Linkedin    Twitter   Facebook

Get Started
Log In

Linkedin

Articles Tagged with: Gen AI

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. 


The Insurance ABF Stack: Panel Takeaways

We opened the insurance panel at the RiskSpan Summit earlier this month with an interesting statistic: according to Moody’s, almost a third of the $6 trillion in cash and invested assets held by US life insurers is now allocated across private credit sub-asset classes. Nancy Mueller Handal of Bayview confirmed it tracks with consideration to private placements, commercial mortgages, and infrastructure debt alongside direct lending. Bill Moretti of Equitable added a useful caveat: the industry still doesn’t have a settled definition of “private credit,” and ABF has only recently been pulled under that umbrella. However you count it, the scale is real and it’s happened fast.

I was joined by Nancy, Bill, and Larry Yang of Global Atlantic / KKR — three practitioners operating through fundamentally different models. What made the conversation worth having was exactly that diversity of structure. Here’s what stuck with me.

You may not know you’re in a bad vintage while you’re living through it.

This was the sharpest exchange of the panel. The group flagged the current non-QM environment specifically: massive appetite, compressed spreads, and layered risk that has never been tested in a declining home price environment. Non-QM has existed entirely within a rising housing market. There was even speculation that we may be experiencing a vintage risk event in data centers right now – and the industry won’t know for years. The most uncomfortable version of vintage risk isn’t the one you can see. It’s the one you’re inside.

The data problem is harder than the analytics problem.

One panelist observed that in structured finance, “you could get to a point where you don’t know what you own.” Larry described building proprietary end-to-end infrastructure — including custom waterfall models — because off-the-shelf systems aren’t granular enough. Nancy’s observation was the most memorable: her research team has grown as capabilities have improved, not shrunk. More tools enable more questions. The data was always there. The limiting factor has always been the ability to extract, manage, and act on it in real time.

AI is a co-pilot, not a substitute.

The panel’s most grounded take: “Seasoned people know what questions to ask — and know when the answer is wrong.” Junior analysts prompting a model without that judgment aren’t getting the same output. The firms making real progress are the ones embedding AI into durable processes, not just re-prompting the same task each month.

And on why they’re still leaning in: Nancy called it “the most exciting market out there.” Larry cited the breadth: in ABF, you’re always encountering asset classes you’ve never seen before. The undisputed line of the day: “Insurance companies are now sexy.”

Hard to argue with that.

The RiskSpan Summit 2026 brought together practitioners across insurance, asset management, and structured finance.


How Mortgage Teams Are Using RiskSpan’s AI Agent to Answer MBS Data Questions in Seconds 

Agency MBS data has never been more available and acting on it has never been harder. 

Mortgage servicers, secondary marketing desks, and risk teams are sitting on a mountain of GSE data that could be driving sharper decisions every day. The challenge isn’t access. It’s time. The typical path from question to answer — submit a data request, wait for an analyst, review a report — can take days. By then, the market has moved. 

RiskSpan built our AI MBS Data Agent to close that gap. Below are four ways mortgage teams are using it today. 

1. Secondary Marketing: Walking Into Negotiations With Data 

Spec pool pay-ups are negotiated daily, but the benchmarks that justify them — CPR performance by coupon, LTV band, servicer cohort — are rarely at your fingertips. 

Secondary marketing teams using the AI MBS Data Agent can ask questions like: 

“Show me our CPR performance for 3.5% coupons from high-FICO, low-LTV collateral over the past 24 months — compared to the market average.” 

The agent returns the answer in seconds, with auditable SQL and export-ready output. That’s the difference between entering a negotiation informed and entering it exposed. 

2. Risk & Underwriting: Surfacing Concentrations Before They Become Problems 

Portfolio risk doesn’t announce itself. It builds quietly in LTV bands, geographic clusters, and servicer cohorts — until a delinquency spike makes it visible. By then, options are limited. 

Risk and underwriting teams use the agent to run surveillance queries that would have previously required analyst hours: 

“Flag servicers in my cohort where 90+ day delinquency rates have increased more than 15% year-over-year — broken out by LTV band and geography.” 

The agent handles the data work. The team focuses on the decision. High-LTV surveillance, adverse selection monitoring, Ginnie vs. GSE comparisons — all available on demand, in plain English. 

3. Default Servicing: Early Warning Signals, Not Lagging Reports 

Default servicing teams need to see trouble coming — not confirm it after it arrives. But most teams are working from monthly data pulls that are already weeks old by the time they reach a desk. 

With the AI MBS Data Agent, monitoring becomes continuous and conversational. Teams can query delinquency trends across servicer cohorts, compare their performance against market benchmarks, and catch emerging signals before they escalate — all without submitting a ticket. 

This is especially valuable for servicers managing Ginnie portfolios, where early delinquency detection can directly affect buyout timing and loss severity. 

4. Executive Reporting: Competitive Benchmarking on Demand 

For mortgage executives, competitive intelligence is often a quarterly exercise — because assembling it takes weeks of analyst time. Market share by investor channel. CPR performance vs. peers. Spec pool trends. These are questions that should have instant answers. 

The AI MBS Data Agent makes competitive benchmarking a conversation, not a project. Executives can ask: 

“What is our market share by investor channel versus the top 10 servicers — and how has it trended over the past 12 months?” 

Fannie, Freddie, Ginnie. Any vintage. Any cohort. The answer comes back in seconds, ready for a board deck or an internal review. 

How It Works 

The AI MBS Data Agent connects to agency MBS data through a natural language interface. Users ask questions in plain English. The agent translates them into SQL, retrieves the data, and returns structured answers with the underlying query visible for audit. 

There’s no data team required to get started, no onboarding overhead for individual queries, and no waiting. The output is export-ready. 

Try It Free for 30 Days 

RiskSpan is offering a 30-day free trial of the AI MBS Data Agent — full access, no credit card required. 

🔗  https://riskspan.com/ai-mbs-agent-offering/ 


AI’s Uneven Impact on Labor Demands a Local Housing View

By: Scott Anderson and Bernadette Kogler

AI is already disrupting parts of the U.S. labor market. The more contested question for residential mortgage investors is how and where. The dooms-day camp sees broad structural displacement where white-collar knowledge workers are replaced by AI agents. The optimists counter that AI makes workers more productive, that falling labor costs will expand demand leading to net employment gains. Both camps agree on one thing: the impact will not be uniform. Disruption will concentrate in pockets — specific occupations, metro centers, and income bands.

