a QC check, misread a schema change as an instruction, or format a number differently on run 47 — silently. In regulated finance, “the agent usually gets it right” is not a control. Anyone who has sat through a model governance review knows that the question isn’t whether the process works; it’s whether you can prove what it did, every time.

To be fair, this argument cuts both ways, and pretending otherwise would be dishonest. LLM-generated code can also be wrong — subtly, in edge cases, in ways that survive a casual glance. The principle isn’t “code good, agent bad.” The principle is review once, run many. A script gets one round of concentrated human scrutiny, and that scrutiny amortizes across thousands of executions. An agent-in-the-loop asks you to trust a fresh, unreviewed performance every single run. My boss didn’t just automate my job — he automated it with error handling, and someone reviewed that logic before the operator ever touched it. That review was the whole point.

Use the LLM for Judgment, Not Execution

The obvious objection is that nobody serious should run a fully deterministic job through an agent loop. Fair. But many tasks are not purely deterministic. They look repetitive on the surface while still containing moments of judgment: triaging anomalies, deciding whether a failed reconciliation is a data problem or a real signal, or writing the two sentences of commentary that make a report useful. Those are legitimate places for an LLM.

But that does not make the LLM the operator.

The right pattern is LLM as designer, author, and escalation point — not as the thing that runs the job every day. The operating layer should be code: scheduled, versioned, reviewed, monitored, and boring. Code should extract the data, validate it, compute the results, apply the rules, and distribute the output. The LLM should be called only when the process reaches a judgment node: when a check fails, when a number looks strange, when the rules do not cover the case, or when a human-quality narrative is actually needed.

That is the same pattern my boss landed on in the 1990s. The operator ran the job. The human did not sit there rerunning the same steps every morning; the human got called when something broke or required judgment. The modern equivalent is not “replace the operator with an LLM.” It is “let code be the operator, and let the LLM help design the system, explain exceptions, and handle ambiguity.”

This is also where the tooling ecosystem is already headed. The model vendors themselves are pushing patterns where the model writes and executes code rather than acting as a chatty runtime that makes dozens of repeated tool calls. That is not a rejection of LLMs. It is a more disciplined role for them: use the model to create and improve the machinery, not to impersonate the machinery every time it runs.

Where to put the production version

Once the LLM has written your pipeline and a human has reviewed it, deployment is genuinely easy now:

Cloudflare Workers + Cron Triggers. Zero infrastructure, generous free tier, ideal for API-glue jobs — fetch, transform, post. If your job is stitching services together, start here.

Railway. Dead simple for Python jobs with real dependencies — pandas, database drivers, the works. Push a repo, set a schedule, done. Roughly $5 a month.

AWS Lambda + EventBridge. Pennies per run, and the path of least resistance if your firm already lives in AWS and your security team wants everything inside the VPC.

GitHub Actions scheduled workflows. Underrated. If the code already lives in a repo, a cron-scheduled workflow is free for many light jobs and keeps the pipeline next to its own version history.

One honest caveat: deployed code is not free either. Schemas drift, credentials rotate, upstream APIs change, and someone owns the pager. But maintenance cost is bounded and occasional. Agent cost is unbounded and per-run. That asymmetry is the whole argument.

A simple approach

If you can write down the steps, write code — or better, have the LLM write it, then review it once and deploy it. If the task genuinely requires judgment on every run, keep the LLM in the loop — but only at the judgment node, not driving the whole pipeline.

Thirty years ago, my boss promoted me out of a job by automating it, and the firm got both a cheaper process and a modeler out of the deal. Do the same for your LLM. Let it write the script. Let the script run. Save the tokens — and the trust — for the work that actually needs a mind.

Models & Markets Update: June 2026

Key Takeaways

Prepayment model continues to perform well; discount coupon speeds (WAC 5.5 and below) remain stable across Fannie/Freddie and GNMA, driven by housing turnover

Premium coupon speeds (WAC 6.0+) declined as elevated rates reduced refi incentive; UPB rising steeply on new gross issuance of low-seasoning loans

FN/FH S-curve has flattened materially from April to June; GNMA saw similar compression but at tighter intervals and with less magnitude

Non-QM prepay speeds declined across all doc types in May remits; Non-QM DQ60+ rates ticked down marginally

Non-QM Credit Model CM 7.1 in beta; available July 15; dedicated webinar coming before release

Fed held rates at June 17–18 meeting; year-end expectation shifted to 375–400 bps; meaningful probability of a hike as early as September; dot plot revised to 3.8% for 2026

Mortgage rates near 6.5%, back to summer 2025 levels; 6% viewed as 2026 floor; 10-yr TSY expected above 4% for years

CPI at 3.8% YoY, real wage growth -0.4% (first negative since 2022); U-6 underemployment at 8.1%

Home prices near-flat nationally (+0.67% YoY); inventory-constrained, not demand-driven; significant geographic divergence

You can read the recap below or click here for the entire recording.

Prepayment Model Back-Testing: June Factor Data Update

The prepayment model continues to track realized speeds closely across Agency collateral (results available in Edge under Vertex). The June factor data captures May remittances.

Fannie/Freddie — Discount Coupons (WAC 5.5 and Below)

Model CPR closely tracks observed CPR, confirming the model is well-calibrated in the turnover regime. With minimal refi incentive at current rate levels, prepayment activity is driven almost entirely by housing turnover, which has held steady.

Figure 1: FN/FH Discount Coupon Back-Testing — Model CPR vs. Observed CPR

Fannie/Freddie — Premium Coupons (WAC 6.0 and Higher) and S-Curve

Premium speeds declined due to elevated mortgage rates and fewer day counts in May vs. April. UPB is also rising steeply, reflecting gross issuance of newly originated loans with limited seasoning that further dampens prepay. The S-curve flattening tells the fuller story: comparing April, May, and June factor months, the curve has collapsed from the April peak (green), with borrowers responding progressively less to the same level of refi incentive.

Figure 2: FN/FH Premium Coupon Back-Testing — Model CPR vs. Observed CPR

Figure 3: FN/FH S-Curve Flattening — April, May, and June Factor Months

GNMA — Discount and Premium Coupons

GNMA discount speeds remain supported by turnover. Premium speeds are declining but proved more resilient than conventional — approximately 14% reduction vs. 24% for Fannie/Freddie. The GNMA S-curve is compressing, but at tighter intervals than the more pronounced slope collapse seen in conventional.

Figure 4: GN/G2 Discount and Premium Coupon Back-Testing — Model CPR vs. Observed CPR

Non-QM Historical Performance

Based on Cotality data through the Edge platform’s historical performance module, capturing the June factor date (May remittances).

Prepayment Speeds

Speeds declined across all three major doc types reflecting the rate environment: bank statement 29→22 CPR, DSCR 21→16 CPR, full doc 17→16 CPR. Full doc loans carry a lower WAC (~5.1% vs. 6.9–7.1% for bank statement and DSCR) and are more locked in, though they show a steeper S-curve response when refi incentive is held constant. Many DSCR loans remain within prepayment penalty terms, though 2023 originations with three-year terms are now beginning to exit that window, which could lift speeds going forward.

