Stacie Mintz – Turning Qualitative Fundamentals into Quantitative Factors (S7E33)

Finance-focused summary (Stacie Mintz – “Turning Qualitative Fundamentals into Quantitative Factors”)

Disclosures / compliance

• Podcast disclaimer (Corey Hoffstein / Flirting with Models / Newfound Research):
  • Informational only; not investment advice.
  • Newfound Research may hold positions in discussed securities.
  • Opinions are solely those of participants.
• Corey Hoffstein notes he will not discuss Newfound Research funds due to industry regulations.

Presenter(s) / source(s)

• Corey Hoffstein — Co-founder & Chief Investment Officer, Newfound Research
• Stacy Mintz — Managing Director, Head of Quantitative Equity, PGIM Quantitative Solutions

Key finance ideas and takeaways

1) Portfolio construction: “alpha + risk” and why in-house risk models matter

• Mintz frames the portfolio as the combination of:
  • Alpha model
  • Risk model
• Why build risk models in-house (since 1999) vs. using off-the-shelf tools (e.g., Barra)?
  • Off-the-shelf risk models can treat desired exposures (e.g., value) as risk, which may lead to unintended imbalances between alpha themes.
  • In-house control enables tailoring risk to target the risks they care about and to balance exposures intentionally.
• Example: crowding & “quant ecosystem” concentration
  • If multiple quants use similar models, they may end up liking the same stocks.
  • When portfolios across the “quant ecosystem” are optimized, this can over-concentrate in the same names.
• Stress behavior example (“quant quake” — August 2007)
  • PGIM’s approach was described as acting more like a diversification engine.
  • Rather than trying to infer which individual stock is “less risky” from history, it diversified idiosyncratic risk differently using distinct risk modeling.

2) Post-GFC differentiation: “not enough to be a quant… be a different quant”

• After the GFC, Mintz describes a shift toward differentiation in a crowded factor world, including:
  • Use of alternative data
  • Expanded quality research beyond simple metrics like ROE
  • A research process aimed at producing scalable, feasible, differentiated signals

3) “Fundamental quant” vs. generic quant

• Practical distinction: they aim to emulate what a traditional stock picker would do—invest in:
  • Good companies
  • With growth prospects
  • High quality
  • At a reasonable price
• Emphasis on theoretical underpinning:
  • If a factor works, they want to understand why it works, when it works, and how it behaves across cycles.
  • They avoid purely backtest-driven signals that lack clear rationale (“backtest well but unsure why”).
• Resilience goal:
  • In stress, holdings should be resilient because they are good businesses, not just because they generate favorable statistical exposures.

Factor framework / taxonomy at PGIM Quantitative Solutions

Top-level factor groups

Mintz organizes factors into broad groups (useful for organization and client communication/attribution), while the internal model includes more complex interactions:

• Growth
  • Captures faster-growing companies and potential under-reaction to new information / slow diffusion.
• Linkages
  • A structural/information-flow framework across connected firms.
  • Examples described:
    • Industry shocks (moves together)
    • Customer–supplier relationships and multi-layer effects (“suppliers of suppliers”)
  • Goal: estimate how shocks at one firm can propagate to others (positive or negative).
• Quality
  • How well a company is run (management/competitive threat), including signals from markets.
  • Example: interpret options market information as a source of insight.
  • Also described as drawing from information sourced by other informed investors.
• Valuation
  • Ensures companies are reasonably valued—undervalued or under-appreciated.

Notably excluded: Momentum

• PGIM does not automatically include standard price momentum.
• Reason: price momentum blends:
  • company-relevant information
  • plus noise unrelated to the firm (e.g., speculation or shocks elsewhere)
• Mitigation via “information momentum”:
  • They seek company-relevant momentum tied to key events and how investors respond to those events.
• Claimed performance/risk attribute:
  • As effective over time as price momentum
  • with ~half the downside / crash risk (as stated)

Dynamic weighting by company type (systematic, not discretionary)

• Factor weights change at the stock level based on where a company sits on a growth/value spectrum.
• Conceptual mechanism:
  • For fast-growing / early-cycle companies:
    • weight growth factors more
    • reduce emphasis on valuation
  • For mature / stable companies:
    • weight valuation factors more (bargains)
    • rely less on “hype” growth
• Evidence/implementation principle:
  • “It is all systematic, not discretionary at all.”
  • They evaluate the company’s growth rate relative to the universe and adjust growth vs. valuation accordingly.