That specificity is precisely what makes this a mortgage credit problem. Local labor shocks have always transmitted to local housing markets through a familiar mechanism: income impairment reduces demand which pressures home prices, and price declines erode the collateral cushion that protects investors. What varies is the speed, the depth, and whether prices recover once the shock stabilizes.

Early indicators are consistent with the thesis. A Wall Street Journal analysis found the unemployment rate for recent college graduates ages 20–29 at 7.1% as of October 2024 — nearly double the 4.1% rate for the general population — with computer science graduates at 6.1% unemployment and computer engineering majors at 7.5%. The Federal Reserve of NY also published data showing labor market conditions worsened for recent college graduates at the end of 2025, with the unemployment rate climbing to about 5.7% in Q4 2025 — and underemployment rising to 42.5%, its highest level since 2020.

The occupations most exposed to AI displacement cluster in precisely the metros where high-balance prime mortgages are most concentrated, changing the traditional metrics that investors rely on for early warning. AI-driven labor disruption is not a macro thesis to be managed with national averages. While mortgage borrower occupation data is not complete or granular enough to drive AI job loss assumptions, property location data is. The risk needs to be identified, stress-tested, and monitored at a granular level — before it shows up in delinquency data.

The good news is that the data, models and historical reference points exist. With the growth of Non-QM lending, RTL and loan products that carry credit risk without an insurance wrapper, treating AI-driven labor disruption as a present-tense portfolio risk is not optional. To move from thesis to numbers, we ran a portfolio-level stress test on the universe of securitized Non-QM mortgages – starting with how AI exposure is measured at the local level.

Key Findings for investor consideration:

  1. AI labor disruption is expected to have a higher disproportional impact on white-collar professional jobs that are concentrated in certain “knowledge-work” centers.
  2. The largest job markets and the deepest credit pools are systematically the most AI-exposed, suggesting AI risk and portfolio risk overlap rather than diversify.
  3. Stress-testing the $196B universe of securitized Non-QM mortgage loans shows the magnitude is meaningful: under an AI-driven stress scenario calibrated to GFC-level shocks but re-allocated to AI-exposed geographies, RiskSpan’s credit model projects universe-wide cumulative defaults to nearly double and losses more than quadruple.
  4. Securitized deals benefit from built-in geographic diversification, which dampens the absolute impact across deals. Whole-loan investors, who construct their own geographic mix, face concentration risk that accumulates silently if exposure isn’t an explicit input into selection and surveillance.
  5. Act now — put the infrastructure in place to use granular data to monitor unemployment, HPI and loan performance and apply scenarios where relevant.

Measuring AI Exposure at the Local Level

Leveraging research published by the Brookings Institution, RiskSpan used their geographic AI exposure scores as the foundation for a credit stress scenario. The resulting county- and metro-level scores follow a similarly clear geographic pattern to that published by Brookings. Highly educated, high-paying knowledge-work centers — Santa Clara (42.8% exposure), King County WA, New York County, DC, Boston — show the highest exposure. Smaller industrial and rural counties score considerably lower. As a sanity check on the methodology, we aggregated and plotted state-level AI exposure against the share of each state’s adult population with a bachelor’s degree or higher (a well-understood proxy for knowledge-work concentration). The two track closely, which is the result one would expect — it confirms the Brookings measure is capturing what it intends to and isn’t producing arbitrary geographic rankings.

Article content
Figure 1

Unlike earlier automation, generative AI concentrates in exactly the metros that house the largest mortgage markets: as the chart below shows, the most AI-exposed MSAs are also among the deepest non-agency RMBS markets.

Article content
Figure 2

One important caveat before moving on to how we used this data. As Brookings is careful to note, “‘exposure’ does not speak only to the displacement of workers; it also may involve their ‘augmentation’ through rapidly improving AI tools.” The same coder whose role is highly exposed could be made dramatically more productive rather than displaced; the same financial analyst could spend less time on rote modeling and more on judgment-intensive work. Whether exposure plays out as net displacement, net augmentation, or some mix will vary by occupation, employer, and how quickly the technology evolves. For purposes of this exercise, we focus on the displacement story: the scenario where AI exposure translates into labor market shocks that transmit to housing markets and mortgage credit. That’s not the only possible outcome, but it’s the one investors need to be prepared for.

Translating that exposure into a portfolio-level credit stress test on the Non-QM universe, the largest segment of the non-agency RMBS market, required two steps.

Step 1: Construct an AI-driven stress scenario calibrated to the GFC.

Rather than forecast AI displacement magnitude — a genuinely uncertain exercise — we sized the stress to a known reference point: the total HPI decline and unemployment increase observed during the GFC. The GFC offers a useful calibration episode because the empirical relationship between local unemployment shocks and HPI declines is strong and well-documented, regardless of which direction the causation primarily ran. The chart below shows that relationship across MSAs — a clear negative slope linking the spike in unemployment rates to the depth of the HPI decline.

Article content
Figure 3

The critical observation for our purposes is what happened in 2008: the metros most exposed to AI today (DC, Boston, San Francisco, Seattle) cluster toward the middle of that scatter plot. They experienced more moderate impacts than the hardest-hit metros, because the speculative and underwriting excesses that drove the GFC were concentrated elsewhere — Las Vegas, Phoenix, Miami, the Inland Empire. An AI-driven shock would invert that geography.

We took the GFC’s national-level shock magnitudes and redistributed them, concentrating the stress in MSAs with high AI exposure. Starting from Brookings’ occupation-level AI exposure scores rolled up to MSA and state-level, we scaled both the unemployment and HPI shocks to each MSA in proportion to its AI exposure share. The result is a per-MSA stress profile that adds up to GFC-magnitude pain at the national level but lands very differently on the map.

The chart below translates that scaling into implied HPI shock magnitudes across the metros where non-agency RMBS exposure actually sits. The metros at the bottom end of both axes (high portfolio balance and severe implied HPI shock) are where the AI thesis would hit hardest if it materializes.

Article content
Figure 4

Step 2: Run the loan universe through two scenarios.