Figure 5: Non-QM CPR by Documentation Type (July 2023–June 2026)

Delinquencies

DQ60+ rates ticked down marginally in May remits: bank statement ~4%, DSCR ~3%, full doc ~0.72%. A wide gap persists across doc types even after controlling for FICO and LTV. The 2023 vintage remains the highest-delinquency cohort, driven by somewhat looser underwriting and a credit burnout effect — stronger borrowers in that high-WAC vintage have already paid down or refinanced, leaving a residual pool more likely to be credit-impaired.

Figure 6: Non-QM DQ60+ by Documentation Type (July 2023–June 2026)

Figure 7: Non-QM DQ60+ by Vintage and Loan Age

Non-QM Credit Model: CM 7.1 Update

CM 7.1 is in beta testing and will be generally available on July 15, 2026. A dedicated webinar is planned before the production release. The model uses a three-stage architecture with four independently estimated transition models (Bank Statement, DSCR, Full Doc, Other) feeding a unified liquidation timeline and severity model — capturing the meaningfully different performance characteristics across Non-QM documentation types seen in the historical data above.

Figure 8: CM 7.1 Model Structure

Macroeconomic Update: June 2026

Federal Reserve — On Hold, With a Hike Now on the Table

The Fed held rates at its June 17–18 meeting (350–375 bps). The big shift is in forward expectations: comparing CME FedWatch probabilities from May 20 to June 17, the likelihood of a rate hike has risen materially, with a meaningful probability of an increase as early as September. Year-end 2026 market expectation has shifted from 350–375 to 375–400 bps. The Fed’s own June dot plot revised the 2026 median funds rate to 3.8%, up from 3.4% in March, as inflation surprised to the upside and strong payrolls data reduce pressure to ease.

Figure 9: CME FedWatch Conditional Probabilities (May 20 vs. June 17) and Fed Dot Plot

Rates, Inflation, and Home Prices

Treasury and mortgage rates: The 10-year TSY consensus forecast peaks near 4.6% by year-end 2026 and stays above 4% for the next several years. Mortgage rates have been notably volatile since late February, climbing from ~6% back to ~6.5% as of mid-June. The team views 6% as the effective 2026 floor.

Inflation and labor: CPI at 3.8% YoY vs. wage growth of 3.4% produces real wage growth of -0.4% — the first negative year since 2022. U-3 unemployment is 4.3%; U-6 underemployment is 8.1%. Labor force participation is at its lowest since 2021. May payrolls came in at 172K vs. 80K expected but gains were concentrated in leisure/hospitality and local government. The Fed’s June projections raised 2026 PCE inflation to 3.6% (from 2.7% in March).

Home prices: Case-Shiller National at +0.67% YoY (March 2026); 20-City Composite at +0.83%. Gains appear to be inventory-constrained rather than demand-driven. Geographic divergence is significant — some MSAs are in negative territory, which is a relevant risk factor for Non-QM collateral concentration.

Figure 10: 10-Year Treasury Consensus Forecast and Mortgage Rate Trend

Figure 11: Inflation, Wage Growth, and Labor Market Dashboard

Figure 12: Case-Shiller National and 20-City Composite Home Price Indices

Summary

Topic	Key Takeaway
Prepayment Model	Performing well; discount speeds stable (turnover-driven); premium speeds declined on elevated rates and fewer May day counts; UPB rising steeply on new gross issuance
FN/FH S-Curve	Materially flatter from April to June; borrowers responding less to refi incentive; newer vintages pulling down aggregate
GNMA Performance	Discount speeds stable; premium speeds down but more resilient than conventional (14% reduction vs. 24% for FN/FH); S-curve compression at tighter intervals
Non-QM Prepay	Bank statement 29→22 CPR; DSCR 21→16 CPR; Full doc 17→16 CPR; rate-driven; DSCR partly shielded by prepay penalty terms
Non-QM DQ	DQ60+ ticked down marginally: bank statement ~4%, DSCR ~3%, full doc ~0.72%; 2023 vintage remains highest-DQ cohort (loose underwriting + credit burnout)
CM 7.1 Credit Model	Beta testing now; available July 15; dedicated webinar before release; four doc-type transition models feeding unified liquidation and severity model
Fed Policy	Held at 350–375 bps (June 17–18 meeting); year-end expectation shifted to 375–400; hike possible as early as September; dot plot revised to 3.8% median for 2026 vs. 3.4% in March
Rates	10-yr TSY consensus peaks ~4.6% year-end, stays above 4% for years; mortgage rates ~6.5%, back to summer 2025 levels; 6% viewed as 2026 floor
Inflation & Labor	CPI 3.8% YoY; wage growth 3.4%; real wages -0.4% (first negative since 2022); U-3 4.3%, U-6 8.1%; May payrolls 172K vs. 80K expected
Home Prices	Case-Shiller National +0.67% YoY (March 2026); gains inventory-driven, not demand-driven; significant geographic variation with some MSAs negative

We continue to add additional analytics reports on the RiskSpan Platform. Please visit www.riskspan.com/request-access to request free access.

As always, please feel free to contact us to discuss or learn more.

Models & Markets Update: May 2026

Key Takeaways

Prepayment models continue to perform well; April discount coupon speeds remain stable, driven primarily by housing turnover

Premium coupon speeds fell sharply in May factor data as the March mortgage rate sell-off reduced refinancing incentive and flattened the S-curve

GNMA premium coupons showed a moderate speed decline but proved more resilient than conventional counterparts

Non-QM Credit Model CM 7.1 enters beta testing in June and is targeted for production release around July 10th; a dedicated webinar is planned for the second half of June

No Federal Reserve rate cuts are expected in 2026; CME FedWatch now shows a meaningful probability of a rate hike later this year or in early 2027

Mortgage rates have climbed back to levels last seen in summer 2025, with the Freddie Mac survey rate at 6.51% and Mortgage News Daily at 6.65%

The 10-year Treasury rose approximately 50 basis points since February; market consensus expects it to reach ~4.8% by year-end and remain above 4% for the next 2–3 years

CPI inflation rose to 3.8% year-over-year in April; core CPI at 2.8% — both well above the Fed’s 2% target

Home prices stagnating nationally (~0.67% YoY per Case-Shiller); San Francisco has turned negative while New York continues to grow at ~4.8%

Consumer stress deepens: low- and middle-income households carry credit card debt roughly 3x their monthly spending at ~23% APR; buy now, pay later obligations add further hidden risk not captured in credit bureau data

You can read the recap below or click here for the entire recording.

Prepayment Model Back-Testing: May Factor Data Update

The prepayment model continues to track realized speeds closely across Agency collateral. Results are available on the Edge platform under the Vertex module.

Fannie/Freddie — Discount Coupons (WAC 5.5 and Below)

April speeds for Fannie/Freddie discount coupons remained relatively stable. Because these lower-coupon loans carry little to no refinancing incentive, prepayment activity is driven almost entirely by housing turnover, which has held steady.

Figure 1: FN/FH Discount Coupon Back-Testing — Model CPR vs. Observed CPR

Fannie/Freddie — Premium Coupons (WAC 6.0 and Higher)

Premium coupon speeds fell sharply in the May factor data, reflecting the March mortgage rate sell-off. Rising rates reduced refinancing incentive and caused a notable flattening of the S-curve. The May S-curve sits meaningfully below both the April curve and the long-run historical average (January 2014–May 2026), with the gap widening at higher incentive levels. A diminishing media effect in the May cohort contributed to the flatter shape.

