Model Blind-Spot Override
Override systematic risk models only when they cannot represent a known event
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 5
- Confidence
- 96%
The Model Blind-Spot Override separates routine uncertainty from events outside a model's representational scope. The default is to trust a disciplined systematic process that predicts portfolio volatility and correlation risk. Humans intervene only after identifying information or a mechanism the model cannot know, such as a discontinuous policy event or a visible deleveraging cascade. The decision is not whether an event feels alarming, but whether the model has incorporated its market effects sufficiently quickly and whether its risk estimate remains credible. When the model is genuinely blind, informed judgment can reduce exposure. When volatility and correlations have already adapted, intervention may do more harm than good. This creates a bounded role for discretion while protecting the systematic process from emotional overreaction.
Origin
Bouchaud explains that CFM generally relies on systematic risk models but has occasionally intervened when humans knew something the models could not, including Brexit-related risks, Liberation Day tariffs, and the 2007 quant deleveraging.
Core principles
- 01Use systematic controls under normal conditions
- 02Accept that models cannot represent every event
- 03Intervene only for a specific, explainable blind spot
- 04Judge whether the model has adapted to new information
- 05Avoid treating every unusual event as unprecedented
How to run it
- 1
Detect an exception
Look for an event or portfolio pattern inconsistent with the assumptions behind the normal risk model. Describe the anomaly concretely rather than labeling it merely unprecedented.
Pro tip Use portfolio behavior, market microstructure, and known event mechanics together.
Watch out Losses alone do not prove that the model is blind.
- 2
Inspect model coverage
Determine whether volatility, correlations, liquidity, and other relevant effects have already entered the model's estimates. Separate failure to predict the event from failure to measure its consequences.
Pro tip Ask whether the model has adapted sufficiently fast, not whether it forecast the news.
Watch out No risk model is expected to predict every geopolitical event.
- 3
Specify the blind spot
State exactly what humans know that the model cannot represent. Connect that missing information to a plausible portfolio-loss mechanism.
Pro tip Require an explicit causal pathway from missing information to risk.
Watch out Vague fear is not a valid override signal.
- 4
Make a bounded judgment call
Adjust exposure only enough to address the identified blind spot. Document why the normal model was inadequate and what evidence would reverse the decision.
Pro tip Prefer temporary, reversible risk reductions.
Watch out Do not convert an exceptional override into open-ended discretionary trading.
- 5
Guard against overreaction
Compare the event with historical episodes and recognize that unusual shocks occur regularly. Maintain the override only while the model remains materially unreliable.
Pro tip Review whether the supposedly exceptional environment has become observable market data.
Watch out Repeated intervention can destroy the consistency that gives systematic investing its value.
In the wild
CFM judged that the Liberation Day tariff announcement created market behavior to which its models were initially blind. The team manually intervened because the model's normal risk estimate did not adequately represent the shock.
→ Human judgment supplemented the systematic process during a specific model failure.
Unusual portfolio behavior appeared weeks before the worst day of the quant quake. CFM inferred that another fund's systematic deleveraging could produce the observed pattern and cascade through correlated portfolios, so it reduced leverage before the acute event.
→ A structural interpretation of anomalous signals prompted an early risk reduction.
Common mistakes
Overriding after every loss
Ordinary model losses are not evidence of a blind spot. Intervention should require a mechanism the model demonstrably cannot capture.
Demanding event prediction
A model can manage an event's risk without forecasting the event itself. Evaluate adaptation in volatility, correlation, and exposure estimates.
Calling every shock unique
Unexpected events occur repeatedly in market history. Treating each one as wholly unprecedented encourages excessive and inconsistent intervention.
Is it for you?
Best for
It is best for mature systematic investment teams with established models, governance, and experienced human oversight.
Not ideal for
It is not ideal as a justification for discretionary interference whenever a model experiences ordinary losses.
From the transcript
“we decided that our risk model couldn't know about things that we humans knew”
“I think the idea really is that humans should use their best judgment in these cases”
“I think the the tricky part is not to overreact”
From the episode
The Intersection of Science and Finance with CFM's Jean-Philippe Bouchaud