We took the $196B universe of Non-QM mortgage loans (~500K loans across 810 active Non-QM deals as of April 2026) and ran these loans through RiskSpan’s proprietary credit model under two scenarios: a baseline and the combined AI-driven Unemployment + HPI stress scenario described above. Every loan received a projected cumulative default rate and cumulative loss rate under each scenario, isolating the AI-driven impact at the loan level.

Findings

Under the AI-driven combined stress scenario, universe-wide cumulative defaults nearly double (3.6% → 7.4%) and cumulative losses more than quadruple (0.5% → 2.2%). The steeper loss response reflects the familiar non-linearity between defaults and severities under HPI stress. The mechanism behind those aggregate numbers becomes clearer when you look at it geographically.

Finding 1 — AI exposure drives MSA-level credit performance.

This finding is by construction — we built the stress scenario by re-allocating shocks to AI-exposed MSAs, so the model output should track AI exposure. The MSA-level chart confirms it does. MSAs at the high end of AI exposure (DC, San Francisco, Boston, Seattle) see over an 8% percentage point (pp) increase in cumulative defaults compared to MSAs at the low end. Riverside, near the lower end of the AI exposure range at just under 30%, sees only a 2.6% pp increase in cumulative defaults. There’s MSA-specific noise from collateral mix, but the geographic risk concentration the AI thesis predicts shows up cleanly in the model output. This is the core mechanism: re-allocate the stress to AI-exposed geographies, and the credit impact tracks the re-allocation.

Article content
Figure 5

Finding 2 — The aggregate impact is shaped primarily by leverage.

Decomposing the universe results by current LTV bucket explains where the headline numbers come from. The first thing worth noting is what the LTV distribution looks like: less than 5% of total non-QM balance sits above 80% LTV, and roughly half is below 60%. That’s a meaningfully larger equity cushion than the borrower base of the mid-2000s, when the proliferation of low- and no-down-payment products left investors with thin protection against any HPI decline. Today’s Non-QM universe starts the AI stress scenario with a lot more skin in the game.

Within that distribution, stress impact scales monotonically with leverage, exactly as one would expect: low-LTV loans (≤40%) see a 1.3pp increase in defaults (from 0.96% to 2.24%), while high-LTV loans (>80%) see a 6.7pp increase (from 5.97% to 12.62%), though the relative impact on cumulative defaults is broadly stable across LTV buckets, with defaults about doubling in each bucket. The same pattern holds for losses, where the highest LTV buckets see losses rise by over 3pp under stress versus 0.16pp for the lowest LTV bucket. FICO declines slightly as LTV increases (from 756 for the lowest LTV bucket to 745 for the highest), so the projected cumulative default differences across LTV buckets are driven, at least partly, by underlying borrower credit as well.

Article content

Finding 3 — Deal-level concentration matters, but only in relative terms.

When we rolled up by deal rather than by MSA, the picture got more interesting. The absolute (pp) impact across deals shows essentially no relationship with the deal’s AI exposure share — high-AI-exposure deals don’t show systematically larger pp increases. But the relative impact (stress / base − 1) is clearly correlated with AI exposure.

One plausible explanation is that deal-level base default rates vary widely based on traditional credit characteristics — LTV mix, FICO mix, doc type (Full Doc vs. DSCR vs. Bank Statement), occupancy. That variation in base rates may be dominating the AI signal in absolute terms while leaving it visible on a relative basis. Either way, the relative measure is the cleaner lens for isolating AI exposure’s marginal contribution to deal-level risk.

There’s also a more reassuring story embedded in these charts, at least for securitized mortgage assets. Compare the deal-level scatter to the MSA-level scatter and you’ll notice the deal-level points cluster in a meaningfully narrower range — both on the AI exposure axis (X) and on the impact axis (Y). That tighter clustering is geographic diversification at work: even the most AI-concentrated deal in the universe pulls in loans from enough different MSAs that its blended AI exposure sits closer to the universe average than the most exposed individual MSA does. Diversification within deals is doing what it’s supposed to do — dampening the variance of outcomes — even though the unit of analysis is smaller than an MSA. The magnitude of the differential across deals is bounded by how concentrated any individual issuer is willing to get.

Article content
Article content

The implication for whole-loan portfolios is sharper. Whole-loan investors don’t get the benefit of pre-built deal diversification — they’re constructing their own geographic mix one loan or pool at a time, and concentration risk accumulates silently if it isn’t an explicit input into selection and surveillance. AI exposure is one such risk, and a relatively new one, but it sits alongside more familiar concentrations — natural disaster exposure, regional employment dependencies, single-industry metros — that warrant the same lens. The toolkit is the same in each case: granular geographic data, stress scenarios calibrated to defensible reference points, and loan-level credit modeling. Whether the goal is initial portfolio construction or ongoing surveillance of a seasoned book, the ability to identify these concentrations early — before they show up in delinquency data — is what separates active risk management from passive observation.

What This Means for Investors

History offers instructive analogies including the energy-patch metros of the mid-2010s when oil prices collapsed — Houston, Oklahoma City, and Calgary experienced home price corrections in the range of 5 to 15 percent. Regional manufacturing declines showed a similar story over a longer time period — the long contraction of auto-sector employment in Detroit, for example, played out in local housing markets over years while national HPA remained positive. In neither case did the national trend warn you that specific MSAs were in distress.

Acting now means two things in practice. The first is tracking — building the early-warning infrastructure before you need it. That means monitoring MSA-level unemployment, not just the national rate; watching labor force participation shifts; and tracking occupational mix changes in the markets where your collateral is heaviest. These indicators move before delinquency does. The second is scenario preparation — running stylized AI displacement shocks against the portfolio today using MSA-level or county-level HPA overrides and occupational exposure overlays. The goal is not a forecast but rather a map of your exposure — which corners of the portfolio are most sensitive to particular labor market stress, and what the loss distribution looks like if the more adverse scenarios begin to materialize. That map is worth building now, while the cost of being wrong about the timing is low.

RiskSpan is releasing a new Non-QM-specific credit model in the coming weeks, estimated directly on Non-QM performance data and designed around the underwriting features (DSCR, bank statement, expanded LTV) that distinguish these segments. Re-running this scenario and other stress scenarios on the new model is on the near-term roadmap and may sharpen some of the findings reported here. For investors who want to evaluate their own exposure, the analytical infrastructure described here can be applied to specific portfolios on request.

