A model recommends denying a loan application. A portfolio engine rebalances a client's retirement account overnight. A capital allocation system shifts reserves between business lines. In each case, someone eventually asks the same question: Why did the system do that?
In financial services, that question isn't optional. It's the difference between a decision you can defend to a regulator, a client, or an auditor, and one that you can't.
What explainability means
Explainability is the ability to state, in terms someone can verify, why a system produced the output that it did. That's a higher bar than "interpretability," which is often just about understanding how a model works in the abstract.
Rather than asking how the model behaves in general, explainability asks why this model made this exact call, for this exact customer, on this exact day. For a decision to be truly explainable, you need three things: the factors that were weighed, the relative importance of each, and a clear line from those factors to the final outcome.
Miss any one of those, and you're left with a decision that might be correct, but which no one can actually justify.
Why explainability is crucial in financial services
Regulation is one reason: frameworks like the Federal Reserve's SR 11-7 on model risk management, the Equal Credit Opportunity Act's adverse action requirements, and Basel capital rules all require institutions to show their work. And no regulator is satisfied by “the model said so.”
There's money and trust on the line at the same time: a wealth manager rebalancing client accounts needs to explain a trade in plain language (especially one that triggers a tax event), and a bank denying credit needs specific, defensible reasons for doing so.
Get the explanation wrong and you don't just risk compliance exposure—you risk the client relationship itself. And because these decisions are high-stakes and often irreversible, a mispriced risk model or an unexplainable trade can compound into material financial loss before anyone notices something wrong. There's little room for “we're not sure why it did that.”
This is exactly where black-box machine learning (ML) models run into trouble. A neural network might predict default risk more accurately than a simpler model, but if it can't tell you which factors drove a specific rejection, it's not usable for a regulated decision. Accuracy without an audit trail still leaves you exposed.
Where optimization has the advantage
Mathematical optimization was built to solve exactly this problem, and the explainability comes from its structure. A mixed-integer or quadratic programming model states its objective and constraints explicitly: maximize this, subject to these limits on risk, exposure, tracking error, or regulatory requirements. Every input to the decision is a named, auditable term in the model.
When a portfolio changes, you can trace the change to a specific constraint or coefficient. When a problem is infeasible, tools like Gurobi's Irreducible Inconsistent Subsystem (IIS) identify a minimal set of conflicting constraints causing the problem, rather than leaving you to guess. And with hierarchical, multi-objective optimization, you can show not just what decision was made, but the priority order that produced it: minimize tracking error first, then transaction costs, then portfolio complexity or investor preferences.
That last point is critical for organizations like MyGoals, a Canadian business that builds a highly detailed, individual-level financial model for each of ther clients. The team uses Gurobi-powered mathematical optimziation to determine the optimal sequence of contributions and withdrawals over time, based on the client’s specific goals. As Raphi Zaionz, founder of MyGoals, put it: “It’s not just about getting the best answer; it’s about giving advisors and clients a plan they can understand and trust.”
That's what turns “the algorithm decided” into “here is the exact order of priorities that produced this trade,” a sentence that satisfies a client, a compliance officer, and an internal audit team all at once.
The bigger picture
None of this positions optimization against machine learning. In fact, the two work together, along with the LLMs that are increasingly layered on top of both. ML models are often excellent at prediction; optimization is what turns predictions into decisions you can defend. In an industry where every decision might eventually be reviewed by a regulator, a client, or a courtroom, that combination is what makes a decision defensible in the first place.
If your organization is choosing between decision systems, accuracy alone doesn't answer the harder question: Can you explain that answer to the person standing in front of you?
Learn more about how Gurobi optimizes the financial services industry and explore our dedicated documentation for finance.
