What this article covers
Connecting dv01 to an AI assistant is the easy part. Getting reliable answers out of it is a skill, and it's mostly about understanding what sits behind the connection. This article covers how to ask for dv01 data well: what the connection actually exposes, how to frame a request so it returns what you meant, how to refine it, and how to check the answer before you put it in front of anyone.
Everything here applies to any MCP-compatible client. MCP is an open protocol, so the same dv01 connection works across assistants — the wording of your request matters, the choice of client does not. Setup steps differ by client and are covered in Connecting to dv01 via MCP.
What you're actually querying
The connection gives your assistant four distinct capabilities. Almost every good request is a combination of them, and knowing which one you're invoking is the difference between a precise answer and a vague one.
| Capability | What it does | Ask for it like this |
|---|---|---|
| Discovery | Finds which deals, benchmarks, and platform aggregates exist in your organization — and which ones you're entitled to see. | “What non-QM deals do I have access to?” |
| Metadata | Tells you what's available for a given entity: which metrics, which dimensions you can stratify by, and the date range covered. | “What can I stratify this benchmark by, and how far back does it go?” |
| Performance data | The main event: aggregated loan performance. You pick an entity, some metrics, an optional breakdown, and a date. | “Pull 60+ DQ and CPR for this deal, by month, for the last year.” |
| Cashflow projections | Forward-looking collateral and structure projections under scenarios you specify. Runs as a job, taking longer than a data pull. | “Project collateral cashflows for this deal at 15 CPR / 2 CDR.” |
The practical upshot: when you're unsure what's possible, ask a discovery or metadata question first instead of guessing at a data question. “What dimensions can I break this benchmark out by?” costs you one exchange and prevents a whole conversation spent guessing at dimension names that don't exist.
The five-part request
A performance-data request has five slots. Leave one empty and the assistant will fill it with an assumption — sometimes the right one, sometimes not. Filling all five is the single highest-leverage habit in this article.
| Slot | What it answers | If you leave it out |
|---|---|---|
| Entity | Which deal, benchmark, or platform aggregate? | You get a name-matching guess, or a clarifying question. |
| Metrics | Which measures — DQ, CPR, CDR, losses? | You get a generic default set, rarely the one you wanted. |
| Breakdown | One row, or split by FICO / LTV / vintage / state? | You get a single pool-level row. |
| Time | A point in time, or a series over a window? | You get the latest available month — which may differ by entity. |
| Output | Inline table, chart, Excel, slide? | You get an inline table and re-ask for the file. |
All five, filled in:
“For the Fitch-dv01 Non-QM Benchmark, show 60+ DQ, CDR, and current LTV, broken out by current FICO bucket and weighted by balance, as of July 2026 — as a table, then export it to Excel.”
Two choices that change the answer
These two are worth understanding properly, because getting them wrong produces numbers that look plausible and are quietly misleading.
Calendar month vs. loan age
Calendar month asks “what happened in July?” It's the right lens for surveillance and for anything tied to the market environment — rates, seasonality, macro.
Loan age asks “how does this pool behave 18 months in?” It aligns every vintage to a common starting point, which is the only fair way to compare a 2023 cohort against a 2025 one. Use it for loss curves and any vintage-versus-vintage question.
Comparing cumulative losses across vintages on a calendar axis is the most common analytical mistake made through this connection — the older vintage always looks worse, purely because it's had longer to accumulate. Say “by loan age” explicitly.
Weighted vs. unweighted
Ask for “weighted by balance” whenever you want a figure that represents the pool. An unweighted average treats a $2 million loan and a $60,000 loan as equals, which is almost never what you mean for a pool-level metric like WAC, LTV, or FICO. State it — don't assume the default matches your intent.
Related: cumulative metrics (cumulative loss, cum net loss to liquidation) are point-in-time to your as-of date, not to a fixed window. If you're reconciling against a figure from another system, confirm both are on the same as-of and the same time basis before concluding anything.
Start broad, then narrow
The most effective pattern is not one perfect request — it's a short chain of them. Each exchange keeps the context of the last, so refinements are cheap and you should use them freely.