Note: The Brookings Institution has been tracking AI’s labor market impact since 2019. A 2024 Brookings piece analyzed data from OpenAI to measure occupational AI exposure as the share of tasks where AI could reduce human completion time by at least 50%. AS one would expect, the cognitive/manual divide is evident in the rankings, with computer and mathematical work atop the list at 75% while construction sits at the bottom at 5.6%.

Their 2025 piece extends that work to geography by multiplying each occupation’s exposure rating by its share of local employment in each county or metro, then aggregating. Counties with large workforces in highly exposed occupations (software developers, financial analysts, lawyers, marketing professionals) score higher than those whose employment skews toward construction, agriculture, or in-person services.

Key References:

Generative AI, the American worker, and the future of work | Brookings

The geography of generative AI’s workforce impacts will likely differ from those of previous technologies | Brookings

There Is Now Clearer Evidence AI Is Wrecking Young Americans’ Job Prospects – WSJ

www.newyorkfed.org/research/college-labor-market


AI Isn’t Coming to Structured Finance. It’s Already Here.

At SFVegas 2026, RiskSpan had a front-row seat to one of the most consequential conversations happening in our industry right now. I moderated a session on Agentic AI and the Securitization Lifecycle, while our CEO, Bernadette Kogler, participated in a panel on AI applications in structured finance. Across both rooms, the message was the same: we are done experimenting. We are redesigning workflows. 

The distinction that framed everything was simple but important. Copilots assist. Agents orchestrate. We are moving from AI that enhances what humans do to AI that executes complex, multi-step workflows — with humans supervising the process rather than driving every step. That shift changes everything about how we think about structured finance operations. 

And structured finance, it turns out, is almost perfectly suited for this moment. It is document-heavy, multi-party, process-driven, and data-intensive. From origination through surveillance, the lifecycle is fundamentally workflow-based. Panelists across both sessions shared real examples already in production: legal document review compressed from days to hours, AI-powered loan tape scrubbing before cash flow calculations, prompt-driven scenario generation replacing manual model configuration, and surveillance scaled across hundreds of deals per month. One striking observation: the future interface for structured finance may not be a UI at all — it may be entirely driven by prompts. 

But the panels were equally clear-eyed about what has to come first. Sixty-five percent of financial services firms are actively using AI, yet only 13% have deployed it across production processes. The gap between those two numbers is largely a data problem. The most sophisticated AI cannot overcome poor inputs, inconsistent loan tapes, or legacy system constraints. Firms that want to lead need to fix the foundation before layering on orchestration. 

Explainability matters just as much. With the EU AI Act and US fair lending enforcement raising the stakes, auditable, transparent models are not optional. And governance is shifting from “human in the loop” to “human over the loop” — a subtle but meaningful difference that requires defined accountability, model drift monitoring, and clear operational guardrails. 

The one-year outlook from both panels was notably concrete. Expect AI agents managing defined surveillance workflows, deeper cross-platform integration, and a sharper divide between early movers and everyone else. The competitive advantage will go to firms that clean their data, build explainability into their models from day one, and embed AI into operations — not just into pilot programs. 

Structured finance has always rewarded process discipline and deep domain expertise. That doesn’t change in an agentic world — if anything, it becomes more critical. The quality of an AI agent’s output is only as good as the prompts and parameters guiding it, and designing those well requires people who understand cash flow waterfalls, covenant structures, and credit risk at a fundamental level. Agentic AI doesn’t replace that expertise. It amplifies it. The firms that understand that will define the next chapter. 

At RiskSpan, we have spent years building at exactly this intersection — combining deep structured finance domain knowledge with purpose-built analytics infrastructure. That foundation is what makes it possible to deploy AI that actually works in production, not just in demos. The opportunity in front of our clients right now is significant, and we are focused on helping them capture it. 


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. 


The Data Model That Powers Private ABF: Why Purpose-Built Architecture Changes Everything 

Private asset-backed securities don’t follow the same rules as public securitizations. The structures are more diverse. The triggers are more nuanced. The collateral is more diverse. And yet, most market participants still try to manage these instruments with tools designed for a different world entirely—or worse, with disconnected spreadsheets that multiply risk with every manual handoff. 

RiskSpan’s Private ABF platform was built specifically to solve this problem. Not as an adaptation of existing tools, but as a purpose-built data architecture designed from the ground up for the unique demands of private structured finance. The difference isn’t incremental—it’s foundational. 

Built for the Full Lifecycle 

RiskSpan’s Private ABF platform isn’t just a database—it’s a relational architecture where every entity maintains its identity and relationships across the entire deal lifecycle. The platform connects deals to their tranches, tranches to their collateral, collateral to individual loans, and all of these to the triggers, waterfalls, fees, and reserves that govern cash flow distribution. 

This matters because private ABS transactions aren’t static. Collateral performs. Triggers trip. Waterfalls redirect. Reserves build and release. A platform that can’t maintain these relationships in real time isn’t managing deals—it’s creating snapshots that are stale before they’re finished. 

Loan-Level Depth That Powers Real Analysis 

At the heart of any ABS transaction is the underlying collateral. RiskSpan’s Private ABF platform maintains comprehensive loan-level data with over 200 attributes per asset, tracking everything from origination characteristics through current performance status. This includes credit metrics like FICO scores and debt-to-income ratios, property and collateral details, payment history and delinquency tracking, modification and loss mitigation status, and ARM reset schedules and rate mechanics. 

The platform currently manages nearly half a billion loan records across active transactions. Each loan maintains its full history—not just current state, but the complete trajectory that informs forward-looking projections. When you run a cash flow model, you’re not working from aggregated pool statistics. You’re working from the actual loans, with their actual characteristics, generating projections that reflect real collateral behavior. 

Depth for Esoteric Structures 

RiskSpan’s Private ABF platform currently supports over several collateral types and structures. Each asset class has its own performance characteristics, prepayment behaviors, and loss dynamics. The platform’s architecture accommodates this diversity without forcing artificial standardization. 

But diversity in collateral is only part of the challenge. Private ABS triggers represent some of the most complex conditional logic in structured finance. Unlike standardized agency delinquency tests, private deal triggers can involve multi-step calculations with lookback periods, cure provisions that allow temporary breaches, step-rate adjustments that phase in over time, and early amortization events with distinct severity levels. 

