Machine Intelligence and Risk Management
What this covers
Patrick interviews Jeremiah Lowin, who works between machine learning and financial risk management. They start with what a model is. Patrick frames it as a Model as Imperfect Analogy that always carries an error term, using the football field standing in for an acre. Jeremiah's test is whether a model approximates without going too far: the normal distribution suits introductory statistics even though stock returns aren't normal. The technical middle covers closed-form linear regression versus iterative methods, why optimization with known error derivatives beats brute-force or evolutionary search, recurrent networks building context from sequence order, dropout deliberately breaking a model during training, and Google Translate's statistical rather than rule-based approach.
The philosophical thread is where it turns contentious. Jeremiah describes watching Geoffrey Hinton's Boltzmann machines morph one handwritten digit into another, first reading it as intelligence and then as a deterministic product of construction. He invokes Searle's Chinese Room and argues nobody knows what intelligence is. He also forecasts which work is automatable: repetitive tasks that resist explicit rules, and advisory jobs that map experience to rule-based answers.
On the finance side, he calls overfitting the greatest danger, noting that a seven-layer network takes two lines of code and that trusting a strong backtest is itself overfitting. Risk, he says, is shifted or mutated, never eliminated. The Lehman weekend, when the firm ran two sets of books, illustrates this. The episode closes on the claim that "passive" investing is an active choice, with robo-advisors as a middle ground.
Lowin argues that machine learning and human intelligence are both pattern-fitting processes whose 'intelligence' label keeps moving, and that risk management is a skeptical, qualitative practice of understanding a portfolio's full outcome distribution rather than eliminating risk.
- Models are imperfect tools; the skill is knowing when a model is good enough and avoiding overfitting, including false confidence in backtests
- Definitions of intelligence and AI shift as capabilities become routine, so 'intelligence' is hard to define by what a system does
- Risk cannot be wished away, only shifted or mutated, so managers must track where exposure goes and use conversations and fire drills alongside quantitative tools
Risk management means measuring and neutralizing unintended exposures to test whether an edge survives.
- Portfolio exposures unintended by the process can creep in over time, and risk management involves measuring them along multiple dimensions, then deliberately removing or at least monitoring them, since the most important exposures may be unknown unknowns that no dimension captures.
“Your dimensions don't contain probably the most important one.”
- Investing processes seek an edge, and naive portfolio construction drags in unintended risks such as sector or currency exposure; effective risk management identifies those byproducts and checks whether the edge survives neutralizing them.
“can your edge survive neutralizing those risks?”
- If every manager uses the same machine learning tools on the same data, the managers using them may not outperform, and a manager using nothing might win on relative performance.
“Unless, again, thinking about relative skill, like everyone's using the same thing and then the dummy's that's not using anything might win.”
- Absolute skill across managers is at an all-time high, but alpha depends on relative skill, which is at an all-time low, measured by metrics such as Sharpe ratios, creating a paradox for active managers.
“absolute skill's at an all-time high but for alpha, all that matters is relative skill”
- The biggest market risks are often those nobody is talking about, so the best risk thinkers are those who anticipate unknown unknowns, which is the core skill of separating signal from noise.
“the biggest risks in the market are the ones literally nobody is talking about”
- Patrick frames risk management as reducing or suppressing unwanted exposures so a portfolio looks more like a benchmark, such as the S&P 500 or zero exposure, while asking whether Lowin's hands-on work goes further.
“risk manager frankly as somebody that is trying to suppress exposures or to make exposures look more like the thing you were trying to beat”
Every model is an imperfect metaphor that always carries an error term relative to reality.
- Models are metaphors that map an abstract idea onto a concrete one, and like the football-field-as-acre analogy, they always carry an error term; a football field with end zones removed is about 4,500 square feet larger than an acre.
“a model is effectively an analogy or a metaphor just like language is”
- A model is effectively an analogy or metaphor, like language, that captures a large part of an underlying truth but never all of it, so every model carries an error term relative to reality.
“a model is effectively an analogy or a metaphor just like language is”
- Patrick says he has seen plenty of bad backtests because he runs them himself, which contrasts with Lowin's joking claim that he has never seen one.
“I've seen plenty of bad backtests.”
Human questions and relationships stay durable as machine capability advances faster than expected.
- As tools make it faster to reach answers, the durable human contribution is deciding which questions to ask, since the questions remain the domain of aware, conscious humans.
“you still have to come up with the questions. That still seems firmly to be the domain of humans”
- Machines have become better than radiologists at their job faster than expected, and progress is exponential rather than linear, which makes it hard to anticipate which skills will stay valuable.
“machines are actually better at doing the radiologist's job than the radiologist. That happened fast, really fast”
- Those who can design machine learning systems are paid very well, possibly high six figures coming out of college, because few people have the skills even though many jobs need them.
“you might get paid high six figures coming out of college if you're really good at this sort of thing”
- The Turing test asks whether a machine can convince a human in conversation that it is human, which is a goalpost-moving standard rather than a test of intelligence itself.
“the machine needs to be able to deceive within five minutes or a series of text”
- Patrick argues that cultivating relationships is a durable human skill that machines will not replace, even if people come to prefer AI companions.
“you think that there's a basic need for relationship and you can always do well if you were good at cultivating relationships”