Figure 2: FN/FH Premium Coupon Back-Testing — Model CPR vs. Observed CPR

Figure 3: EDGE Historical Performance — FN/FH S-Curve (Refi Incentive vs. CPR)

GNMA — Discount and Premium Coupons

GNMA collateral showed a similar pattern. Discount coupon speeds remained supported by turnover activity, while premium coupon speeds saw a moderate decline consistent with the higher-rate environment. Notably, GNMA premiums proved more resilient than their conventional counterparts when compared against the conventional S-curve, reflecting structural differences in the GNMA borrower population.

Figure 4: GN/G2 Discount and Premium Coupon Back-Testing — Model CPR vs. Observed CPR

Non-QM Credit Model: CM 7.1 Update

CM 7.1, RiskSpan’s new Non-QM Credit Model, is on track to enter beta testing in June with a targeted production release around July 10, 2026. A detailed webinar covering the model will be held in the second half of June.

Model Structure

CM 7.1 uses the same three-stage architecture as RiskSpan’s agency credit model:

Transition Models (four, one per documentation type) — each independently estimated

Bank Statement

DSCR

Full Doc

Other

Liquidation Timeline Model — applied once a loan enters default

Severity Model — estimates final losses on the defaulted balance

Each documentation type is modeled independently at the transition stage, then fed into a unified liquidation timeline and severity model. This segmentation reflects meaningfully different performance characteristics across Non-QM documentation types.

Figure 5: CM 7.1 Model Structure — Four Transition Models Feed into Unified Liquidation and Severity Models

Macroeconomic Update: May 2026

Federal Reserve — No Rate Cuts Expected; Hike Risk Emerging

The Fed funds rate remains at 350–375 bps. CME FedWatch futures indicate it is highly unlikely to be cut in 2026. More notably, the conditional probabilities have shifted over the past 4–6 weeks to reflect a meaningful likelihood of a rate hike in the latter part of 2026 or early 2027. With persistent inflation and a new Fed chair, the market sees little room for easing.

Figure 6: Federal Funds Target Range — Upper Limit (Source: FRED) and CME FedWatch Conditional Probabilities

Treasury Yield Curve — Significantly Higher Than February

The Treasury yield curve has shifted materially upward since the February 2026 trough, when the 10-year yield was at its lowest recent level and mortgage rates briefly approached 6%. Since then:

The 10-year and 30-year Treasuries rose approximately 50 basis points

The 2-year Treasury rose approximately 75–80 basis points

As of May 21, the 10-year Treasury stood at approximately 4.60%

This move is attributed to geopolitical dynamics (including the situation around Iran) and a declining global appetite for U.S. Treasuries, with recent auctions clearing at progressively higher yields. Market consensus projects the 10-year to reach approximately 4.8% by December 2026 and to remain above 4% for the next 2–3 years.

Figure 7: Treasury Yield Curves — January through May 2026

Mortgage Rates — Back to Summer 2025 Levels

Mortgage rates have given back much of the progress made earlier in the year. As of the call date, the Freddie Mac primary survey rate was 6.51% and the Mortgage News Daily rate was 6.65% — levels last seen in August 2025. The expectation is that mortgage rates will remain at or above 6.25% for the foreseeable future, with a sub-6% rate considered unlikely in the near term.

Figure 8: 10-Year Treasury Yield Forecast and Primary Mortgage Rate Trend (Mortgage News Daily, MBA, Freddie Mac)

Inflation — Staying Elevated

The April 2026 CPI print came in at 3.8% year-over-year; core CPI (excluding food and energy) ran at 2.8%, well above the Fed’s 2% target. The PCE index was not yet published at call time but was expected to confirm continued inflationary pressure. Combined with stable unemployment, this leaves the Fed with limited flexibility to ease.

Figure 9: PCE Inflation (ex. Food & Energy) and CPI — Year-over-Year % Change

Home Prices — Stagnating, with Pronounced Regional Variation

National home price growth has slowed to near-zero. The Case-Shiller National Index showed approximately +0.67% year-over-year as of February 2026, while the 10-City Composite came in at +1.74%, suggesting urban markets are modestly outperforming. Regional divergence is pronounced: San Francisco has recorded negative price growth for approximately the past six months (-0.34% YoY), while New York remains solidly positive at +4.78%.

Figure 10: Case-Shiller National and 10-City Composite Home Price Indices — Year-over-Year % Change

Figure 11: Case-Shiller San Francisco and New York Home Price Indices — Year-over-Year % Change

Consumer Stress: Evidence from Credit Card Spending

This month’s call featured a deep dive into consumer financial stress, drawing on research from the Federal Reserve Bank of Boston. The analysis is particularly relevant to mortgage credit risk given evidence of rising delinquencies in FHA and Non-QM collateral.

Credit Landscape

Agency loans (excluding FHA) continue to show low delinquency rates with no significant deterioration. FHA, however, is exhibiting elevated delinquencies that remain high even after accounting for the trial modification policy introduced in October 2025, which holds more loans in delinquent states during the modification process. The Non-QM universe has also begun to show a rising delinquency trend over the past 12–18 months.

Boston Fed Analysis: Spending and Debt by Income Group

A study from the Federal Reserve Bank of Boston segments credit card behavior across three income cohorts and reveals a striking disparity between spending levels and outstanding balances:

Low-income ($0–$39K): monthly spending ~$25B vs. revolving debt ~$80B — a 3x ratio; this group is primarily revolving rather than paying off balances

Middle-income ($59K–$83K): spending ~$37B vs. debt ~$105B — also approximately a 3x ratio

High-income ($121K+): spending ~$170B vs. debt ~$185B — roughly 1:1, consistent with transactor behavior (spend and pay off monthly)

Low- and middle-income households are therefore carrying roughly three months’ worth of spending as permanent revolving debt, at credit card APRs recently running around 23% on new issuances. Total credit card outstanding nationally has reached approximately $1.25 trillion.

Figure 12: Aggregate Credit Card Spending by Income Group, January 2015–May 2025 (Source: Boston Fed / Federal Reserve Y-14M)

Figure 13: Aggregate Credit Card Debt by Income Group, January 2015–April 2025 (Source: Boston Fed / Federal Reserve Y-14M)

Buy Now, Pay Later: An Untracked Risk

Buy now, pay later (BNPL) services have grown rapidly and appear to be used disproportionately by lower-income households. Because BNPL obligations are not reported to credit bureaus, they represent an invisible liability not reflected in standard debt figures. The team flagged this as a developing risk to monitor, particularly for its potential impact on borrower liquidity and mortgage performance in the FHA universe.

Summary

Topic	Key Takeaway
Prepayment Model	Performing well overall; April discount speeds stable (turnover-driven); May premium speeds fell sharply on March rate sell-off and S-curve flattening
GNMA Performance	Discount speeds supported by turnover; premium speeds declined moderately but more resilient than conventional counterparts
Non-QM Credit Model	CM 7.1 beta in June; production release ~July 10; dedicated webinar in second half of June
Mortgage Rates	Freddie Mac at 6.51%; Mortgage News Daily at 6.65%; back to summer 2025 levels; sub-6.25% rate unlikely near-term
Fed Policy	No cuts expected in 2026; CME FedWatch shows meaningful probability of hike later in 2026 or early 2027
Treasury Yields	10-year up ~50 bps since February; 2-year up ~75–80 bps; consensus at ~4.82% by year-end
Inflation	CPI 3.8% YoY (April); core CPI 2.8%; PCE similarly elevated; well above 2% target
Home Prices	National ~+0.67% YoY; San Francisco negative; New York +4.78%; highly geography-dependent
Consumer Stress	Low/mid-income households revolving 3x monthly spending at ~23% APR; BNPL obligations add hidden risk; FHA and Non-QM delinquencies trending higher

We continue to add additional analytics reports on the RiskSpan Platform. Please visit www.riskspan.com/request-access to request free access.