The platform models triggers as executable logic, not static documentation. When a trigger breaches, the system knows what happens next—which waterfall priorities shift, which reserve requirements change, which reporting obligations activate. This is the deal’s immune system, and it needs to function in real time. 

Time-Series Without the Chaos 

Every entity in RiskSpan’s Private ABF platform exists in time. The platform maintains separate performance histories for tranches, collateral pools, fees, and reserves—each keyed by reporting date to enable point-in-time reconstruction of any deal state. 

This architecture solves one of the most persistent problems in private ABF operations: answering the question “what did we know, and when did we know it?” Whether for regulatory compliance, investor reporting, or internal risk management, the ability to reconstruct historical deal states isn’t a luxury—it’s a requirement that spreadsheet-based approaches simply cannot meet reliably. 

The Waterfall as Working Code 

Cash flow waterfalls in private ABF can run to dozens of steps with conditional branches, pro-rata splits, and priority reversions. RiskSpan’s Private ABF platform models these waterfalls as executable payment sequences—not flowcharts, but actual logic that routes cash from sources to destinations based on current deal state. 

Each waterfall step defines its priority in the payment sequence, its source of funds and destination, its allocation basis (pro-rata, sequential, or targeted), and the conditions under which it activates or suspends. 

When combined with the platform’s cash flow engine, these waterfall definitions become working models. You can project payments under any scenario, stress collateral performance, and see exactly how cash moves through the structure period by period. 

An Architecture Built for AI 

The same architectural principles that make RiskSpan’s Private ABF platform effective for traditional analytics make it exceptionally well-suited for artificial intelligence and machine learning applications. This isn’t a coincidence—it’s a consequence of building a data model that prioritizes structure, relationships, and semantic clarity. 

AI systems thrive on clean, well-organized data with explicit relationships. RiskSpan’s Private ABF platform delivers exactly this: normalized entities with consistent identifiers, clear hierarchies from deals down to individual loans, and temporal versioning that distinguishes current state from historical snapshots. When an AI model needs to understand a transaction, it doesn’t have to infer structure from unstructured sources—the relationships are already defined and traversable. 

The platform’s semantic richness enables natural language interfaces that actually work. Because every field has meaning within a consistent schema, AI can translate questions like “show me deals where the senior OC trigger is within 50 basis points of breach” into precise queries without ambiguity. The loan-level depth means AI models can identify patterns across hundreds of millions of records—finding correlations between origination characteristics and performance outcomes that would be invisible to traditional analysis. 

Time-series architecture is particularly critical for AI applications. Machine learning models for credit risk, prepayment prediction, and loss forecasting require historical sequences, not point-in-time snapshots. RiskSpan’s Private ABF platform maintains this temporal context natively, enabling training datasets that capture how loans and deals evolve over their lifecycles. 

RiskSpan is already deploying AI capabilities on this foundation: automated anomaly detection that flags unusual performance patterns, intelligent document extraction that populates deal records from offering documents, natural language querying that makes complex analytics accessible to non-technical users, and predictive models that leverage the full depth of loan-level history. The platform doesn’t just store data—it organizes knowledge in a form that AI can reason about. 

Designed for the Enterprise 

RiskSpan’s Private ABF platform operates as a multi-tenant platform with granular access controls. Every record carries client and user identifiers that enable sophisticated permission model—issuers see their deals, investors see their positions, servicers see their portfolios, and rating agencies see what they’re authorized to review. 

This isn’t just about security (though it is that). It’s about enabling collaboration across the structured finance ecosystem without compromising confidentiality. A single platform can serve all participants in a transaction, each with their appropriate view of the data. 

Built for What Comes Next 

Private ABS is an evolving market. New asset classes emerge. New structures get tested. New regulatory requirements arrive. RiskSpan’s Private ABF platform accommodates this evolution through flexible schema design that allows custom attributes without requiring database modifications. 

When a client brings a novel structure—say, a securitization of an asset class we haven’t seen before—the platform can ingest and model it without waiting for a software release. This extensibility is what allows a platform to stay current with market innovation rather than constantly playing catch-up. 

What This Means in Practice 

The architectural decisions in RiskSpan’s Private ABF platform translate directly to operational capabilities. For deal structuring, you can model waterfall variations and see their impact on tranche economics before going to market. For pricing, you can run scenarios against actual loan-level collateral, not simplified pool assumptions. For risk management, you can monitor trigger proximity and project breach timing under stress. For surveillance, you can track every metric that matters, with full audit trails and historical reconstruction. 

Private ABF deserves purpose-built infrastructure. RiskSpan’s Private ABF platform delivers it. 


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. 


Use Case: RiskSpan’s Agentic AI for MBS Data Tool 

Breaking Down VA vs FHA Prepayment Speeds 

This use case demonstrates how agentic interaction with the MBS Data Tool helps investors move from a general observation (“VA speeds are fast”) to a data-backed, multidimensional explanation tied directly to credit, LTV, and program structure. 


Business Question 

Why are VA loans — particularly those with low FICO and high LTV profiles — prepaying so much faster than historically, and faster than FHA loans with comparable characteristics? 


For the full output of our AI MBS Agent, read on or watch our two-minute demo video:


Workflow with Agentic AI 

1. Baseline Comparison 

Based on the data, I can now provide a comprehensive analysis of VA vs FHA loan prepayment trends since 2020:

Key Findings:

1. Current Prepayment Speeds (as of September 2025):

– VA loans: 11.74% CPR

– FHA loans: 8.68% CPR

– VA loans are currently prepaying about 3 percentage points faster than FHA loans

2. Historical Trends:

2020-2021 (Refinance Wave):

– Both loan types experienced very high prepayment speeds

– Peak CPRs:

* FHA: 38.48% (April 2021)

* VA: ~30-31% (Late 2021)

– VA loans consistently showed higher prepayment sensitivity during this period

2022-2023 (Rising Rate Environment):

– Both loan types saw significant slowdown in prepayment speeds

– Lowest CPRs:

* FHA: 4.42% (February 2023)

* VA: 5.07% (February 2023)

– The spread between VA and FHA prepayment speeds narrowed but remained positive

2024-2025 (Current Environment):

– Moderate prepayment speeds for both loan types

– VA loans maintain faster prepayment speeds

– Recent range:

* FHA: 5-10% CPR

* VA: 7-14% CPR

3. Key Observations:

– VA loans consistently prepay faster than FHA loans across different rate environments

– The spread between VA and FHA prepayment speeds tends to widen in more favorable refinancing environments

– Both loan types show similar seasonal patterns but with different magnitudes

– VA loans show greater prepayment sensitivity to market conditions

4. Portfolio Sizes:

– FHA: Currently ~7.45 million loans with $1.47 trillion in balance

– VA: Currently ~3.77 million loans with $1.03 trillion in balance

This analysis suggests that investors should expect consistently faster prepayment speeds from VA loans compared to FHA loans, particularly during periods of favorable refinancing opportunities. The higher prepayment sensitivity of VA loans likely reflects the generally better credit quality of VA borrowers and their greater ability to refinance when rates are favorable.


