WBS Valletta: Automated Pricing Model Validation via AADC Tape Introspection
Dmitri Goloubentsev presented at the 22nd WBS Quantitative Finance Conference. When a pricing model runs on AADC, the whole calculation is recorded with every value tied to the source file and line that produced it. The talk showed what a validator gets from one such recording of an unmodified QuantLib swaption: 140 hardcoded constants with their coordinates, 85 frozen branches, which market quotes a discount factor depends on and which it provably does not, and where a replay stops being valid. The evidence is laid out against PRA SS1/23, SR 26-2 and the ECB guide, and every number in the slides reproduces from a pip install. Earlier that day Dmitri joined Alexander Sokol, Nicole Königstein and Achintya Gopal on the opening panel, Is Active Deception Part of AI Model Risk?
QuantLib 1.41 built with AADC source tracing, so a Python script can ask an unmodified C++ pricer which constants, branches and inputs produced a number. Linux x86-64.
The 2.x line is a new engine: record once, replay the plain tape for a one-off valuation, or compile it for repeated use and batches. Wheels for Linux x86-64 and aarch64, Windows x64 and, since 18 September, Apple Silicon, so a model developed on a Mac runs unchanged on the grid. The Community Edition is free for non-commercial and academic use and runs the interpreter and every JIT backend at full speed; an Enterprise licence adds the C++ SDK, AVX-512, the ahead-of-time code generators and production use. Intel Macs are not supported.
Innovation of the year (tech), Energy Risk Awards 2026
Energy Risk named MatLogica innovation of the year in the technology category at the 2026 awards, announced at the Energy Risk USA dinner in Houston on 7 May, for the gas storage work in which a neural policy is trained through the recorded simulation.
SNAPO on arXiv: optimal control via differentiable simulation
Dmitri Goloubentsev and Natalija Karpichina publish SNAPO (Smooth Neural Adjoint Policy Optimization): a neural policy inside a differentiable simulator with smooth constraints and exact adjoint gradients, demonstrated on natural gas storage, pension fund asset-liability management and pharmaceutical manufacturing.
WBS Palermo: accurate Greeks for autocallables with AAD
At the 21st WBS Quantitative Finance Conference, Dmitri Goloubentsev presented a production-ready way to compute autocallable Greeks with smoothing and AAD, cutting the computational cost by 90% without losing accuracy.
Ranked 13th in the Chartis Quantitative Analytics 50
Chartis Research again named MatLogica Category Leader in automatic differentiation, with five category awards in total and an overall ranking of 13th in the Quantitative Analytics 50.
On the final evening of QuantMinds International, Dmitri Goloubentsev and George Petropoulos of Delta Capita hosted a roundtable joined by Jesper Andreasen, Peter Jäckel and Serguei Issakov.
News
Technology newcomer of the year, Asia Risk Technology Awards 2024
Risk.net named MatLogica technology newcomer of the year in its Asia Risk Technology Awards.
Intel's developer site published a walkthrough of accelerating simulations and backpropagation with AADC from Python and C++ analytics, using vectorisation and adjoint differentiation.
Chartis QuantTech 2023: AAD and cloud for real-time risk
MatLogica was ranked 10th in the Chartis QuantTech50. Dmitri Goloubentsev's presentation, Leveraging Automatic Adjoint Differentiation and Cloud for Real-time Risk, is on YouTube.
Two conference talks used MatLogica's code generation AAD: Stephan Bosch of ING on comparing AAD techniques and performance, and Svetlana Borovkova of Probability & Partners on estimating expected shortfall sensitivities with AADC.
Chartis Research gave MatLogica four awards: innovation, AAD, data-parallel programming, and innovation in computational frameworks, with an overall ranking of 10th out of 50.
QuantMinds 2022: Automatic IFT, from overnight risk to live risk
Dmitri Goloubentsev's Barcelona talk on the automatic implicit function theorem, based on the paper by Goloubentsev, Lakshtanov and Piterbarg published on Risk.net and SSRN.
AADC used to train neural networks for time series
Prof. Roland Olsson used AADC to train custom neural network architectures for time series analysis. The paper reports better accuracy than the methods it compares against and several times lower training time.
Adopting AADC gave the bank 15 to 20 times faster risk, cut overnight portfolio risk from more than 8 hours to 2, and intraday risk from more than 30 minutes to a few.
An online environment with the multi-curve fitting and AAD risk demo on QuantLib and an external Levenberg-Marquardt library, and a video on reaching a 350× performance gain for XVA pricing on Intel AVX2 with 5 threads.
Dr Antoine Savine, author of Modern Computational Finance (Wiley, 2018) and best known for his work on volatility and interest rate models, became an advisor to MatLogica.
2021
Event
QuantMinds 2021: AAD integration strategies
Dmitri Goloubentsev on AAD integration strategies for top performance and ease of use, presented at QuantMinds in Barcelona.
By continuing to use this site you accept our use of cookies and acknowledge that all content, technology descriptions, and methodologies are protected by copyright and applicable patents. See our Privacy Notice.
