PodcastThe Monetary Matters NetworkDavid Wright

Generative AI Can’t Generate Alpha… But Machine Learning Can | Pictet’s David Wright

Pictet’s quant team relies on large gradient‑boosted tree ensembles trained on 15 years of daily tabular data to forecast company‑level 20‑day relative returns, prioritizing interpretability and explicit guardrails to avoid unintended factor or industry tilts. They deliberately avoid generative language models for production signals—using them only for productivity—because of opacity, hallucination risk and look‑ahead bias with text‑derived features. The research has been scaled into AI‑enhanced ETFs within a $30bn franchise, with roughly 0.40–0.45 of model view intensity converted into long‑only, benchmark‑aware positions and the methodology documented for investor review.

Monetary PolicyArtificial IntelligenceRisk Management

Preview · 11 more points in the app

Read the Full Summary

Key Takeaways

Pict trains very large gradient‑boosted ensembles (thousands of trees) on 15 years of daily features to forecast company‑level 20‑day relative returns, a short‑horizon cross‑sectional objective that directly underpins its AI‑enhanced ETFs (PQNT, PQUS).

The feature library deliberately mixes signal types (~25% sell‑side, market/price, accounting, other) and prioritizes historically alpha‑generative, expandable features; analyst‑derived signals often appear high in tree splits, but simple analyst signals have decayed and are most informative only when observed well before earnings (~10–30+ days).

Because the ETFs are long‑only and benchmark‑aware, the portfolio only maps roughly 0.4–0.45 of the model's view distribution into active positions and cannot fully express extreme negative views, a constraint combined with explicit training guardrails to avoid unintended market, factor or industry tilts while targeting modest 1–2% annual outperformance.

The firm avoids using generative AI for production return forecasting—citing opacity, hallucination risk and difficulty of testing—and similarly resists text‑derived signals because short archival history and severe look‑ahead bias would require impractical daily retraining; instead, generative models are limited to non‑production tasks (coding, summaries, client materials).

TECHNOLOGY

Tree Ensembles — The firm trains very large gradient‑boosted ensembles of decision trees added sequentially so each new tree learns prior errors, producing models with thousands of trees whose early splits have outsized influence on forecasts.

“The form of machine learning that we are using is based on decision trees. Thousands and thousands of decision trees that work very effectively…”—

Feature Mix — Their feature library splits roughly 25% sell‑side, 25% market/price, 25% accounting and 25% other, prioritizing historically alpha‑generating, economically grounded and easily extensible features; analyst‑derived signals often rank high in tree splits and carry strong importance.

“So I understand that in terms of what the signals are, what the features are, roughly 1/4 would be sellside information…”—

20‑Day Forecasts — Strategies forecast company‑level 20‑day relative returns using hundreds of input features and 15 years of daily history, training on rolling 20‑day returns cleaned of market, sector and country effects to support P1 and PQUS objectives.

“What we're essentially wanting to do to try and generate a consistent outperformance above the index in those two strategies you highlighted P1 and PQU US…”—

MARKETS

Fund Scale — Pict’s quantitative franchise manages over $30 billion and has productized years of research into AI‑enhanced ETFs (PQNT, PQUS and regional variants) positioned as hybrid active‑passive funds targeting modest annual outperformance.

“Very pleased today to be joined by David Wright, co-head of Pict Asset Management's quantitive investments franchise, which manages over 30 billion dollars in their quant group…”—

Benchmark Limits — As long‑only, benchmark‑aware ETFs, the funds cannot fully express extreme negative views; they map roughly 0.40–0.45 of the model's view intensity into active positions, a partial mapping intended to target 1–2% annual outperformance.

“If I sort of give you a statistical way of describing this, if you could perfectly reflect the positive views exactly is their distribution into the active positions above the…”—

Related Episodes

Important Disclaimers

Not Investment Advice: For education only — not financial, tax or legal advice.

AI-Generated: Made by AI from source transcripts and may contain errors.

No Guarantees: Past performance does not predict future results; investing carries risk.

Our Commentary · Request Removal