2. High-Credit Borrowers 

Based on the analysis of high-FICO (740+) VA and FHA loans, here are the key findings:

1. Current Prepayment Speeds (September 2025):

– High-FICO VA: 9.37% CPR

– High-FICO FHA: 5.96% CPR

– Spread: 3.41% faster for VA

2. Portfolio Characteristics (as of September 2025):

High-FICO VA Loans:

– Average FICO: 779

– Average WAC: 3.85%

– Average CLTV: 58.1%

– Balance: $443.4B (43% of VA portfolio)

– Loan Count: 1.47M loans

High-FICO FHA Loans:

– Average FICO: 769

– Average WAC: 4.55%

– Average CLTV: 58.4%

– Balance: $188.0B (13.8% of FHA portfolio)

– Loan Count: 879K loans

3. Key Trends (2024-2025):

Prepayment Speed Range:

– High-FICO VA: 4.26% to 14.76% CPR

– High-FICO FHA: 3.56% to 8.68% CPR

Seasonal Patterns:

– Both show seasonal increases in spring/summer months

– VA loans show more pronounced seasonal effects

– Peak prepayment typically occurs in October/November

4. Notable Observations:

– The prepayment speed differential between VA and FHA remains significant even at high FICO levels

– High-FICO VA loans have consistently lower WACs (about 70bps lower)

– Similar CLTV profiles between high-FICO VA and FHA

– VA has a much larger proportion of high-FICO borrowers (43% vs 13.8% for FHA)

– High-FICO VA loans show greater prepayment sensitivity to market conditions

5. Recent Trends:

– Both loan types show increasing prepayment speeds in 2025

– The spread between VA and FHA prepayment speeds has remained relatively stable

– High-FICO VA loans maintain faster prepayment speeds even in rising rate environments

This analysis suggests that even when controlling for credit quality (FICO ≥ 740), VA loans consistently prepay faster than FHA loans. This is likely due to:

1. Lower average note rates (WAC)

2. VA’s no-down-payment benefit making it easier to refinance

3. Generally better borrower demographics and income profiles

4. More streamlined refinancing options for VA loans

Investors should expect continued faster prepayment speeds from high-FICO VA loans compared to high-FICO FHA loans, particularly during favorable rate environments or seasonal peaks.


3. Low-Credit Borrowers 

Based on the analysis of low-FICO (<640) VA and FHA loans, here are the key findings:

1. Current Prepayment Speeds (September 2025):

– Low-FICO VA: 21.04% CPR

– Low-FICO FHA: 12.07% CPR

– Spread: 8.97% faster for VA

2. Portfolio Characteristics (as of September 2025):

Low-FICO VA Loans:

– Average FICO: 611 – Average WAC: 4.77%

– Average CLTV: 66.2%

– Average Age: 52.3 months

– Balance: $77.5B (8.3% of VA portfolio)

– Loan Count: 316K loans

Low-FICO FHA Loans:

– Average FICO: 609

– Average WAC: 4.83%

– Average CLTV: 60.4%

– Average Age: 62.8 months

– Balance: $290.7B (21.3% of FHA portfolio)

– Loan Count: 1.49M loans

3. Key Trends (2024-2025):

Prepayment Speed Range:

– Low-FICO VA: 9.08% to 27.54% CPR

– Low-FICO FHA: 6.31% to 12.73% CPR

Notable Patterns:

– Much higher volatility in VA prepayment speeds

– Larger seasonal effects in VA loans

– Consistently wider spread between VA and FHA vs high-FICO segment

4. Key Differences:

Portfolio Composition:

– FHA has a much larger concentration in low-FICO loans (21.3% vs 8.3% for VA)

– FHA’s low-FICO portfolio is nearly 4x larger than VA’s by balance

Performance Characteristics:

– VA loans show significantly higher prepayment speeds despite similar credit profiles

– VA loans exhibit greater prepayment volatility

– VA loans have slightly better average FICO scores within the low-FICO segment

– VA loans have higher CLTVs but lower WACs

5. Recent Trends:

– Both loan types show increasing prepayment speeds in 2025

– VA prepayment speeds are consistently higher and more volatile

– Seasonal patterns are more pronounced in VA loans

– The spread between VA and FHA prepayment speeds has remained wide

6. Notable Observations:

– The prepayment speed differential between VA and FHA is actually larger in the low-FICO segment than in the high-FICO segment

– Low-FICO VA loans show surprisingly high prepayment speeds despite lower credit scores

– FHA loans show more consistent, less volatile prepayment patterns

– VA loans maintain faster speeds despite higher CLTVs

This analysis reveals that the prepayment behavior difference between VA and FHA loans is even more pronounced in the low-FICO segment than in the high-FICO segment. This could be due to:

1. VA’s stronger refinancing program benefits for lower-credit borrowers

2. More flexible VA streamline refinance requirements

3. Better servicing support for VA borrowers

4. Lower average note rates in the VA program

5. Potentially different borrower demographics within the same FICO band

Investors should expect:

– Higher and more volatile prepayment speeds from low-FICO VA loans

– More stable, lower prepayment speeds from low-FICO FHA loans

– Larger seasonal effects in VA loans

– Greater market sensitivity in VA loans


4. High-LTV Loans 

Based on the analysis of high-LTV (≥95%) VA and FHA loans, here are the key findings:

1. Current Prepayment Speeds (September 2025):

– High-LTV VA: 12.37% CPR

– High-LTV FHA: 7.77% CPR

– Spread: 4.60% faster for VA

2. Portfolio Characteristics (as of September 2025):

High-LTV VA Loans:

