How Did Your Model Actually Compute It? Validation from the Run

How Did Your Model Actually Compute It? Validation from the Run

Speaker: Dmitri Goloubentsev, CTO, MatLogica

Model validation needs to establish what pricing code actually computes, not just what the source code and documentation say it should. Yet validating complex models still relies heavily on manual code inspection. As AI accelerates code generation and modification, understanding what actually executes becomes even more important.

In this webinar, Dmitri Goloubentsev, CTO of MatLogica, introduces AADC Tracer, which records calculations as they execute in the production build and links each operation to the source code that performed it.

Using QuantLib as a worked example, including a barrier option whose price depends on more day-count conventions than the engine code suggests, Dmitri will demonstrate how to:

• See the formula actually computed, with each term linked to its source line.
• Trace the inputs, hard-coded constants and market-dependent branches behind a price.
• Replay calculations under different scenarios with exact Greeks, and see where a replay stops being valid.
• Compare what model documentation says with what the run computed.
• Preserve execution records as evidence for validation under PRA SS1/23, SR 26-2 and the ECB guide.
• Use AI to analyse execution evidence, with every claim linked to a source file and line for checking.

The session shows how validation can move beyond reading code to examining what the model actually did.