Selective Model Evolution Protocol
Separate structural changes from one-off shocks before updating models
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 6
- Confidence
- 98%
The protocol starts by distinguishing a novel event from a durable change in the market-generating process. A singular political event should not automatically be encoded as though it will recur frequently; instead, the existing portfolio is stress-tested under several plausible outcomes and risk is adjusted if necessary. Persistent developments—such as changing correlations, volatility levels, monetary-policy regimes, or newly relevant factors—should inform the model going forward. This separation reduces two opposing errors: leaving a model frozen as its environment changes and overfitting it to an exceptional shock. Each case is evaluated independently, with scenario analysis governing immediate exposure and evidence of persistence governing structural model updates.
Origin
Vassalou described this protocol through her experience managing quantitative strategies across the financial crisis, the British referendum, and changing monetary-policy regimes. Extracted from Masters in Business.
Core principles
- 01Models must evolve as persistent market structure changes
- 02Rare events should be stress-tested rather than automatically learned
- 03Recurring shifts in volatility and correlation belong in adaptive models
- 04Every proposed model change requires case-specific judgment
How to run it
- 1
Detect the change
Identify the event, variable, or relationship that differs from the models established environment.
Pro tip Describe the change in measurable terms whenever possible.
Watch out A dramatic headline does not necessarily represent a structural shift.
- 2
Classify persistence
Decide whether the development is a rare event or evidence of a recurring change in correlations, volatility, policy, or market structure.
Pro tip Ask whether the model is likely to encounter comparable conditions repeatedly.
Watch out Treating a one-off event as recurring can produce severe overfitting.
- 3
Stress-test exceptions
For rare events, create plausible scenarios and estimate how the existing model and portfolio behave in each.
Pro tip Include adverse but credible scenarios, not only the expected outcome.
Watch out Historical simulation alone may omit genuinely unprecedented outcomes.
- 4
Set immediate risk
Use stress-test results to decide whether to reduce, retain, or hedge current exposure.
Pro tip Tie the decision to predefined loss tolerance and liquidity constraints.
Watch out Do not confuse confidence in a forecast with resilience to being wrong.
- 5
Update durable inputs
Incorporate persistent changes into the models factors, parameters, or adaptation rules.
Pro tip Prioritize variables with an observable and economically defensible mechanism.
Watch out Adding every new variable can make the model unstable and opaque.
- 6
Validate continuously
Track whether the revised model behaves as intended and whether the presumed structural change persists.
Pro tip Compare revised and unrevised model behavior during a controlled evaluation period.
Watch out A model update is provisional until it survives new data.
In the wild
A portfolio team faces a referendum with no close historical precedent. Rather than retraining the model on an event unlikely to recur regularly, it models several political and market outcomes, estimates portfolio losses under each, and reduces or retains risk according to those results.
→ The portfolio is prepared for the event without embedding a singular episode into its normal forecasting process.
The team observes a sustained increase in volatility and a persistent change in cross-asset correlations. It updates the relevant estimates and adaptation rules, then monitors whether the new relationships hold in subsequent data.
→ The model responds to a durable market change rather than operating on stale assumptions.
Common mistakes
Build it and forget it
A fixed model eventually reflects an obsolete market environment and can misstate both opportunity and risk.
Learn from every shock
Automatically adapting to exceptional events teaches the model to expect patterns that may never recur.
Add factors indiscriminately
New information should enter only when its relevance and persistence justify the added complexity.
Is it for you?
Best for
It is best for quantitative teams maintaining live models through shocks, policy changes, and shifting market structure.
Not ideal for
It is not ideal for discretionary processes that lack explicit models, scenarios, or measurable risk limits.
From the transcript
“you can just build it and forget it, but it has to be done in a way that keeps up with the developments in the…”
“that's an instance where you want to take your model and stress tested to see how it will behave depending on different scenarios that may…”
“you have to be selective and really treat each case separative.”
From the episode
Maria Vassalou on the Small-Cap Effect
Maria Vassalou