– Average FICO: 714

– Average WAC: 4.33%

– Average CLTV: 70.9%

– Average Original LTV: 100.3%

– Average Age: 52.9 months

– Balance: $607.2B (61.1% of VA portfolio)

– Loan Count: 2.11M loans

High-LTV FHA Loans:

– Average FICO: 679

– Average WAC: 4.61%

– Average CLTV: 62.9%

– Average Original LTV: 97.6%

– Average Age: 71.7 months

– Balance: $1.04T (72.0% of FHA portfolio)

– Loan Count: 5.08M loans

3. Key Trends (2024-2025):

Prepayment Speed Range:

– High-LTV VA: 5.60% to 19.73% CPR

– High-LTV FHA: 4.50% to 8.89% CPR

Notable Patterns:

– Higher volatility in VA prepayment speeds

– Strong seasonal patterns in VA prepayments

– More stable FHA prepayment patterns

– Consistent spread between VA and FHA

4. Key Differences:

Portfolio Composition:

– Both programs have significant high-LTV concentration

– FHA has larger absolute exposure but similar relative exposure

– VA allows higher original LTV (100%+) compared to FHA

Performance Characteristics:

– VA loans consistently prepay faster despite higher LTVs

– VA borrowers have notably higher FICO scores (+35 points)

– VA loans have lower WACs (-28bps)

– VA loans are significantly younger (-19 months)

5. Recent Trends:

– Both loan types show increasing prepayment speeds in 2025

– VA prepayment speeds remain more volatile and responsive to seasonality

– Peak prepayment periods occur in spring/summer months

– The spread between VA and FHA has remained relatively stable

6. Notable Observations:

– High-LTV concentration is a defining characteristic of both programs

– VA’s higher prepayment speeds persist despite higher LTVs

– Better credit quality of VA borrowers likely contributes to faster speeds

– Younger average loan age in VA portfolio may influence prepayment behavior

– Lower WACs in VA portfolio suggest more efficient refinancing history

Implications for Investors:

1. Portfolio Management:

– Expect faster and more volatile prepayments from high-LTV VA loans

– Consider seasonal patterns in VA prepayments for timing decisions

– Factor in the significant WAC differential when evaluating prepayment risk

2. Risk Considerations:

– Higher prepayment risk in VA despite similar LTV profiles

– Greater market sensitivity in VA portfolio

– More stable, predictable prepayment patterns in FHA

3. Valuation Impact:

– Need to apply larger prepayment speed differentials for VA vs FHA

– Consider stronger seasonal effects in VA valuations

– Account for credit quality differences within high-LTV segment

The analysis suggests that even in the high-LTV segment, VA loans maintain significantly faster prepayment speeds than FHA loans, driven by:

1. Better borrower credit quality

2. Lower note rates

3. More flexible refinancing options

4. Younger loan age

5. More efficient VA refinancing program

These factors appear to overcome the theoretical constraints of high LTV ratios on refinancing ability.


5. Program Features as Drivers

The data strongly supports that program features drive faster VA prepayment speeds. Here’s the detailed analysis:

1. Early Prepayment Behavior (0-24 months):

– VA loans age 13-24 months: 25.29% CPR

– FHA loans age 13-24 months: 17.04% CPR

– The 8.25% difference reflects VA’s more flexible refinancing rules

– No seasoning requirement for VA vs 210 days for FHA

2. Refinance Incentive Response:

– High incentive (>100bp) VA: 50.43% CPR

– High incentive (>100bp) FHA: 26.55% CPR

– The dramatic difference (23.88%) shows VA borrowers can more easily act on rate incentives

– Even medium incentive (0-100bp) shows large gap:

* VA: 29.59% CPR

* FHA: 18.37% CPR

3. Credit Quality Impact:

– VA average FICO: 715-724 across age buckets

– FHA average FICO: 672-681 across age buckets

– Better VA credit quality + easier program rules = faster speeds

4. Cost Structure Effects:

VA Advantages:

– Lower funding fee (0.5%)

– No ongoing MIP

– Can include closing costs

– Results in lower all-in refinancing costs FHA Constraints:

– Higher upfront MIP (1.75%)

– Ongoing annual MIP

– More restricted closing cost inclusion

– Higher total refinancing costs

5. LTV Considerations:

– VA has no maximum LTV for IRRRL

– FHA generally limited to 97.75% LTV

– Current data shows:

* VA high-LTV loans (>95%): 12.37% CPR

* FHA high-LTV loans (>95%): 7.77% CPR

6. Program Flexibility Evidence:

Response to Rate Environment:

– VA loans show stronger reaction to refi incentives

– Higher CPRs across all incentive buckets

– Faster response to market opportunities

Seasoning Effects:

– VA shows earlier prepayment spikes

– More pronounced early-age prepayment behavior

– Less constrained by seasoning rules

7. Servicer Impact:

– VA servicers may be more proactive due to program ease

– Lower documentation requirements reduce operational friction

– Simpler approval process enables faster execution

The data demonstrates that VA’s program features create a more efficient refinancing mechanism:

1. Lower Costs:

– Reduced fees

– No ongoing MIP

– More included costs

2. Fewer Restrictions:

– No LTV limits

– No seasoning requirement

– Limited credit checks

3. Simpler Process:

– Less documentation

– Faster approval

– Lower operational burden

These advantages manifest in:

– Consistently faster prepayment speeds

– Stronger response to rate incentives

– Earlier refinancing activity

– Higher refinancing success rates

For investors, this means:

1. Expect faster VA prepayments across market cycles

2. Anticipate stronger response to refinancing opportunities

3. Factor in program differences when valuing VA vs FHA MBS

4. Consider seasoning rules when evaluating young loans

The combination of program features and data evidence clearly shows that VA’s more flexible refinancing rules are a primary driver of faster prepayment speeds compared to FHA.