As always, please feel free to contact us to discuss or learn more.

Models & Markets Update: April 2026

Key Takeaways

Prepayment models continue to perform well, with March speeds driven by a February rate rally and day count effects

A new Non-QM Credit Model (CM 7.1) is on track for release near end of Q2 2026, with a dedicated webinar planned for end of May or early June

Housing turnover analysis reveals rate sensitivity at positive refinancing incentive levels — a finding that will inform the next prepayment model

Mortgage rates hit a six-month high in March before pulling back; rates are expected to remain above 6% through 2026 and 2027

No Federal Reserve rate cuts are expected in 2026; the consumer remains under pressure from elevated rates and rising credit card balances

You can read the recap below or click here for the entire recording.

Prepayment Model Back-Testing: April Factor Data Update

The prepayment model continues to track realized speeds closely across Agency collateral. Results are available on the Edge platform under the Vertex module.

Fannie/Freddie — Discount Coupons (WAC 5.5 and Below)

Discount coupons showed a modest uptick in March speeds, driven primarily by two factors:

A day count effect: March had three more collection days than February

Seasonal turnover patterns typical of the spring housing market

Figure 1: FN/FH Discount Coupon Back-Testing — Model CPR vs. Observed CPR

Fannie/Freddie — Premium Coupons (WAC 6.0 and Higher)

Premium coupons saw a sharper increase in prepayment speeds in March, driven primarily by the rates rally in February. With rates subsequently moving higher in March, May factor data is expected to show a decline in speeds — a clear convex response consistent with model expectations.

Figure 2: FN/FH Premium Coupon Back-Testing — Model CPR vs. Observed CPR

GNMA — FHA and VA Segmentation

GNMA performance showed a similar pattern to Fannie/Freddie across discount and premium coupons. A notable enhancement this month: the team has introduced the ability to split GNMA back-testing results by FHA vs. VA segments on the Vertex report, providing additional analytical granularity.

Two segment-level observations:

FHA: FHA: The model has shown a slight drift in prepayment speeds over the past year. This is primarily attributed to the FHA trial modification policy change in 2025, under which servicers are no longer required to buy out delinquent loans — a policy shift that has meaningfully reduced prepayment speeds relative to historical levels.

VA: VA: A discrepancy between modeled and actual speeds reflects the default VA-to-PMMS spread assumption being too tight. Users can adjust the VA spread to current market levels within the platform to bring results into closer alignment with observed speeds.

Figure 3: FHA Segment Back-Testing

Figure 4: VA Segment Back-Testing

New Non-QM Credit Model: CM 7.1

Guanlin Chen from the Quantitative Modeling Group presented an overview of the upcoming Non-QM Credit Model, version 7.1, expected to be available to all users near end of Q2 2026.

Model Structure

CM 7.1 follows the same three-component framework as RiskSpan’s agency credit model:

1. Transition Model — generates a time-varying transition matrix estimating transition probabilities across delinquency states and time periods

2. Liquidation Timeline Model — applied once a loan enters default

3. Severity Model — estimates final losses on the defaulted balance

A key design decision: the model is built with separate sub-models for each documentation type, consistent with the segmentation used in the Non-QM prepayment model: Bank Statement, DSCR, Full Doc, and Other. This segmentation reflects the meaningfully different performance characteristics across these loan types and allows for more accurate, documentation-specific projections.

Transition Matrix Design

The transition matrix tracks loans across delinquency states — current (0), one-month delinquent (1), two-month delinquent (2), foreclosure (F), and REO (R) — with additional granularity for delinquency history.

Back-Testing Results

Initial back-testing of the 30-day delinquency transition demonstrates that the model captures the overall trend in Non-QM credit performance well across all four documentation types. The COVID period was intentionally excluded from model training — including it would have caused extreme unemployment levels to dominate the model and distort sensitivity to other risk factors.

Figure 5: NonQM Credit Model — Current to 30DPD Transition: Actual vs. Projection by Documentation Type

A dedicated webinar covering CM 7.1 in detail is planned for end of May or early June. Please stay tuned.

Housing Turnover Analysis: Rate Sensitivity at Positive Refinancing Incentive

Shane Lee from the Quantitative Modeling team presented new research on housing turnover behavior in a positive refinancing incentive environment.

Background

Total prepayment has two components: housing turnover (prepayment driven by home sales) and refinancing. In current model design, housing turnover is assumed to be weakly rate sensitive — and in the positive refinancing incentive regime, sensitivity is held at zero. The question the team set out to answer: is that assumption correct?

Data Sources

NAR: National Association of Realtors (NAR) Existing Home Sales — measures the number of homes sold including single-family, condo, and co-op properties (including sales without mortgages)

Equifax ADS: Equifax ADS Data — tracks trade lines per consumer, allowing the team to identify housing turnover by flagging cases where an existing mortgage closes and a new mortgage originates at a different ZIP code for the same borrower

Figure 6: ADS vs. NAR Data Comparison — Prepaid Mortgages vs. Home Sale Units

Key Finding

During the post-COVID refinancing boom, housing turnover activity increased significantly — by almost 50% above the baseline level. This elevated turnover coincided with the period of low rates and high refinancing activity, driven in part by the work-from-home migration wave.

Figure 7: Housing Turnover CPR — ADS Data (nearly 50% above baseline during COVID refi boom)

This finding suggests that housing turnover is more rate-sensitive in a positive refinancing incentive environment than current models assume — a potential source of underestimation when projecting prepayment speeds in a low-rate environment. Research is underway to incorporate this into Prepayment 4.0.

Macroeconomic Update: April 2026

Federal Reserve — No Rate Cuts Expected in 2026

CME FedWatch futures currently indicate no Federal Reserve rate cuts this year. The overall expectation is that the Fed funds rate (currently 350–375 bps) will remain unchanged through year-end 2026.

Figure 8: Federal Funds Target Range — Upper Limit (Source: FRED)

Mortgage Rates — Elevated and Volatile

Mortgage rates hit a six-month high in March, with Freddie Mac’s primary rate reaching 6.45% and Mortgage News Daily data showing rates approaching 6.64%. Rates have since pulled back modestly as some geopolitical uncertainty subsided. The 10-year Treasury rate is expected by market consensus (econforecasting.com) to remain above 4% for the next three to five years — implying mortgage rates are unlikely to fall significantly below 6%.

Figure 9: 10-Year Treasury Yield — Historical and Market Consensus Forecast

Figure 10: Primary Mortgage Rate Trend

Unemployment and Inflation

March unemployment rate: 4.3% — trending upward

PCE (excluding food and energy): approximately 3% — still above the Fed’s 2% target

Figure 11: Unemployment Rate

Figure 12: PCE Inflation (ex. Food & Energy)

Consumers continue to face pressure from elevated gasoline and oil prices. Credit card balances have risen significantly over the past two years, adding to the financial strain on households.

