This interview covers the evolution of quantitative investing (using computer models to pick stocks). Eric Sorensen, who manages $46 billion, says quant strategies must constantly adapt. He's bullish on biotech/pharma because his team built a model using Markov chains (a probability tool) to analyze drug trial success rates—it's been running for 10 years. He warns value investing works long-term but can falter recently. Key holdings: ① Biotech/pharma companies—their model buys firms with above-average success rates; ② Commercial banks—they used FDIC (US deposit insurer) loan data to outperform in 2008; ③ China A-shares—they used NLP (reading text) to gauge chatroom sentiment, but didn't profit, just learned.
Eric Sorensen (CEO of Panagora Asset Management, managing over $46 billion in assets) discussed the evolution of quantitative investment strategies on the program. Key points include: quantitative research has evolved from early simple factors (such as value and momentum) to complex machine learning
Eric Sorensen (CEO of Panagora Asset Management, overseeing over $46 billion) brings a background as an Air Force pilot and 40 years of experience in quantitative research. This episode explores the evolution of quantitative strategies from simple factor models to cutting-edge technologies such as machine learning and natural language processing. The core judgment is: quantitative research must continuously iterate, establishing a "research graveyard" to record failed signals, while active management should integrate expert systems with portfolio construction—Eric Sorensen believes that "discovery" and "dollars" are the dual forces driving outstanding quantitative institutions.
Eric Sorensen divides quantitative financial research into four evolutionary stages:
1. Asset Pricing Equilibrium Theory — The CAPM model (Treynor, Sharpe, Mossin, Lintner), describing "how the world should be"
2. Efficient Market Hypothesis — Fama and others, describing "competition makes markets difficult to beat"
3. Anomalies and Behavioral Finance — Empirical findings such as small-cap stocks and earnings surprises, with behavioral explanations (people buy and sell at the wrong times)
4. Asset Pricing Econometrics — Sorensen's own research area: "Can we better know where bonds or stocks should trade?"
Sorensen notes that his doctoral dissertation studied municipal bond pricing, using multiple regression models with bond yields/prices as dependent variables and independent variables including credit ratings, municipal information, yield curve slope, and issuance structure. Key principle: The model must have intuitive causality — "Even when using advanced tools, it cannot simply be a correlation between price and anything that might be related."
Eric Sorensen proposes a lifecycle framework for risk premium strategies—the "5C":
| Stage | Description |
|---|---|
| Character | Creative ideas are discovered and may genuinely work |
| Capacity | Strategies are replicated, capital floods in |
| Collateral | Becomes crowded, supply of investable assets is limited |
| Conditions | External catalysts (e.g., Fed policy) invalidate assumptions |
| Capitulation | Liquidity dries up, eventual collapse |
Historical case: The portfolio insurance strategy (LOR strategy by Leland, Rubinstein, and O'Brien) during the 1987 crash. This strategy used futures markets to hedge equities but grew too large. When the market fell in October 1987, portfolio insurance programs triggered a cascading decline in both futures and cash markets, ultimately causing the Dow to plunge 22% in a single day. Key lesson: Using options hedging (e.g., BEA selling out-of-the-money puts) could have avoided this disaster—since that day, the implied volatility of S&P at-the-money puts has always exhibited a skew and has never been cheap again.
Sorensen believes value investing is a "true long-term proposition"—if one had started value investing 30–40 years ago and held clients for five years, wealth would have consistently accumulated (stochastic dominance). However, the problem is that signals sometimes fail (e.g., poor performance of quantitative equity strategies in recent years), possibly due to crowding or other factors.
Eric Sorensen emphasizes "smart data" over "big data":
> "I call this smart data and big beta. Data must be smart. Big data vendors have huge average revenues, but many lack clearly defined business models—they just happen to know certain data."
Key principles: Smart data must satisfy three conditions:
1. Intuitiveness — makes sense (e.g., what sell-side analysts are doing)
2. Collectability — can be prioritized for access (not purchased from well-known vendors)
3. Reasonable payback period — rarely uses "tomorrow's price" as the dependent variable (non-stationarity)
Case 1: Biotech Model
The team built an expert system simulating analysts' understanding of the FDA approval process. Using Markov chain analysis and probability models to assess drug trial success rates—biotech companies have extremely low success rates, and the model buys companies with "above-average success rates." The model has been running for 10 years and is now an independent biotech/pharmaceutical portfolio.
Case 2: Financial Stock Model (2007–2008)
The team discovered that the FDIC (Federal Deposit Insurance Corporation) collected delinquency data on commercial bank loan portfolios (percentage of 1-month, 2-month, and 3-month delinquencies). By accessing this data, they could discount bank asset values for more accurate valuations. Sorensen noted: "In 2007–2008, our financial stocks performed much better than most." However, this data has now been commercialized—vendors purchase it from the FDIC and resell it.
Case 3: NLP in China's A-Share Market
The team used natural language processing to read Chinese chat room content and identify sentiment. Due to censorship in China, investors use slang to express negative sentiment (e.g., "辣鸡" instead of "垃圾"). Unsupervised machine learning was used to identify these expression patterns. Sorensen admitted: "I don't think we made money from it, but we learned a lot."
Eric Sorensen takes a cautious view of AI/machine learning applications in financial markets:
> “No matter how many possibilities there are in chess or autonomous driving, they are finite, rule-based, and deterministic. In financial markets, neither of these holds true — you have Brownian motion, stochastic processes, the outcomes for security prices are nearly infinite, and they are not rule-based.”
