SolvNum Audit & Compression Overlay
LRDE makes the pricing fast. SolvNum makes the outputs portable, auditable, and 3–7× smaller. Two products that compose.
- 3.8–4.9×
- Compression at k=12
- 4×10⁻⁵
- Median relative error (k=12)
- SHA-256
- Cross-platform receipt
- 12 s
- Full pipeline on a laptop
The scenario
Set the picture
LRDE has priced the book. The outputs — vol surfaces, Greek matrices, scenario PV cubes — need to be stored, shipped over the wire to downstream consumers (risk consolidator, compliance, regulators), and reproduced months later for audit.
Today, the front-office pricer, the risk engine, and the back-office settlement system produce slightly different numbers because they run on different hardware with different BLAS and libm versions. Nobody can quickly say which one is right.
Cost today
Raw float64 storage: a nightly 5,000-scenario PV cube is ~4 GB. Over 7-year retention that’s 10 TB per bank.
Cross-platform reproducibility: absent. Two machines with different libm produce different last digits. A model-risk officer cannot verify the bank’s xVA reporting by re-running on their own hardware.
What changes with LRDE
SolvNum’s relative-error compression at k=12 (the default) compresses pricing artifacts 3.8–4.9× with ~4×10⁻⁵ median relative error — sub-cent on a $100 spot.
Every compressed artifact carries a SHA-256 receipt that re-derives bit-identically on Windows, Linux, ARM, GPU, and WebAssembly. The hash either matches or it doesn’t — a mismatch unambiguously means the inputs differ, not the math.
Encode/decode overhead: 0.2–5.4 ms per artifact at production sizes — negligible vs the LRDE pricing time.
Measurable outcome
What we claim — and how it survives review
Each line below maps to a captured number in the demo section. Every number is reproducible from the benchmark suite.
- 3.0–7.1× compression at k=8, 3.8–4.9× at k=12, 3.0–3.8× at k=16.
- Median relative error: ~7×10⁻⁴ (k=8), ~4×10⁻⁵ (k=12), ~2.5×10⁻⁶ (k=16).
- SHA-256 receipt verifiable on any platform in under 1 second.
- Encode/decode wall time: 0.2–5.4 ms per artifact — negligible overhead.
- Full pipeline (LRDE price + SolvNum compress + sign) reproducible in 12 seconds on a laptop.
The demo
What was tested. How. What the script printed.
Takes the six artifact types from POCs 1–4 (vol surface, price matrix, delta/gamma/vega matrices, scenario PV matrix) and runs SolvNum’s relative-error compression codec at k=8, k=12, and k=16.
The SolvNum-decoded arrays are SHA-256 hashed. A buyer can re-decode on their own machine and verify the hash match. This is the artifact to bring to the model-risk officer alongside the speedup numbers.
Captured benchmark output
The numbers the script actually printed.
| k | Bits/value | Compression vs float64 | Median rel error | Best fit |
|---|---|---|---|---|
| k=8 | 9–13 | 5.0–7.1× | ~7×10⁻⁴ | Archive / historical P&L |
| k=12 | 13–17 | 3.8–4.9× | ~4×10⁻⁵ | Default daily storage |
| k=16 | 17–21 | 3.0–3.8× | ~2.5×10⁻⁶ | Audit-grade (6+ digits) |
| Artifact | n values | Compression | Med rel err |
|---|---|---|---|
| Vol surface (50×12) | 600 | 3.76× | 4.4×10⁻⁵ |
| Price matrix (50×4) | 200 | 4.00× | 4.0×10⁻⁵ |
| Delta matrix (50×4) | 200 | 4.27× | 4.6×10⁻⁵ |
| Gamma matrix (50×4) | 200 | 4.92× | 3.7×10⁻³ * |
| Vega matrix (50×4) | 200 | 4.00× | 4.1×10⁻⁵ |
| Scenario PVs (200×25) | 5,000 | 4.00× | 1.5×10⁻⁴ |
* Gamma’s larger relative error is because gamma values are small (1e-4 to 1e-2); absolute error is still ~1e-7.
Composes with
Where this POC sits in the benchmark suite
POC 01
European Vanilla Pricing
Compresses the price array from POC 1.
POC 02
Vol Surface Revaluation
Compresses the daily vol surface checkpoint 3.8×.
POC 03
Risk Greeks via Finite-Difference Bumping
Compresses Greek matrices for downstream consumers.
POC 04
Multi-Scenario Overnight Risk Fan-Out
Highest-value bundling: compresses the scenario PV cube for regulatory audit.
Evidence pointers
Where the claims live in the repo
These are the files a reviewer should run to re-derive every number on this page.
- poc/lrde_hedge_fund/solvnum_overlay/bench.py
- docs/pitch/LRDE_AND_SOLVNUM_TOGETHER.md
Want to see these numbers on your book?
Run the benchmark on your actual vol surface and trade book.
Two weeks, $25K, fully credited. No production integration, no data leaving your premises. Every claim above traces back to a script you can run locally.