Home Prices — Stabilizing but Elevated

Home price growth remains positive but has decelerated substantially from the 20%+ year-over-year peaks observed in mid-2022:

Case-Shiller National Index: approximately 1% year-over-year growth

10-City Composite: slightly above the national index

FHFA All-Transaction Index: somewhat stronger, indicating variation across market segments

Figure 13: Case-Shiller U.S. National Home Price Index — Year-over-Year % Change

Two regional case studies highlight the range of outcomes: Austin, TX and Boise, ID both experienced peak growth of 30–35% in 2021–2022, followed by sharp declines through 2023, and are now returning to modest positive territory. Housing supply remains severely constrained.

Summary

Topic	Key Takeaway
Prepayment Model	Performing well overall; March speeds driven by February rate rally and day count effects
GNMA Segmentation	New FHA/VA split available in Vertex; FHA drift tied to 2025 trial mod policy change
NonQM Credit Model	CM 7.1 on track for end of Q2 2026; dedicated webinar coming end of May / early June
Housing Turnover	Rate sensitivity confirmed in positive refi regime; ~50% above baseline during low-rate period; research underway for Prepayment 4.0
Mortgage Rates	Hit six-month high of 6.64% in March; expected to remain above 6% through 2026–27
Fed Policy	No rate cuts expected in 2026; Fed funds rate at 350–375 bps
Home Prices	Growth slowing (~1% nationally); supply constraints persist

We continue to add additional analytics reports on the Platform. Please visit www.riskspan.com/request-access to request free access.

As always, please feel free to contact us to discuss or learn more.

Models & Markets Update: March 2026

Register here for next month’s call: Thursday, April 16th, 2026, 1 p.m. ET.

Key takeaways from this month’s call:

Non-mortgage credit is deteriorating more rapidly than mortgage credit

BNPL usage may be masking underlying financial strain

Macroeconomic conditions are likely to remain restrictive, reinforcing current trends

Prepayment models remain well-calibrated, even as borrower behavior begins to shift

You can read the recap below or click here for the entire 20-minute recording.

Credit Performance by Asset Class

The data shows a clear divergence between mortgage and non-mortgage credit:

Mortgage delinquencies remain relatively low, supported by tighter underwriting standards

Credit card delinquencies have increased meaningfully since 2022

Auto loan delinquencies are approaching levels observed during the Global Financial Crisis, particularly among younger borrowers

The following charts from NYFed illustrate how younger age cohorts are consistently exhibiting higher delinquency rates across credit types (mortgages, credit cards, and autos).

Buy Now, Pay Later (BNPL) usage continues to expand.

Adoption is highest among younger borrowers.

A meaningful portion of usage is for essential expenses such as groceries

Because BNPL obligations are not consistently captured in traditional credit metrics, they may obscure underlying levels of consumer leverage and stress.

Macroeconomic Outlook: Rates Expected to Remain Elevated

The macroeconomic environment continues to support a “higher-for-longer” rate outlook.

Market expectations suggest no Federal Reserve rate cuts through 2026.

The 10-year Treasury rate is expected to remain above 4% over the next several years.

Mortgage rates, after declining earlier in 2026, have risen again and are expected to remain near or above 6%.

At the same time:

Inflation remains above target levels

Unemployment is trending upward

These conditions suggest a continued tightening backdrop for borrowers, with limited relief from monetary policy in the near term.

Housing Market: Moderation Continues

Home price growth remains positive but has slowed:

Case-Shiller index shows modest annual growth (~1.3%)

FHFA index indicates somewhat stronger growth (~3.3%)

Differences between indices suggest variation across market segments, with relatively stronger performance in more affordable segments and geographic differences on home prices.

Prepayment Trends: Stable Performance with Emerging Shifts

Against this macro and consumer backdrop, prepayment behavior continues to evolve.

Prepayment models remain closely aligned with realized speeds across FN/FH and GNMA collateral, as shown in the coupon-level comparisons.

Refinance behavior is well captured, including sensitivity to changes in mortgage rates.

There are, however, early indications of shifting borrower behavior:

Prepayment speeds increased in February despite fewer collection days, suggesting a gradual weakening of the mortgage rate “lock-in” effect.

Short-term rate increases may moderate this trend, but the directional change is notable.

GNMA Segmentation Enhancements

The introduction of FHA and VA segmentation in GNMA back-testing provides additional analytical detail.

FHA performance shows some divergence, likely reflecting recent policy changes affecting delinquent loan buyouts.

VA results are more sensitive to spread assumptions and can be adjusted to align with market conditions.

We continue to add additional analytics reports on the Platform. Please visit www.riskspan.com/request-access to request free access.

As always, please feel free to contact us to discuss or learn more.

From Household Debt to Non-QM Credit: February Models & Markets Recap

Register here for next month’s call: Thursday, March 19th, 2026, 1 p.m. ET.

In this month’s Models & Markets call, RiskSpan’s quantitative modeling team tackled:

The record debt levels now carried by U.S. households (and the consumer stress that is building beneath the surface);

The likely persistence of higher rates;

RiskSpan’s forthcoming non-QM credit model, and;

(as always) how RiskSpan’s prepayment model is performing

You can read the recap below or click here for the entire 20-minute recording.

$18 Trillion in Household Debt (and Growing)

 U.S. household debt reached $18.8 trillion at the end of 2025 and continues to climb. Mortgages account for the largest share at roughly $13 trillion, with auto loans, student loans, credit cards, and HELOCs making up the balance.

Using conservative assumptions for average interest rates for each category, we estimate that these balances equate to roughly $1.1 trillion in annual interest payments and $2.5 trillion in total annual debt service payments – the approximate cost required each year just to keep households current.

A Distributional Problem

The aggregate debt figure masks meaningful stress at the lower end of the income spectrum:

A median household (~$80K gross income) may devote roughly 37% of disposable income to debt service.

Bottom-quartile households (~$32K gross income) may spend 40–55% of disposable income servicing debt.

Lower-income households are disproportionately exposed to higher-rate revolving credit and subprime auto loans, as opposed to 3–4% fixed-rate mortgages. The averages therefore understate the severity of strain on the more vulnerable segments.

Are Reported Delinquencies Understating Stress?

Delinquencies are rising across income levels, particularly in lower-income areas. Lenders, however, may be quietly modifying or re-aging loans, particularly in consumer credit categories (e.g., auto loans). Such modifications can:

Push missed payments to the back of the loan

Reset accounts to “current” status

Avoid immediate charge-offs

While this suppresses reported delinquency statistics, borrower balances may continue to grow. This implies that reported delinquency rates may really be more of a floor, as aggregate DSCR and household stress may be greatly understated.

Consumer strain is real and potentially worse than what is suggested by the headline metrics.

Coming in Q2: RiskSpan’s Non-QM Credit Model!

RiskSpan’s forthcoming non-QM credit model will feature four distinct documentation categories:

Bank Statement

Full Documentation

DSCR/Investor

Other (e.g., VOE, asset depletion)

Each segment is modeled independently through a transition-matrix framework covering:

Current

30-day DQ

60-day DQ

90+ DQ

Termination (voluntary and involuntary)

Prior delinquency figures prominently in the model, with clean loans having relatively low base transition rates from current to delinquent, while loans with prior delinquency history can experience transition probabilities up to 10x higher. Capturing this conditional risk dynamic is central to the model’s design.