6. Scenario Stress Test 

Based on the current portfolio characteristics and historical refinancing behavior, here’s how a 100bps rate drop would likely impact VA vs FHA prepayment speeds:

1. Current Portfolio Rate Distribution:

VA Portfolio:

– 7.5%+ WAC: $4.6B (0.4%)

– 7.0-7.49% WAC: $18.5B (1.8%)

– 6.5-6.99% WAC: $71.1B (6.9%)

– 6.0-6.49% WAC: $105.3B (10.2%)

– 5.5-5.99% WAC: $118.7B (11.5%)

– <5.5% WAC: $713.0B (69.2%)

FHA Portfolio:

– 7.5%+ WAC: $23.6B (1.6%)

– 7.0-7.49% WAC: $68.5B (4.7%)

– 6.5-6.99% WAC: $163.8B (11.2%)

– 6.0-6.49% WAC: $158.4B (10.8%)

– 5.5-5.99% WAC: $149.0B (10.2%)

– <5.5% WAC: $903.5B (61.5%)

2. Projected Prepayment Speeds by Refi Incentive:

Based on current behavior with similar incentives:

High Incentive (>100bps):

– VA: 50-55% CPR

– FHA: 25-27% CPR

– Differential: ~25-28% faster for VA

Medium Incentive (50-100bps):

– VA: 35-40% CPR

– FHA: 20-25% CPR

– Differential: ~15% faster for VA

Low Incentive (0-50bps):

– VA: 25-30% CPR – FHA: 15-18% CPR

– Differential: ~10-12% faster for VA

3. Portfolio Impact of 100bps Rate Drop:

VA Portfolio:

– ~$200B would move to >100bps incentive

– ~$119B would move to 50-100bps incentive

– Weighted average CPR projection: 32-37% CPR

FHA Portfolio:

– ~$256B would move to >100bps incentive

– ~$149B would move to 50-100bps incentive

– Weighted average CPR projection: 18-22% CPR

4. Key Factors Driving Divergence:

Program Features:

– VA IRRRL has no LTV restrictions

– VA has lower funding fees

– VA has simpler documentation requirements

– VA has no seasoning requirement

Borrower Characteristics:

– VA has higher average FICOs (714 vs 675)

– VA has lower average WACs

– VA has younger average loan age

5. Projected Timeline of Response:

Months 1-3:

– VA speeds increase rapidly to 30-35% CPR

– FHA speeds lag, reaching 15-18% CPR

– Differential: 15-17% CPR

Months 4-6:

– VA peaks at 35-40% CPR – FHA reaches 20-22% CPR

– Differential: 15-18% CPR

Months 7-12:

– VA moderates to 30-35% CPR

– FHA stabilizes at 18-20% CPR

– Differential: 12-15% CPR

6. Risk Factors to Consider:

Upside Risks (faster speeds):

– Pent-up refinance demand

– Improved technology/processing

– Competitive lender environment

Downside Risks (slower speeds):

– Housing market uncertainty

– Credit tightening

– Operational capacity constraints

7. Investment Implications:

1. Premium Pools:

– Expect significantly faster VA prepayments

– Greater price impact on VA premium pools

– Wider VA-FHA OAS spreads

2. Current Coupon Pools:

– Moderate prepayment differential

– Less price impact

– Smaller OAS spread widening

3. Discount Pools:

– Minimal impact on speeds

– Limited price effect

– Stable OAS spreads

Recommendations:

1. Portfolio Positioning:

– Reduce exposure to premium VA pools

– Consider FHA alternatives in premium segment

– Maintain VA exposure in current coupon/discount segments

2. Risk Management:

– Increase prepayment protection in VA holdings

– Consider specified pool stories

– Monitor servicer behavior closely

3. Relative Value:

– Look for opportunities in FHA premium pools

– Consider VA-FHA switches in premium segment

– Evaluate specified pool pay-ups

The analysis suggests a significant divergence in VA vs FHA prepayment speeds following a 100bps rate drop, with VA speeds likely to be 12-25% CPR faster depending on the segment and timing. This differential is driven by both program features and borrower characteristics, with the impact most pronounced in premium pools.


Questions or Need Support? 

Our team is standing by to help you make the most of these new features. Contact us at techsupport@riskspan.com or reach out to your dedicated account manager. 


Interested in a Complimentary Trial or Demo?

Contact us here.


RiskSpan Launches Agentic AI for MBS Data — Instant, Transparent Insights for Agency MBS Professionals 

Arlington, VA – September 2, 2025 RiskSpan, a leading provider of data analytics solutions for the structured finance industry, today announced the release of Agentic AI for MBS Data.

Watch a one-minute demo video or read on to learn about our new, AI-powered MBS analyst that delivers instant, sourced, and context-rich analysis in plain English.

Solving the MBS Data Bottleneck

While Agency MBS loan data is readily available, meaningful insights for many users often remain locked behind complex SQL or Python queries. This creates a barrier for traders, portfolio managers, and risk teams who need answers quickly to avoid missing opportunities. 

“Our clients told us they needed faster, clearer access to this data,” said Suhrud Dagli, RiskSpan’s Co-Founder and Chief Innovation Officer. “Agentic AI removes the technical barrier and delivers the kind of immediate, transparent insight that our clients have come to demand.” 

A 24/7 AI-Powered MBS Analyst 

With Agentic AI for MBS Data, users simply ask a question in plain English — such as “Which Ginnie Mae 2020 production pools had the fastest speeds in Q2?” — and get back clear, visualized, and fully-cited answers in seconds. Ask Complex Questions, Get Instant Answers – No SQL or Python required. 

Key benefits include: 

  • Full Transparency – All sources cited for confidence and compliance. 
  • Deeper Insights – Surfaces patterns and drivers you might not think to look for. 
  • Narrative-Ready Reporting – Polished summaries and visuals for executives and clients. 
  • Efficiency – Eliminates dependence on technical teams for everyday queries. 

Designed for How MBS Professionals Think 

From spotting prepayment anomalies to explaining performance shifts, Agentic AI identifies not just the what but the why. It transforms hours of manual analysis into seconds of actionable insight, enabling faster, more informed decision-making. 

Agentic AI for MBS Data is available now. RiskSpan is offering live demos to showcase how the platform transforms the way MBS data is interrogated and reported. 


About RiskSpan 

RiskSpan delivers a single analytics solution for structured finance and private credit investors of any size to confidently make faster, more precise trading and portfolio risk decisions and meet reporting requirements with fewer resources, and less time spent managing multiple vendors and internal solutions. 

Learn more at www.riskspan.com.  


Get Started
Log in

Linkedin   

risktech2024