Technology
The computational foundation behind explainable risk analytics
MASTIX uses Adjoint Algorithmic Differentiation (AAD) to compute exact sensitivities as part of the valuation itself. That is what makes fast attribution, scenario analysis, and traceable risk analytics possible.
Tracking every driver of every number means calculating thousands of sensitivities. For a portfolio of 100,000 instruments with 1,000 risk factors, that means 100 million individual calculations just to get a full sensitivity set.
Traditional finite-difference methods require a separate run for each sensitivity: bump one input, recalculate everything, measure the change. At portfolio scale, that rules out anything faster than an overnight batch.
Traditional Approach
Adjoint Algorithmic Differentiation (AAD) computes sensitivities to all inputs in a single structured pass through the calculation. Instead of asking 'what happens if I change this input?' over and over, it answers all of those questions simultaneously.
The cost is roughly a small multiple of a single valuation — regardless of how many sensitivities you need. Whether you need 10 or 10,000, the computational cost is essentially the same.
AAD approach
Why the cost profile changes
Because sensitivities are propagated through the same computation graph as the valuation, adding more risk factors does not create a new portfolio run for each one. Runtime follows the structure of the model rather than the length of the bump list, so broad sensitivity coverage becomes usable during analysis.
AAD is not a reporting add-on. The calculation engine itself has to track dependencies through every operation, so sensitivities are produced as part of the valuation rather than estimated afterward.
That architectural choice matters downstream. If valuations, sensitivities, scenarios, and attribution share one foundation, the numbers stay aligned and the explanations stay connected to the calculation that produced them.
Architectural consequence
When sensitivities come from the valuation itself, attribution and scenario analysis stop being separate reconciliation problems.
Because scenarios reuse the same analytical foundation, teams can explore lighter changes during the working session. Broader balance-sheet scenarios can still run as controlled portfolio-scale analyses rather than overnight requests.
Because sensitivities come from the valuation itself, attribution stays connected to the calculation that produced the result. Movements can be decomposed into rates, volumes, model changes, and new deals without rebuilding the explanation afterward.
Because outputs, drivers, and assumptions share one engine, the path from result back to input remains visible. That makes review and model-governance discussions about the calculation itself, not a black-box report.
Because the engine avoids one full rerun per factor, portfolio-scale analysis does not have to start from a distributed-compute assumption. The design supports single-machine setups for workloads that would otherwise be pushed into heavier batch infrastructure.
For readers who want more technical detail, these are the best next steps.
12 min read
How the computational foundation of ALM is changing, and what an architecture built for interactive, auditable analysis looks like in practice.
Read more6 min read
How MASTIX Derivatives Studio uses Adjoint Algorithmic Differentiation to produce fast, sensitivity-based P&L explain across portfolio, time, and market-data effects.
Read moreThe best way to understand the difference is to see it. Get in touch and we can walk through interactive attribution on a portfolio that looks like yours.