Back-testing (shown below for the Full Doc segment) indicates the model is tracking historical delinquency transition rates reasonably well, though development remains ongoing.

Macro Considerations

Consistent with prior months, the macro backdrop continues to reinforce a “higher-for-longer” rate environment.

Fed and Policy Outlook

Fed Funds expectations imply limited cuts in 2026.

No immediate expectation of a March rate cut.

10-Year Treasury

Consensus forecasts suggest the 10-year Treasury will remain above 4% for the next 2–3 years.

Recently, it has hovered around ~4.1%, down slightly from prior highs.

Mortgage Rates

Primary mortgage rates are approaching 6%, but not sustainably breaking below it. In our view, mortgage rates are likely to remain around or above 6% through 2026, possibly into 2027.

Labor, Inflation, and Home Prices

Unemployment ticked down slightly.

Job creation surprised to the upside.

Inflation remains sticky in the 2.5–3% range

National home prices showed modest year-over-year growth (~1.4%).

Traditional seasonal adjustments may be less reliable in today’s inventory-constrained housing market. Turnover seasonality appears to be shifting earlier in the year, with implications for both pricing and prepayment dynamics.

Prepayment Model Performance: Stable & Improving

Despite macro headwinds and rising consumer stress, RiskSpan’s prepayment models continue to perform well.

GSE Discounts (WAC 5.5 and Below)

Prepayments declined slightly in the most recent month, primarily due to fewer collection days and normal January seasonality (lower turnover). Overall model fit remains strong.

One identified refinement: the model’s seasonal peak appears slightly delayed (June/July shifting toward August). This will be addressed in the next version update

GSE Premiums (WAC 6 and Above)

Refinance speeds have been largely unchanged over the past two months. S-curve comparisons between December and January show no material differences once recount adjustments are made. A modest ~1.5 CPR change in recent data appears driven by turnover rather than refi activity.

Ginnie Mae: FHA vs. VA Enhancement

Performance across Ginnie segments remains solid, with recent prepayment dips again attributable to fewer collection days. However, we have observed divergence between FHA and VA: Modeled FHA speeds tend to be overestimated, while modeled VA speeds tend to be underestimated compared to recent historicals.

To address this, RiskSpan is adding a loan guarantor filter to the back-testing report, enabling FHA and VA splits (expected early March). This enhancement will improve transparency and precision in Ginnie performance analysis.

We continue to add additional analytics reports on the Platform. Please visit www.riskspan.com/request-access to request free access.

As always, please feel free to contact us to discuss or learn more.

Why AI Won’t Kill Asset-Backed Finance Software — and Why the Last Mile is the Moat

Every wave of financial technology innovation brings the same prediction: software will be commoditized. Today, that prediction is being applied to AI. If AI models can reason, summarize, and generate code, the thinking goes, B2B vertical SaaS becomes unnecessary.

That conclusion is inherently wrong. ABF platforms are not feature layers, they are governed systems.

The last mile of AI deployment isn’t friction—it’s the moat.

ABF Is Not a “Promptable” Problem

ABF platforms sit directly in the flow of capital allocation, risk management, and regulation. For asset managers deploying institutional capital, this creates a very high bar for reliable data, validated models and domain-specific workflow.

The real question isn’t whether a system can produce answers. It’s whether it can produce results that are:

Consistent over reporting periods and market cycles

Explainable under stress and investor scrutiny

Defensible and robust enough for LPs, investment committees, and regulators

That high bar changes everything. It explains why technology adoption in financial markets moves cautiously and why legacy systems persist. These systems embed decision rights, controls, and institutional logic that can’t simply be recreated with better prompts. Any platform that ignores this reality will struggle to scale beyond pilots.

Which leads to the obvious question: if AI is so powerful, where does it actually help?

AI Accelerates Workflow — Not Accountability

Applied correctly, AI can materially improve ABF workflows. It can ingest complex credit agreements faster, reconcile data across counterparties, flag covenant breaches, and reduce manual reporting work. In other words, AI increases operational leverage.

But AI does not remove the need for explicit deployment configuration and governance. Institutions still must define who owns key assumptions, which decisions can be automated, and where accountability sits when outcomes affect capital. These embedded design choices (not prompts) ultimately determine whether a platform is trusted.

AI compresses timelines, but responsibility remains fixed. Once this distinction is recognized, the broader implication becomes clear: AI does not eliminate the need for software. It raises the bar for it.

Software Remains the System of Record

The idea that AI replaces SaaS also misunderstands where SaaS enterprise value lives. Enterprise value in ABF doesn’t live in isolated insights. It lives in controlled systems of record and durable platforms that provide:

Governed data and the system of record

Embedded domain expertise

Repeatable processes that survive personnel turnover

A shared source of truth across counterparties, investment, risk, accounting, and investor relations

AI without software discipline creates speed without stability. With it, AI becomes force-multiplying. The question, then, is what separates platforms that successfully integrate AI from those that don’t.

The Real Differentiator: Deployment Intelligence at Scale

What separates enduring platforms from feature-rich tools is not model sophistication—it’s deployment intelligence — the ability to integrate AI into live production environments without weakening controls. That requires:

Controlled data pipelines designed for real-world imperfections

Configuration layers that adapt to fund-specific structures without breaking controls

AI outputs that are transparent, and auditable

Implementation treated as a repeatable product, not bespoke services

This is where defensibility emerges. Deployment intelligence compounds with each client rollout. Each successful implementation strengthens the next, deepening institutional trust and operational resilience. AI amplifies this flywheel but cannot replace it.

The Mispriced Risk of “AI-Only” Narratives

In private credit, trust is earned slowly and lost quickly. It is built through consistent valuations, defensible reporting, and reliability during market dislocations.

A system that produces faster answers, but weaker confidence does not displace incumbents. It increases operational and reputational risk. Investors should be wary of platforms that promise instant replacement without acknowledging institutional reality of fiduciary-grade infrastructure.

The Investment Takeaway

AI is not commoditizing ABF software solutions. It is widening the moat for platforms that integrate AI responsibly into governed systems.

The next phase of growth for category leaders such as RiskSpan will be driven by combining deep domain knowledge with AI-native architecture. Leaders will treat the last mile – data integration, workflow configuration, and control design — as a core product capability, not an implementation afterthought.

In markets where trillions in capital allocation depend on data integrity and institutional trust, the last mile isn’t an implementation detail.

It’s the moat.

Rates, Prepays and Consumer Stress: What the Data is Telling Us at the Start of 2026

 Register here for next month’s call: Thursday, February 19th, 2026, 1 p.m. ET.

In the January Models & Markets call, our quantitative modeling team hosts their first monthly deep dive of the year into prepayment model performance, an updated analysis of second liens and HELOCs using Equifax data, and the evolving macroeconomic backdrop shaping mortgage markets.

Here’s a quick recap in case you missed it.

(Click here for the entire 20-minute recording or continue reading for a summary.)

Revised HELOC and HEL Results Using Equifax ADS Data

  We performed a comprehensive analysis of second liens and HELOCs using Equifax’s Analytic Data Set (ADS), which represents a 10% anonymized sample of U.S. consumer credit data at the tradeline level.

