1. What is model risk on a trading desk?

Working definition:

Model risk = risk of wrong decisions or wrong numbers because the model, inputs, or usage are wrong.

Sources:

  • Wrong structure
    • Using Black–Scholes for instruments with early exercise, strong path-dependence, or jumps/spikes.
  • Wrong parameters
    • Mis-calibrated vol surfaces, stale correlations, wrong mean-reversion speeds.
  • Wrong implementation
    • Coding bugs, sign errors, interpolation mistakes.
  • Wrong use
    • Using a short-rate model beyond its calibrated tenor, ignoring use-limits.

For a senior quant dev, the key is: you don’t “trust the model”; you control it.

2. Core controls around pricing models

2.1 Documentation and assumptions

Each model should have:

  • Clear product scope: what it does / does not cover.
  • Assumptions:
    • Dynamics (GBM vs OU vs jump),
    • Market regime (liquid, arbitrage-free, etc.),
    • Exercise style (European only, no path-dependence).
  • Known limitations:
    • “Breaks down for deep OTM options with very long maturity.”
    • “Not calibrated to stress regimes.”

This is what you say in an interview:

“We explicitly document assumptions and use-limits; that’s part of model control.”

2.2 Independent validation

Model validation checks:

  • Theoretical soundness:
    • Is the model coherent with no-arbitrage and desk use-case?
  • Implementation:
    • Code review, unit tests, finite-difference checks.
  • Calibration quality:
    • Fit errors, stability, parameter plausibility.
  • Backtesting / benchmarking:
    • Compare model prices vs market, model Greeks vs realised P&L.

Validation outcome:

  • Approved with limits, or
  • Rejected / needs remediation.

Limits could be:

  • Product types, tenors, underlying sets,
  • Vol/parameter ranges,
  • Purpose (e.g. pricing allowed, capital disallowed).

3. Drift monitoring and regression control

Once in production, the real game is drift:

3.1 Golden regression tests

  • Fix:
    • A set of reference trades,
    • A reference market snapshot,
    • A specific model configuration.
  • Store:
    • PV, Greeks, selected diagnostics.

Every code or config change:

  • Re-run tests; compare new vs golden values.
  • Enforce tolerances:
    • |new - old| ≤ abs_tol + rel_tol * |old|.

If breach:

  • Either:
    • There is a bug, or
    • A deliberate model change that needs:
      • Documentation,
      • New golden values.

This is your regression shield.

3.2 Calibration and performance drift

Monitor over time:

  • Calibration error distributions:
    • Are we drifting from market quotes?
  • Parameter paths:
    • Vol-of-vol, mean-reversion, correlations.
    • Are they stable? Are there jumps that indicate regime change or data issues?
  • P&L residuals:
    • From P&L explain:
      • Risk + new trades + model changes vs actual P&L.
    • Persistent residual patterns → model or mapping problems.

You want:

“If the world or our behaviour changes, I see the drift in a dashboard before it hits P&L badly.”

4. Model risk vs trading P&L

Important conceptual point:

  • True trading P&L should come from:
    • Views, positioning, risk-taking.
  • Systematic P&L from model bias is not alpha; it’s model risk P&L.

Examples:

  • Using a model that systematically undervalues a structured product you’re long in:
    • Booked P&L looks great; market P&L may not.
  • Wrongly specified smile:
    • Hedging using wrong vega/gamma; residuals show up as unexplained P&L.

Risk function wants to separate:

  • Trading performance vs
  • Model performance.

5. How to phrase it in an interview

Short, desk-flavoured summary:

“Model risk comes from structure, parameters, implementation, and misuse. We control it with a validation process that documents assumptions, tests implementation with unit and finite-difference checks, and benchmarks against market data. In production, we use golden regression suites and numerical tolerances to detect unintended behaviour changes, and we monitor calibration errors and P&L residuals over time. Persistent residuals are treated as model risk P&L, not trading alpha, and trigger model review.”

That’s usually enough to sound like an adult in the room.