Methodology / step-by-step framework explicitly described

A) Research agenda process (“research shark tank”)

• An annual December event (“research shark tank”):
  • Researchers pitch 3–4 ideas
  • Team Q&A and critique (“poke holes” / challenge)
• Cross-team participation:
  • Portfolio management team plus researchers (from senior to junior)
• Idea filters:
  • Align with philosophy
  • Scalable
  • Feasible
  • Differentiated
  • Relevant

B) Backtesting discipline (especially around shocks like COVID)

• Evaluate factors as if they were actually managed in real time.
• Backtesting can be valuable but can also mislead.
• Red flag concept:
  • A factor that does “well every year” is “scariest,” suggesting overfitting or regime insensitivity.
• COVID “bar shift”:
  • COVID “compressed everything,” revealing mismatches between strategy behavior and stated philosophy.
  • If a strategy is truly philosophy-driven, it may not behave “as expected” during a regime like COVID.
• Short histories for newer factors:
  • Some factors/datasets have only ~5–7 years of history (vs. a prior preference for ~20 years).
  • Approach: use theory + understand weaknesses; consider lower weight initially.

C) Integrating LLMs (and NLP) into factor research safely

• Framework approach:
  1. Start with a concept or dataset to add.
  2. Use the simplest tool first; add complexity only if needed.
  3. Use LLMs to extract candidate signals, then validate rigorously.
• Risk controls:
  • Tight control of datasets during testing to reduce:
    • look-ahead bias
    • memorization
    • hallucinations
• Model selection:
  • Use models “published recently enough” to reflect modern techniques, while still validating out-of-sample with time separation.
• Validation approach:
  • Use multiple models and check response convergence.
  • Test stability under question rephrasing.
• Fundamental alignment:
  • Tie extracted insights back to the company’s sales or earnings to maintain “fundamental quant” alignment.

Emergent shocks: resilience and real-time adaptation goals

• Objective: produce a consistent stream of alphas across market conditions by emphasizing resiliency.
• Approach:
  • Use multiple sources of alpha / different factor types that perform in different regimes.
  • Identify information shocks before they show up in financial statements.
• Manager behavior under shocks:
  • Prefer limited discretion; remain within the model’s philosophy.
  • Example of exception: potential PM intervention for extraordinary events such as the Russia invasion of Ukraine.
• Learning loop:
  • Reassess outcomes and feed insights back into models so future shocks are handled better (not necessarily preventing all shocks).

Qualitative → quantitative using LLMs (board composition, innovation)

Why qualitative signals can be quantified

• NLP/LLMs benefit from abundant cross-country text data (e.g., earnings transcripts, news).
• Main challenge:
  • Extract valuable information from text and ensure apples-to-apples comparability across firms, sectors, and geographies.

Board composition example (mapping to linkages)

• Move beyond board “credentials” and examine connections among board members.
• Connectivity network as a proxy for information flow:
  • Well-connected high-quality board members sit on multiple boards.
• Stated finding:
  • Struggling companies with well-connected boards show more resiliency than peers.

Performance metrics / explicit numbers mentioned

• Risk model build-in-house: 1999
• Alpha/bias paper: 1999
• Strategy launch: 1996
• Research shark tank: every December
• Backtest history targets:
  • previously ~20 years
  • now sometimes only ~5–7 years
• “Information momentum” claim:
  • designed to have ~half the downside / crash risk vs traditional price momentum
• Signal count:
  • ~25–40 signals (“concepts”) depending on region
  • economically meaningful concepts: “more in the 40s”
  • contrast noted: 400 signals would be hard to attribute meaningfully

Instruments / tickers mentioned

• None explicitly mentioned.

Direct explicit recommendations/cautions

• Don’t replace an entire model with LLMs immediately:
  • start with concepts/data and use simpler approaches first.
• In research:
  • manage LLM risks through controlled data, model selection, multiple-model validation, and question perturbation tests.
• In backtesting:
  • treat “too perfect” performance cautiously—especially around shocks that can invalidate factor behavior.
• In shocks:
  • avoid throwing the model out; allow limited PM intervention only for truly exceptional events.