Following the resolution of data quality issues identified in an earlier analysis, the revised results now align much more closely with economic intuition. Prepayment speeds behave consistently across vintages, credit score bands, and refinancing regimes.

One key takeaway holds that higher credit score borrowers tend to prepay faster, particularly during refinancing waves, while lower credit score segments remain slower. This pattern is especially evident in post-COVID vintages. Overall credit quality for HELOCs and second liens remains strong, with performance clustering closer to the highest credit score bands.

Another notable observation is the role of seasonality in newer HELOC vintages. In a high-rate environment with limited refinancing activity, turnover-driven prepayments become more prominent. Baseline prepayment speeds for HELOCs are running around 15 CPR, higher than what is typically observed in first-lien portfolios under similar conditions. These dynamics provide useful signals for understanding how first-lien behavior may differ when second liens or HELOCs are present on the same property.

  We plan to expand this analysis further, including deeper investigation into correlations between first- and second-lien prepayment behavior.

Mortgage Rates Remain Likely to Stay Higher for Longer

The broader economic outlook remains one of persistence rather than relief. Federal Reserve projections point to unemployment stabilizing around the low-4% range and real GDP growth near 2% over the medium term. Meanwhile, expectations for the fed funds rate suggest limited room for significant cuts beyond 2026.

Longer-term rates tell a similar story. Consensus forecasts indicate the 10-year Treasury is unlikely to fall meaningfully below 4% over the next two to three years, implying mortgage rates are likely to remain near (and potentially above) the 6% level for much of the period ahead. Temporary dips tied to policy announcements or market events have proven short-lived, with rates quickly reverting back toward recent levels.

Consumer Stress Continues to Build

While headline spending remained strong during the most recent holiday season, the composition of that spending tells a more cautious story. Consumers increasingly favored lower-cost retailers, suggesting budget sensitivity and selective spending behavior.

Survey data reinforces this theme. Year-over-year consumer sentiment and expectations have declined meaningfully, and perceptions of job insecurity (particularly among college-educated workers) have become more negative. These dynamics could have important implications for credit performance and housing activity as economic uncertainty persists.

Prepayment Model Performance: v. 3.7 Continuing to Track Market Performance Well

RiskSpan’s prepayment models continue to perform well across Agency collateral.

RiskSpan’s Prepayment Model v3.7 continues to demonstrate strong performance across collateral types. Recent back-testing shows that model projections remain closely aligned with realized speeds, even as seasonal effects and calendar nuances influence month-to-month results.

For conventional 30-year loans with lower coupons, December’s modest uptick in observed CPRs was largely attributable to four additional collection days relative to November. After adjusting for day count effects, actual prepayment speeds continue to trend lower, consistent with expectations in a higher-rate environment.

Premium cohorts also remained largely stable. Despite a brief decline in mortgage rates late last year, the move was insufficient to trigger a meaningful new refinance wave. Most refinance-eligible borrowers have already acted, and the refinancing “pull-forward” effect appears largely exhausted. This dynamic is also visible in the S-curve, which has flattened back toward historical averages after October’s temporary acceleration.

Agency collateral shows similar patterns. Ginnie Mae discount cohorts tracked model expectations closely, while premium cohorts remained flat. One area of ongoing refinement is deep in-the-money, very high-coupon Ginnie Mae loans, where actual speeds have run slightly slower than model projections as refinance incentives flatten out earlier than in prior cycles.

Looking Ahead

In summary:

RiskSpan’s Prepayment Model v3.7 continues to perform well across most collateral segments

HELOC and second-lien analysis using Equifax data now shows economically intuitive and stable results

Mortgage rates are likely to remain near 6% in the absence of a major macro shock

Consumer behavior is showing increasing signs of stress and caution

RiskSpan plans to release additional analytics later this year, including a new non-QM credit model in the first half of the year and a next-generation prepayment model in the second half.

We continue to add additional analytics reports on the Platform. Please visit www.riskspan.com/request-access to request free access.

As always, please feel free to contact us to discuss or learn more.

What a Year of Building AI in Structured Finance Actually Taught Us

The lessons nobody puts in the demo.

In 2025, our team built production AI systems that process billions of performance records for tens of millions of mortgages, develop cash flow models for complex private ABF structures directly from documents, and connect large language models directly to bond analytics APIs.

We built dashboards, connectors, and credit analytics. Some of them worked. Some of them taught us more by failing.

This is what we learned—not the polished conference talk version, but the notes we’d share with a peer team starting the same journey.

The Value Shift Nobody Prepares You For

A portfolio delinquency analysis that used to take three hours now takes twenty minutes.

That sounds like a win. It is a win. But it also raises a question that’s harder to answer than any technical problem we solved this year:

If AI handles in minutes what took us hours, what are we contributing?

When we started pulling this thread, we realized that a significant portion of what felt like skilled analytical work was actually mechanical labor—data extraction, formatting, applying the same methodology we’d applied dozens of times before. The expertise was real, but it was wrapped in hours of execution that masked how much of the work was routine.

Here’s where we landed:

AI handles the “how.” Humans own the “why” and “so what.”

The value now lives in knowing which questions matter. Understanding what the client really needs versus what they say they need. Recognizing when output is wrong because we understand the domain deeply enough to see the error.

That’s an entirely different skill set. It requires judgment, contextual awareness, and domain intuition that deepens over years—the kind of expertise AI can’t simply replicate, unlike procedural analytical work.

Not everyone will make this transition comfortably. The analysts who built their identity around being fast and thorough at execution face a harder adjustment than those who always saw execution as a means to an end.

We don’t have this all figured out yet. But we’ve stopped pretending the shift isn’t happening.

Stop Asking AI to Write Code—Start Asking It to Think With You

For years, we used Claude as a coding assistant. “Write a function that does X.” “Convert this data from format A to format B.” “Generate a script that calculates Y.”

That works. But it captures maybe 20% of the value.

The shift that changed our results was: treating Claude not as a tool to instruct, but as an analyst to think alongside.

The difference looks like this:

Before (instruction mode):

“Write a Python script to calculate delinquency rates from this loan data.”

After (thinking partner mode):

“We need to identify hidden credit risk in this CLO portfolio—issuers that resemble recent defaults but haven’t shown price distress yet. What factors should we consider? What data would we need? Let’s build a scoring model together.”

That second conversation led to identifying hidden exposure across issuers. Claude suggested factors we hadn’t considered—CLO concentration patterns, industry clustering effects, the relationship between coupon levels and distress signals. We debated the weighting. We refined the methodology. The output was genuinely collaborative.

The code that emerged from the second approach was better, but that’s almost beside the point. The thinking was better. The model was better. The insight was better.

This requires a different posture than most of us learned. You have to think out loud. Admit what you don’t know. Explain your reasoning and invite critique. Treat the AI as a colleague who happens to have read every document and doesn’t get tired—not as a sophisticated autocomplete.

The developers and analysts on our team who made this shift produce substantively different work than those who are still in instruction mode. And the gap is widening.

The First Version Will Be Wrong—Plan for It

We built a benchmark analysis comparing a client’s NonQM loan portfolio against the broader market. The analysis looked solid: the portfolio showed a 1.37% delinquency rate advantage versus the universe. Strong results. Ready to present.

