rates-engine-cpp

A C++20 rates engine and curve-risk research lab. It combines OIS discounting, separate projection curves and full-rebootstrap quote DV01 with an experiment that asks: what valuation uncertainty remains when every market quote fits? Independent Python and QuantLib implementations check pricing and risk.

Curve ambiguity audit

The research workflow finds minimum/maximum portfolio PVs compatible with quote intervals and explicit forward-rate bounds. It exports the curves that attain them, checks LP optimality conditions, reprices with QuantLib, and ranks hypothetical additional quotes by how much they narrow the range.

In the synthetic 1 million, 3.5Y payer example, quote DV01 can be hedged below 1e-6 currency/bp while the same-quote valuation range remains unchanged:

Assumed continuous forward bounds	Minimum PV	Maximum PV
−1% to 10%	−66,244.31	63,956.10
2.5% to 4.5%	−5,776.30	11,559.45

The ordinary bootstrap PV is 2,720.35. These are attainable ranges under stated assumptions, not confidence intervals or market forecasts. The broad range demonstrates why the forward assumptions matter.

Research, prior art and reproduction · Inputs · Witnesses and numerical evidence

Admissible curve bounds have established prior art. This project's contribution is an auditable experiment joining valuation ambiguity, a quote-neutral hedge counterexample and conditional quote prioritization; no new mathematical theory or global novelty is claimed. The native API also supports discount uncertainty with a separate projection curve held fixed. Joint curve uncertainty and seasoned SOFR ambiguity are outside this experiment's scope.

Captured SOFR example

The captured SOFR example values a USD 1 million payer, 3.75% fixed, from 1 October 2026 to 1 October 2028, as of 21 October 2026, 18:17:09 UTC:

Seasoned SOFR valuation	Result
Dirty PV	-4,448.43 USD
Signed accrued	144.94 USD
Clean PV	-4,593.36 USD
Daily observations	15 realized / 484 projected

Captured NY Fed fixings are combined with synthetic OIS quotes. The SOFR validation record reports 30,524 comparisons, with a 2e-5 currency/bp tolerance for quote risk; the captured-fixing replay uses 2e-5 currency for PV/cash flows. These are scoped acceptance tolerances, not a general accuracy guarantee. Inputs and reproduction retain the trade's observation conventions and valuation date.

• Pricing: fixed/floating swaps, seasoned SOFR swaps with realized fixings, observation conventions, and compounded overnight basis swaps.
• Risk and hedging: discount/projection quote sensitivities, shared portfolio scenarios, bounded hedge sizing and full-repricing stress analysis.
• Lifecycle: contractual payment reconciliation, cash-collateral ledgers and optional Python/SQLite journals with transaction and replay checks.

Calibrated zero curve	DV01 ladder

These plots show the introductory single-curve, synthetic example. Quote DV01 bumps market inputs and rebuilds the curve; zero-node DV01 shifts node yields. Inputs, outputs and reproduction and chart provenance are retained. The instantaneous forward plot shows the jumps at pillars implied by log-linear discount interpolation.

Current release results

Recorded for v2.3.0, implementation commit 00519f48d767, on 22 October 2026. Each row names its configuration or suite; the counts are not interchangeable.

Check	Recorded result	Evidence
Windows Release CTest, Python bindings enabled	100/100 passed	Runtime record
Independent OIS / separate-projection validation	230/230 comparisons passed, QuantLib 1.44	Methods and tolerances, report
Release-commit CI	7/7 jobs passed: Linux, Windows, macOS, sanitizers, static analysis and coverage	CI record
C++ library coverage, Linux Clang/LLVM 18 Debug	92.50% lines / 80.13% branch outcomes; basis.cpp: 43.46% lines	Scope and diagnostics

Coverage excludes CLI, bindings, Python code and dependencies. LLVM reported 31 functions with mismatched data; the linked report retains that warning. Basis calibration and risk need stronger instrumented C++ test coverage; the aggregate percentage should not obscure that gap. Historical measurements use different code and configurations. The badge above shows the current workflow status.

OIS discounting and separate projection

The C++ and Python APIs bootstrap an OIS discount curve, then calibrate a projection curve against forward-market quotes using that discount curve. Discount-quote shocks rebuild discounting and recalibrate projection to the unchanged forwarding quotes; projection-quote shocks hold discounting fixed. QuantLib independently bootstraps from the same quotes and repeats the shocks.

Use OisCurveBootstrapper, CurveBootstrapper::build_forward, price(discount, projection, swap) and multi_curve_risk for this workflow. The multi-curve guide includes a C++ example and validation details. Seasoned SOFR and basis swaps have dedicated APIs for their observation conventions and curve dependencies. The CSV CLI remains a single-curve introductory tool.

The examples use synthetic OIS/forward quotes. The SOFR fixing demonstration combines captured NY Fed fixings with synthetic OIS quotes. Treasury curves are used only in the separate historical proxy study; they are not an executable swap-market calibration. See the market-data policy and model boundaries.

Build and validate

Requires CMake 3.25+, Ninja, a C++20 compiler and Python 3.10+. Install the build prerequisites first. GoogleTest is fetched on the first test build. For the configured Windows environment:

.\run.ps1 -WithPython -Test
.\.venv\Scripts\python.exe python/validate_multicurve.py --module-dir build/release --out results/multicurve --require-quantlib

On Linux/macOS, with the chosen Python environment active:

python -m pip install pybind11 QuantLib "scipy>=1.14,<2"
cmake --preset release -DRATES_BUILD_PYTHON=ON -Dpybind11_DIR="$(python -m pybind11 --cmakedir)" -DPython3_EXECUTABLE="$(python -c 'import sys; print(sys.executable)')"
cmake --build --preset release --parallel 4
ctest --preset release
python python/validate_multicurve.py --module-dir build/release --out results/multicurve --require-quantlib

For a prebuilt CLI demonstration, extract the v2.3.0 Windows x64 ZIP, open PowerShell in its root and run:

.\bin\rates_cli.exe --quotes examples/synthetic_quotes.csv --swap examples/sample_swap.csv --out results/demo

Default Windows executables embed compiler runtimes. The ZIP includes native tools and a static library; Python workflows require bindings built for the chosen interpreter. Public-download and extracted-ZIP verification records the checksum and system-only PATH checks.

Follow pricing through to a hedge

The SOFR hedging study sizes a bounded, lot-rounded hedge and measures residual quote DV01 and full-repricing shocks. Its risk and scenario chart shows both improvements and shocks that worsen. It uses synthetic prices and assumed entry costs; reducing the chosen sensitivity norm does not establish trading returns or general market-risk reduction.

python python/hedge_study.py --module-dir build/release --out results/hedging --require-quantlib

Portfolio benchmarks demonstrate the design choice of sharing calibrated scenarios across trades versus rebuilding them independently. They do not benchmark against QuantLib. Workload definitions, absolute timings and machine details are in the scenario-reuse study. The solver study measures calibration and risk on the same workload before and after the safeguarded-secant implementation.

Documentation

Topic	Guide
Valuations, cash flows and stress examples	Worked examples
Risk signs, coordinates and conventions	Risk methodology, conventions
Contractual payments and durable journals	Lifecycle guide
Design, C++ consumers and implementation	Architecture, consumer example, walkthrough
Performance experiments and historical results	Benchmark methodology, portfolio study
Release history and scope	Release notes, development history, limitations

MIT licensed; dependencies retain their own licenses. This is a deterministic swaps project with explicit model boundaries and reproducible validation.