Anchor to a benchmark first. The Fitch-dv01 and dv01 benchmarks are curated, sector-wide aggregates. Reading your deal against a sector baseline tells you something; reading it in isolation mostly doesn't. A typical session:
| Step | What you ask | Why |
|---|---|---|
| 1. Orient | “Compare the four Fitch-dv01 mortgage benchmarks on 30+/60+/90+ DQ, CPR, CDR, and current LTV.” | One table, whole sector. Tells you where the stress is before you commit to a direction. |
| 2. Follow the signal | “Non-QM DQ looks elevated. Break that benchmark out by current FICO bucket, weighted by balance.” | Finds out whether it's broad-based or concentrated in one cohort. |
| 3. Add time | “Now show the bottom two FICO buckets by month for the last 18 months and chart it.” | Distinguishes a deteriorating trend from a one-month print. |
| 4. Bring it home | “Add my deal to that comparison on the same metrics and as-of.” | The actual question — but now with a sector baseline behind it. |
| 5. Package | “Put steps 1 and 4 into Excel, one tab each, with the as-of in the header.” | Export once, at the end, when you know what's worth keeping. |
What step 1 returns, from the dv01 Demo Org:
| Benchmark | 30+ DQ | 60+ DQ | 90+ DQ | CPR | CDR | Cur LTV |
|---|---|---|---|---|---|---|
| Fitch-dv01 Prime Jumbo | 1.14% | 0.62% | 0.47% | 10.48% | 0.02% | 64.31% |
| Fitch-dv01 Non-QM | 5.32% | 3.32% | 2.55% | 14.14% | 0.20% | 68.37% |
| Fitch-dv01 Closed-End 2nds | 1.87% | 0.85% | 0.61% | 23.36% | 0.02% | 19.43% |
| Fitch-dv01 HELOC | 2.16% | 1.12% | 0.80% | 21.20% | 0.59% | 27.39% |
dv01 Demo Org, as of 7/1/2026. Rates are returned as decimals — ask for percentage formatting if you want it presented this way.
Six patterns that cover most work
Nearly every request people make is one of these six shapes. Learn the shape, then substitute freely — that's more useful than memorizing example prompts.
| Pattern | Template | Use it when |
|---|---|---|
| Snapshot | “For [entity], pull the latest [metrics].” | You need one current read to anchor on. |
| Comparison | “Compare [entities] on [metrics] as of [date], ranked by [metric].” | Relative performance — deal vs. sector, platform vs. platform. |
| Stratification | “Break [entity] out by [dimension] showing [metrics], weighted by balance.” | A pool-level number moved and you need to know which cohort drove it. |
| Trend | “Show [metric] for [entity] by month over [window] and chart it.” | Separating a trajectory from a single noisy print. |
| Curve by age | “Plot [cumulative metric] by loan age for [entities].” | Comparing vintages or cohorts fairly. |
| Projection | “Project cashflows for [deal] under [scenario].” | Forward-looking work. State your assumptions explicitly. |
Any of the first five can be closed with “…and put that in Excel” or “…make it a slide.”
Where to point them
Benchmarks and metrics by asset class. Coverage also extends to credit card, student loans, small business, home efficiency, and point of sale — ask a discovery question to see what your organization is entitled to.
| Asset class | Benchmarks & aggregates | Metrics that matter most | The question people actually ask |
|---|---|---|---|
| Mortgage | Fitch-dv01 Prime Jumbo, Non-QM, Closed-End Seconds, HELOC. Deal-level: STACR / CAS shelves and many more. | DQ pipeline, CPR, CDR, current and original LTV | Is my deal's delinquency pipeline building faster than its sector — and is it concentrated in a FICO or LTV cohort? |
| Consumer | Consumer Unsecured Benchmark. Platform aggregates: LC Standard, SoFi PL, Marlette, Avant, Prosper, Upgrade. | Monthly charge-off %, DQ pipeline, CDR, cumulative loss | How do this platform's charge-offs compare to the market, and are newer vintages tracking worse at the same loan age? |
| Auto | Auto Benchmark – Prime, Auto Benchmark – Subprime. | Cum net loss to liquidation, DQ pipeline, monthly charge-off % | How wide is the prime/subprime gap right now, and is the subprime DQ pipeline pointing to worse losses ahead? |
Note on prime-leaning vs. deeper-credit platforms: LC Standard and SoFi PL sit at the prime end with the lowest charge-offs; Marlette and Prosper are near-prime and mid-pack; Avant and Upgrade are the deeper-credit, higher-yield end. Picking a comparison set from a single tier is usually more informative than ranking all six at once.
