Jeremiah Lowin – Machine Intelligence and Risk Management - [Invest Like the Best, EP.20]

Below is the interpretation of the Jeremiah Lowin interview transcript, based on the rules and analytical framework you provided.

At a Glance

Guest Jeremiah Lowin is the Director of Risk Management at a private investment firm in the New York area and an expert in machine learning and statistics. This episode revolves around two main themes: the principles and limitations of machine learning, and the philosophy and practice of portfolio risk management. The most impactful judgment in the entire episode is: Jeremiah Lowin believes that current so-called "artificial intelligence" does not possess true intelligence; it is merely a more powerful pattern recognition tool. Furthermore, the core of risk management is not to eliminate risk, but to understand and reallocate it, and the most effective risk management tool is often qualitative conversation, not complex quantitative models.

Models Are Tools, Not Magic: From Linear Regression to Machine Learning

Jeremiah Lowin argues that all models, including the most advanced machine learning models, are essentially "tools"—imperfect maps of reality. Their value lies in helping humans make better decisions, not in replacing them.

• The Essence of Models is Approximation: Lowin uses the analogy of a "football field" versus an "acre" to illustrate that models, like language, are a "necessary approximation." A football field is about 4,500 square feet larger than an acre. When scaled up to 10 acres, the error accumulates to one acre. Therefore, understanding a model's margin of error is crucial.
• From Linear Regression to Iterative Optimization: Linear regression is a "closed-form solution" model—input data yields a direct result. Machine learning models (like neural networks) are "iterative models." They approach an optimal solution by continuously calculating errors and adjusting parameters (e.g., gradient descent). This differs from "brute force" or "evolutionary algorithms," which are more random and computationally expensive.
• Time Series and Context: Lowin emphasizes that for applications in fields like finance, the "order" of data is critical. Citing the progress of Google Translate, he notes that its core breakthrough came from using "recurrent neural networks." These networks, like humans, build context progressively while reading a sentence, rather than processing words in isolation. This captures richer semantic information than traditional methods.

The Art of Risk Management: From Measurement to Conversation

Jeremiah Lowin believes the core of risk management is not eliminating risk (which is usually impossible), but understanding risk and actively reallocating it. He emphasizes qualitative conversation and visualization tools over a pure reliance on quantitative models.

• Defining and Measuring Risk: Lowin defines a risk manager as a "professional skeptic." Using the metaphor of "observing a sculpture with a flashlight in a dark room," he explains that risk measurement is like seeing the sculpture's projection—you can never see the whole picture. Risk is not just "permanent capital loss" or "volatility," but an understanding of how an asset or portfolio behaves under all conditions (up, down, sideways).
• The Core of Risk Management is Conversation: Lowin argues that the most effective risk management tool is "a lot of conversation." By conducting "fire drill"-style stress tests and scenario analyses with investment managers, unknown risks ("unknown unknowns") can be exposed in advance. This allows decision-makers to react based on prior thinking during a crisis, rather than relying on models that may fail.
• Risk Cannot Be Eliminated, Only Transferred: Lowin points out that risk is like a mass; it cannot be eliminated from thin air, only reallocated or transferred (e.g., via derivatives). The art of risk management lies in understanding this "mutation"—when you remove one risk, it appears elsewhere in another form. Therefore, one must continuously track the portfolio's evolution, not just its current state.
• Visualization and Intuition: Lowin shares his experience building a "3D risk exposure map." This map transforms complex, multi-dimensional risks (e.g., exposure to oil, credit) into an intuitive graphic. Its key value is not in showing "where you are now," but in showing "how you evolved to this point," helping decision-makers spot slowly accumulating risks (the "boiling frog" effect).

Reflections on Active vs. Passive Management

Jeremiah Lowin argues that there is no true "passive investing" because choosing an index is itself an active decision. He is increasingly inclined towards a "low-activity" investment approach, as the cost of finding good active managers is rising.

• "Passive" is Also Active: Lowin notes that choosing to invest in the stock market (as opposed to bonds or cash) is itself an active, opinionated decision. Therefore, the line between "active" and "passive" is blurred.
• Increased Difficulty in Finding Good Managers: Lowin observes that due to lower barriers to entry in the hedge fund industry, increased competition, and institutional investors (like pensions) becoming the primary source of capital and altering incentive structures, the difficulty and cost of finding truly alpha-generating active managers are rising. He tends to remain skeptical unless fully convinced.
• The Danger of Machine Learning: Overfitting: Lowin warns that machine learning models (especially deep neural networks) have vast numbers of parameters and are highly prone to "overfitting" historical data, producing perfect backtest results. He cites the "dropout" technique as an example of preventing overfitting and notes that using simple models (like linear regression) is itself an effective form of "regularization," limiting the model's degrees of freedom and avoiding over-interpretation of noise.

Positions Mentioned

This interview did not discuss specific investable positions (companies, funds, etc.), focusing primarily on methodological and philosophical aspects.

Judgments Worth Remembering

1. Models are "necessary approximations," not truth. (Jeremiah Lowin) — Using the "football field" vs. "acre" analogy, he illustrates that models inherently contain errors; understanding the accumulation and boundaries of error is more important than the model itself.

2. The essence of machine learning is iterative optimization, not brute force or evolution. (Jeremiah Lowin) — By calculating the gradient of the error function, the model efficiently knows how to adjust parameters to reduce mistakes, which is key to its efficiency over other methods.

3. Data order equals context, and is key to applying machine learning in finance. (Jeremiah Lowin) — Citing Google Translate's progress, he points out that recurrent neural networks build context by processing ordered data, leading to a leap in translation quality.

4. Risk cannot be eliminated, only reallocated and transferred. (Jeremiah Lowin) — The art of risk management lies in understanding that when you remove one risk, it "mutates" and appears elsewhere in another form, and in continuously tracking this evolution.

5. The most effective risk management tool is qualitative conversation, not quantitative models. (Jeremiah Lowin) — Through "fire drill"-style stress tests and scenario analyses, "unknown unknowns" can be exposed in advance, training decision-makers' intuitive reactions during a crisis.

6. A portfolio's "evolution" is more important than its "current position." (Jeremiah Lowin) — By building a 3D risk exposure map, he found that tracking how risk changes slowly over time (the "boiling frog" effect) is more valuable than knowing the current risk exposure.

7. There is no true "passive investing"; choosing an index is itself an active decision. (Jeremiah Lowin) — He believes the line between "active" and "passive" is blurred; investors should focus on "activity level" and cost, not labels.

8. The greatest danger of machine learning is "overfitting," and using simple models is itself an effective preventive measure. (Jeremiah Lowin) — He warns that complex models can perfectly fit any noise, while simple models like linear regression, by limiting the number of parameters, naturally act as a "regularizer," forcing the model to seek more robust signals.