Applicable Areas:
Sorensen emphasizes: “If the engineer or designer is unaware of the underlying limitations and pitfalls, trees must be pruned — you cannot let them grow fully, otherwise you will overfit and obtain spurious results.”
Eric Sorensen introduced the concept of a "research graveyard"—a repository for all failed hypotheses.
Failure case: Volatility forecasting
The team collaborated with Nobel laureate Robert Engle, using ARCH/GARCH models to forecast volatility. Sorensen admitted: "I knew we wouldn't make money from it, but it was fascinating." The market is quite efficient at bridging the gap between realized volatility and statistically predicted volatility.
Criteria for identifying successful signals:
Decision framework for data procurement:
> "If the vendor says it costs $200,000 a year... you can buy this data, very expensive. Or you can go across the street to a window washer or a plumber—that building has 40 tenants, and that person has access to know who the tenants are. Trade organizations have this information. It's free."
Panagora’s active equity strategy is divided into two types:
1. Dynamic Equity
2. Deep Industry Models
Model Update Frequency: Factors may persist for 4–5 years, with a half-life of approximately 2 years. The team continuously monitors performance and uses commercial models to evaluate the marginal value of new datasets.
Eric Sorensen believes concentrated portfolios are a safe harbor for non-quantitative investors:
> "Concentrated portfolios. Forget alpha-beta separation—that is disappearing. You cannot have a single market-cap-weighted MSCI that everyone must track. Those who can truly pick stocks and hold concentrated portfolios, I believe, have their entire world."
Key conditions:
Modern Olympics analogy: Smart beta is like the specialization of the modern Olympics—no longer the pentathlon (one athlete doing everything), but a collection of specialists: sprinters, hammer throwers, hurdlers, etc. "You wouldn't have a 165-pound sprinter throw the javelin, or a 300-pound shot putter sprint."
Eric Sorensen proposes a Venn diagram framework of "three circles":
| Circle | Content |
|---|---|
| Domain Knowledge | Understand the industry: What do traders do? What drives daily price movements? Understand market makers, buy-side clients, stickiness, cycles |
| Analysis/Algorithms | Computational expertise and algorithms learned in school |
| Data | Understand data quality — commercially available data differs from what a university can afford |
Sorensen notes: "I know someone who manages a large asset management firm. He asked me: 'You have good quantitative analysts. We go to universities to recruit people with advanced statistics degrees. Not all of them succeed.' — They only have one circle (technical skills), without understanding the industry or data quality."
Advice: Understand that all three circles are important. "The story of my colleague George Masali is that he spent a lot of time thinking about what fundamental analysts do. That made him a good quantitative analyst."
| Position | Analyst Stance | Key Data |
|---|---|---|
| Biotech/Pharmaceutical Companies (Industry) | Bullish (with independent model) | Model runs for 10 years, using Markov chain to analyze drug trial success rates |
| Commercial Banks (Industry) | Risk warning (but model effective) | Used FDIC loan delinquency data from 2007-2008, outperformed peers |
| China A-Share Market | Neutral (experimental) | NLP reads 50,000 Chinese characters, identifies slang sentiment such as "trash" |
| Portfolio Insurance Strategy (LOR) | Risk warning (historical case) | Strategy collapsed during the 1987 crash, market fell 22% in a single day |
1. Eric Sorensen: "Quantitative research has gone through four historical stages — from CAPM to the Efficient Market Hypothesis, to anomalies, and then to asset pricing econometrics. We are currently in the fourth stage, but the tools and data are fundamentally different."
2. Eric Sorensen: "Risk premium strategies go through a 5C lifecycle: Characteristics → Capacity → Collateral → Conditions → Capitulation. Portfolio insurance in 1987 is a classic case — the strategy became too large, an external catalyst triggered it, and ultimately liquidity dried up."
3. Eric Sorensen: "Smart data is better than big data. It must be intuitive, have priority access, and offer a reasonable payback period. The FDIC's loan delinquency data is a model of smart data — it is free, unique, has causal logic, but has now been commercialized."
4. Eric Sorensen: "Machine learning has fundamental limitations in financial markets — the outcomes for security prices are nearly infinite and are not rule-based. However, decision trees/random forests and NLP are effective applications, provided engineers know the limitations and prune the trees to prevent overfitting."
5. Eric Sorensen: "Factors may last 4-5 years, with a half-life of about 2 years. You need continuous monitoring and use a business model to evaluate the marginal value of new datasets — 'Will it add 3 basis points or 4 basis points? What is the cost?'"
6. Eric Sorensen: "The safe harbor for non-quantitative investors is a concentrated portfolio — forget alpha-beta separation. Those who can truly pick stocks and hold a concentrated portfolio have their own world, but you cannot constrain them (cannot require a beta of 1)."
7. Eric Sorensen: "Young quantitative researchers need to understand three circles: domain knowledge, analysis/algorithms, and data. Those who only know technology (one circle) will not succeed — spend time thinking about what fundamental analysts are doing to become a good quantitative analyst."
8. Eric Sorensen: "It is important to study the graveyard of research — we did a lot of volatility forecasting work (in collaboration with Nobel laureate Robert Engle), knowing it would not make money, but we learned a lot. Failure signals are just as valuable as success signals."