Defi Credit Indices

There's a strange gap in finance. Consumer credit drives a huge share of the economy, and when it turns, it turns quickly. Yet the tools most people use to watch it are surprisingly blunt. The headline delinquency figures come out on a lag, lumped together at a level that hides what's actually happening underneath, and shaped by whatever choices the institution publishing them happened to make. If you want to know how subprime borrowers are really doing this month, you're mostly guessing.

The frustrating part is that the data already exists. Every loan inside an asset-backed security reports its status every month: what got paid, what's late, what charged off, what was recovered. Millions of records, and all of it public. It just isn't usable in any practical sense. It's scattered across thousands of filings, a real chore to pull together, and it's never expressed as a number you could glance at, never mind trade on.

So that's what Serention does. We take the raw loan-level data and build indices out of it. The idea is straightforward: instead of repackaging somebody else's average, compute the index directly from the loans themselves, using a fixed, published rule for how they're combined. Because it's mechanical and built on the primary data, you can see exactly what it measures and why it moved. There's no black box to take on faith. And because it's built from the bottom up, it picks up things the aggregates miss. Stress tends to show up in the loan tape well before it shows up in a quarterly press release.

One thing we were careful about. An index like this can quietly mislead you. As any pool of loans ages, delinquencies naturally drift higher, so a naive measure can look like it's flashing red when really the loans just got older. Serention takes that aging effect out. When the index climbs, it's because borrowers are genuinely getting weaker, not because the math is drifting under you.

The first index is the Auto Subprime Credit Index, and we started there on purpose. Subprime auto is big, it's deeply securitized, and it's one of the earliest places household stress turns up. People tend to fall behind on the car loan after the credit card but before the mortgage, and the subprime end of the market moves first of all. The index pulls the performance of well over a million loans into a single monthly reading. When we checked it against the Federal Reserve's own auto-credit series, it lined up closely. Close enough to trust the method, while showing detail those series can't.

Auto subprime is the starting point, not the whole plan. The same approach carries over to the rest of consumer credit: prime auto, auto leases, credit cards, and on from there. Each one is its own benchmark for its own slice of borrowing, built the same transparent way, so they stay comparable and they mean what they say. Over time that's the actual product. Not any single index, but a consistent set of them that gives the market a common language for where consumer credit stands.

Here's the part that makes this more than a dashboard. The indices are meant to be traded. Each one is published in an open form that other contracts can read as a price, and on top of that you can take a position: go long or short the direction of credit, post stablecoin collateral, and settle against the index automatically. That's something you basically can't do today. You can buy a particular bond, of course, but you can't cleanly take a view on consumer credit itself, separate from any one issuer, and adjust it whenever you want.

The uses are pretty concrete. If you're a lender, or you hold a book of consumer-credit risk, you can finally hedge it. When losses climb, a short position in the right index pays off and offsets them. If you think the consumer is stretched too thin, or quietly getting healthier, you can put that view on directly, with risk you can size, instead of cobbling together a basket of bonds and hoping it tracks. And once people start taking positions, the market does something useful on its own: the price begins to say where credit is heading, not just where it's already been.

There's also a simpler reason to care about how this is built. Because the whole thing is transparent and spelled out, the indices can become shared reference points rather than proprietary numbers a handful of desks pay for and the rest of us squint at. A risk manager, a macro trader, a researcher, and a regulator can all look at the same index and know it means the same thing. That kind of common ground is rare in credit, and it's worth a lot.

None of this depends on inventing new data. The information has been sitting in plain sight for years. The work is in turning it into something honest and readable, and then into something you can actually act on. That's the whole job. We began with subprime auto because that's where the cracks tend to appear first, and we're building out from there.

Serention turns SEC ABS-EE regulatory filings — the loan-by-loan tapes every auto-ABS issuer must file — into a tradeable subprime-auto credit index.

Data pipeline (Python/pandas): a custom parser ingests the raw ABS-EE XML tapes, builds a universe of qualifying subprime trusts (issuance-weighted FICO < 640, >$200M pools, staggered 2017–2024+ vintages), and pools them loan-level into monthly metrics. We stream 32.9M loan-months into Supabase Postgres with a disk-safe download→COPY→delete-per-trust loader that fits the whole panel on an 8 GB laptop, plus a persisted loan_month_agg summary table so the index recomputes in milliseconds instead of re-scanning 33M rows.

Index: two flavors off the same tapes — a VIX-style stress index (rolling-24-month z-score composite of 30+/60+ DPD, roll rate, net loss, recovery) and the tradeable net-yield and gross-yield indexes (interest collected − net losses or not, the economic carry).

Website: a dependency-free static site (vanilla JS + Chart.js) on Vercel at serention.com — live index, methodology, a client-side backtest module, and a comparison page layering z-score, PCA, and net-yield constructions.

On-chain (Solidity + Foundry): a margined, cash-settled total-return derivative on Sepolia (MarginedIndex + MockUSDC) reading a Chainlink AggregatorV3Interface-standard oracle, so any Chainlink-style feed is drop-in. Canton/Daml: a privacy-first version where positions are visible only to the trader and operator, validated on a local Canton ledger.

Partner tech: Supabase (Postgres) is the backbone — it holds the 33M-row loan-level panel and the aggregate that makes the index queryable; Chainlink's AggregatorV3 standard future-proofs the oracle; Vercel ships the static front-end.

Notably hacky: (1) testnet Chainlink Functions sunset (June 2, 2026) mid-build, so we kept the Chainlink AggregatorV3 oracle but fed it via an authenticated off-chain pusher, designed to swap to Chainlink CRE on mainnet. (2) The stress index is a signed z-score, which can't be a price — so we affine-rebase it (100 + 25σ) into a strictly-positive level for the ratio-PnL contract. (3) The Daml SDK ships x86-only, so we ran it under Rosetta 2 on Apple Silicon. (4) Recovery cash-flows lag charge-offs, forcing an explicit net-loss-convention decision (full-netting for the economic carry).