Then someone asked about DSCR loans.

In NonQM lending, DSCR (debt service coverage ratio) loans are a category unto themselves—with measurably better performance than other NonQM products. When we segmented the data, we discovered the universe was comprised of 43% DSCR loans while the client’s portfolio had only 30% DSCR loans.

This changed everything.

The client’s portfolio had less exposure to DSCR loans (the better-performing segment) yet still outperformed the benchmark. That alone was impressive, but our initial analysis understated the true picture. Once we compared performance within segments (DSCR vs. DSCR, non-DSCR vs. non-DSCR), the client’s edge was even larger than we’d initially observed.

If we had presented the first version, we would have undersold our client’s own performance. The insight that mattered most—superior underwriting across both loan categories—would have been invisible.

Lesson: “Wrong” doesn’t mean broken. It means the output doesn’t fully reflect reality. Have a domain expert review the work before drawing conclusions.

Deploying AI Agents for End Users Is a Security Project

Building an AI agent that works in a demo is straightforward. Deploying that agent in a production UI where real users interact with real data took us months.

We built an agent that lets users query our bond analytics platform conversationally. The AI worked. Making it production-ready required solving problems:

Prompt injection: When users can type anything into a text box processed by an LLM, you inherit a new attack surface. We implemented input validation, output filtering, tightly scoped permissions, and logging that captures every agent action for audit.

Rate limiting: A single conversational turn might trigger 50 API calls. We built tiered limits—per-user, per-session, per-token—plus circuit breakers for runaway queries.

Session management: Agent sessions need conversational context across multiple turns, isolated per user, with graceful expiration handling and automatic cleanup.

Audit trails: Regulated industries need to know what the AI did. Every query, tool invocation, and response needs to be logged immutably.

The agent itself was 20% of the effort. Authentication, authorization, input validation, rate limiting, session management, and security review were the other 80%.

Lesson: In production, the agent is the easy part. The security wrapper is the product.

Post script: AgentCore from AWS and Agent Framework from Microsoft are solving the deployment and security headaches.

AI Is Good at Finding Information But Sometimes Overstates What It Means

While building the credit risk analysis, we asked Claude to research distressed issuers—companies that had defaulted or were showing signs of stress. We wanted to understand patterns we could use to identify similar risks in the portfolio.

Claude surfaced real-time signals we wouldn’t have found efficiently on our own: FTC antitrust actions, rating agency downgrades, refinancing walls, fraud allegations. Information that wouldn’t appear in pricing data for months was available in news coverage and regulatory filings. The research phase that would have taken days was completed in hours.

But we also caught Claude drawing confident conclusions from weak sources. In one case, it attributed claims to “industry reports” that didn’t exist when we followed the links. The search results were real. The sources were ‘real’. But the synthesis drew conclusions the sources didn’t support.

The lesson: use AI-powered search aggressively. It’s the difference between stale knowledge and current intelligence, especially in fast-moving situations. But verify specific claims. Click the links. Read the actual sources.

AI is excellent at finding relevant information across large volumes of text. It is sometimes too confident about what that information means when synthesized. The combination of broad retrieval and skeptical verification is more powerful than either alone.

Your Org Chart Isn’t Ready for This

Our AI strategy deck included projections: reduction in onboarding costs, increased client capacity and margin expansion.

The numbers were defensible. The business case was clear.

What the projections didn’t address: the organizational implications of realizing the promised efficiencies.

If analysts can serve five times more clients, do you need fewer analysts—or do you pursue five times more clients? If the answer is “more clients,” do you have the sales capacity? The support infrastructure? The management bandwidth?

If developers now own adoption metrics for the features they build, then what happens to the product managers who previously owned that? Are product managers freed up for more strategic work, or are they defending territory?

If AI drafts client communications, who reviews them? What error rate are we willing to accept? Who’s accountable when the AI gets something wrong?

These aren’t hypothetical questions. We’re navigating them now, and the answers aren’t obvious.

AI doesn’t just improve workflows. It reshapes roles. And most organizations—including ours—are making it up as they go.

The companies that figure out the organizational design will outperform those that simply purchase better software. The differentiation in 2026 won’t come from adopting AI. It will come from redesigning teams, incentives, and accountability structures around what AI makes newly possible.

What We’re Taking Into Next Year

A year of building AI systems in structured finance clarified a few things:

AI is more powerful than the hype suggests—once you integrate it into real workflows rather than treating it as a research toy.

AI is more frustrating than the demos show—the gap between “works in claude.ai” and “works in production” is where most of the time goes.

AI is more dependent on domain expertise than the automation narrative implies—it generates analyses quickly, but distinguishing plausible from accurate requires human judgement that compounds over years. The “why” and “so what” remain stubbornly human problems.

AI changes more than technology—it changes job descriptions, team structures, and how people understand their own value. The skill isn’t operating the tool; it’s knowing when the output reflects reality.

We don’t have all the answers. We’re still learning what this means for how we build software, how we serve clients, and how we organize ourselves.

But we’re no longer wondering whether AI will change our industry. We’re focused on making sure we’re the ones defining how.

12 3 4 5 6

Key Takeaways

Prepayment Model Back-Testing

Conventional Deep Dive

Non-QM Universe and Credit Model 7.1

Macroeconomic Update

The rerun

The economics

The overconfidence risk (AI hallucination) is worse than the bill

Use the LLM for Judgment, Not Execution

Where to put the production version

A simple approach

Key Takeaways

Prepayment Model Back-Testing: June Factor Data Update

Fannie/Freddie — Discount Coupons (WAC 5.5 and Below)

Fannie/Freddie — Premium Coupons (WAC 6.0 and Higher) and S-Curve

GNMA — Discount and Premium Coupons

Non-QM Historical Performance

Prepayment Speeds

Delinquencies

Non-QM Credit Model: CM 7.1 Update

Macroeconomic Update: June 2026

Federal Reserve — On Hold, With a Hike Now on the Table

Rates, Inflation, and Home Prices

Summary

Key Takeaways

Prepayment Model Back-Testing: May Factor Data Update

Fannie/Freddie — Discount Coupons (WAC 5.5 and Below)

Fannie/Freddie — Premium Coupons (WAC 6.0 and Higher)

GNMA — Discount and Premium Coupons

Non-QM Credit Model: CM 7.1 Update

Model Structure

Macroeconomic Update: May 2026

Federal Reserve — No Rate Cuts Expected; Hike Risk Emerging

Treasury Yield Curve — Significantly Higher Than February

Mortgage Rates — Back to Summer 2025 Levels

Inflation — Staying Elevated

Home Prices — Stagnating, with Pronounced Regional Variation

Consumer Stress: Evidence from Credit Card Spending

Credit Landscape

Boston Fed Analysis: Spending and Debt by Income Group

Buy Now, Pay Later: An Untracked Risk

Summary

Credit Performance by Asset Class

BNPL Usage as a Potential Blind Spot

Macroeconomic Outlook: Rates Expected to Remain Elevated

Housing Market: Moderation Continues

Prepayment Trends: Stable Performance with Emerging Shifts

GNMA Segmentation Enhancements

$18 Trillion in Household Debt (and Growing)

Coming in Q2: RiskSpan’s Non-QM Credit Model!

Macro Considerations

Prepayment Model Performance: Stable & Improving

Company

Products

Security & Compliance