Weak request, strong request
Each pair below asks roughly the same thing. The right-hand version is the one that comes back usable on the first try.
| Weak | Strong | What changed |
|---|---|---|
| “How's non-QM doing?” | “For the Fitch-dv01 Non-QM Benchmark, pull 30+/60+/90+ DQ, CPR, and CDR as of July 2026.” | Named the entity, the metrics, and the date instead of leaving all three to inference. |
| “Show me losses by vintage.” | “Plot cumulative net loss by loan age for the 2023, 2024, and 2025 vintages.” | Switched to loan age, so older vintages aren't penalized for seasoning. |
| “Average FICO for the pool?” | “Balance-weighted average updated FICO for the pool, as of the latest month.” | Specified weighting and which FICO — original and updated tell different stories. |
| “Compare my deals to the benchmark.” | “Compare [deal A] and [deal B] to the Non-QM Benchmark on 60+ DQ and CDR, all as of June 2026.” | Pinned one shared as-of, so the comparison isn't mixing reporting months. |
| “Give me everything on this deal.” | “What metrics and dimensions are available for this deal, and what's the date range?” | Asked a metadata question first, rather than triggering an enormous pull. |
Check the answer before you use it
An assistant will present a partial result with exactly the same confidence as a complete one. Three habits catch nearly everything:
- Make it restate the parameters. “Before we go further — which entity, which as-of date, and how many rows did that return?” Cheap, and it surfaces a wrong entity or a stale month immediately.
- Watch for silent truncation. Large pulls can hit a row limit and return a clean-looking table that isn't the whole picture. If a strat's balances don't sum to the pool, or a series has suspiciously round row counts, ask whether the result was capped and narrow the request.
- Sanity-check one number. For anything going into a deck or a memo, verify a single figure against the dv01 platform UI. If it ties, the rest of the pull almost certainly does too.
One more: your organization context persists through a conversation. If you switch orgs partway through, earlier results in that same thread came from the previous org. When it matters, start a fresh conversation rather than switching mid-stream.
Troubleshooting
| What you see | What to do |
|---|---|
| It can't find your deal | Ask it to search rather than guessing the name: “search for deals matching VERUS 2024.” |
| A deal is listed but returns no data | Usually an entitlement boundary, not an error — a name can appear in a catalog while its performance data sits outside your organization's access. Confirm the active org, and contact dv01 if you expected coverage. |
| Results look truncated or totals don't tie | Narrow the request — fewer dimensions, shorter window — and ask explicitly whether the result was complete. |
| Numbers come back as decimals | Rates are returned as decimals (0.0532). Ask for them “formatted as percentages to two decimals.” |
| The as-of date isn't what you expected | Reporting months vary by asset class and servicer. Ask for the latest available as-of before building a comparison, then pin that date across every entity in it. |
| A cashflow request is slow | Projections run as a job — submitted, then polled. Let it finish rather than resending; a duplicate request starts a second job. |
| The connection isn't responding at all | Re-check the connector in your client's settings and re-authenticate if prompted. See Connecting to dv01 via MCP. |
Terms worth pinning down
This is not a catalog of available metrics — there are far more than could usefully be listed here, and the set grows. To find out what a given entity actually supports, ask a metadata question: “What metrics and dimensions are available for this benchmark?” That answer is always current; a list in an article isn't.
What's worth stating here are the few conventions that differ between data providers. If you're reconciling against another system, or reading a number into a conclusion, check these first.
| Term | The convention here |
|---|---|
| CPR and CDR | Both are annualized conditional rates, built from an underlying monthly rate the same way. Don't read either against a monthly figure such as monthly charge-off % without converting — a 6.3% annualized CDR is roughly a 0.54% monthly default rate, not a 6.3% one. |
| Delinquency (30+/60+/90+) | Share of current balance, not loan count. A balance-based and a count-based DQ can differ materially on the same pool. |
| Cumulative net loss vs. cum net loss to liquidation | Different denominators. The first is against original balance; the second against liquidated balance, which is the standard auto loss-severity lens. They are not interchangeable and will differ by a wide margin. |
| Cumulative anything | Point-in-time to your as-of date. Comparing cumulative figures across entities only means something if the as-of is pinned and, for vintage work, if you're on a loan-age axis. |
| Original vs. updated FICO | Two different fields telling two different stories — credit at underwriting vs. credit now. Say which one you want; the gap between them is often the actual finding. |
Figures shown are from the dv01 Demo Org as of 7/1/2026 and are illustrative — values in your own organization may differ.