AI Can Build Your Portfolio—But Can You Trust It? | Systematic Investor | Ep.419
The discussion examines how agentic AI could support institutional portfolio management, while keeping human fiduciaries and boards responsible for review and final decisions. It also weighs the limits of LLM-based backtests, including look-ahead bias, inconsistent outputs, and the risk that agents using the same model produce correlated views. Market themes include unusually high negative betas among S&P 500 constituents, the role of rates in trend following, and the difficulty of timing regimes and volatility. The guest also highlights how AI could help balance competing portfolio objectives, while arguing that investment expertise and decision-making capacity remain essential.
Preview · 22 more points in the app
Read the Full SummaryKey Takeaways
The guest presents agentic AI as a way to abstract parts of institutional asset management, not to remove human fiduciaries: people would review recommendations, retain decision authority, and use a human-written policy statement to guide the CIO agent, with the board making the final decision. The guest also says natural-language debate among agents can make their process easier to audit, even though it remains stochastic.
The proposed agent framework combines regime classification—including confidence and the risk of misclassification—with an adversarial diversifier that takes positions opposite a principal component built from other agents. But using the same major-company LLM across agents could correlate their outputs and reduce the diversification the framework is intended to provide.
The guest warns that historical LLM backtests can contain look-ahead bias because models may encode future knowledge, even when given restricted historical inputs. The speakers add that agents may favor historically successful portfolio schemes and may answer the same prompt differently, complicating both backtest interpretation and reproducibility.
In the paper’s March 2026 example, the CIO chose an allocation different from the agents’ most popular choice, with the guest interpreting the 44% equity weight—versus a 60/40 benchmark—as potentially reflecting mixed macro signals and a late-cycle, stagflationary environment. The guest argues that mixed regime evidence favors diversification, while greater confidence may support concentration, but misclassification can lead to an incorrect allocation and losses.
TECHNOLOGY
“You start seeing how the you know the transition into a gentic AI could have implications in many things that you know historically we used to do using people and…”—
“the first one you have let's say a micro agent so the micro agent will do what will basically look into data the economic IC regime, macro indicators, market indicators…”—
“The last one and I found that very very interesting is called like know an adversarial diversifier…”—
MARKETS
“So what it shows is that sorry it shows basically the percentage of S&P 500 constituents so 500 stocks in the index that have a negative beta to the index…”—
“But I suppose my perspective on this is this kind of tendency amongst analysts to view the fundamentals and then the market reaction as being very much linear and proportionate…”—
“yeah, no at the minimum at the minimum I think rates have been the strongest contributor to trend following strategies.”—
Introduction and what’s on Nick’s radar
Nick highlights research showing nearly 45% of S&P 500 constituents have negative three-month betas to the index, an extreme concentration effect driven partly by AI megacaps.
The extraordinary concentration inside the S&P 500
Alan describes rising global yields and the debate over growth, fiscal deficits, AI investment, and higher-for-longer rates; Nick cautions that markets do not react to fundamentals in a simple linear way.
Why global bond yields are moving higher
Trend-following benchmarks have posted strong monthly and year-to-date gains, with short fixed-income positions contributing most, while other asset classes provide diversification.
A strong month and year for trend following
Despite a substantial rise in yields, bond volatility has remained contained, allowing trend signals and positions to benefit without the move being overwhelmed by choppiness.
Why the bond move has happened without a volatility spike
The speakers discuss research testing volatility-dependent trend lookbacks and note that selecting assets from a small ETF universe differs from a directional, risk-scaled trend-following strategy.
Can volatility make trend following better?
The study finds little reliable improvement from VIX-triggered lookback changes; a few observations and isolated performance episodes make it difficult to establish statistical value.
What the latest research actually tells us
Nick explains that volatility matters through continuous signal scaling, portfolio-level risk targeting, and correlation-aware risk budgets rather than simple regime triggers.
How Nick uses volatility in portfolio construction
Nick argues that adding parameters to trend models may not solve trendless periods; diversifying into complementary strategies can be more prudent than forcing the model to manufacture signals.
Why making trend models more complicated may not help
Faster models may respond well around market turns, but dynamically anticipating reversals is difficult, and infrequent trigger events make their effectiveness hard to prove.
V-shaped markets and the problem with regime triggers
Recent sharp V-shaped markets can challenge unconstrained trend followers, though the speakers note that the latest episodes have not erased the strategy's broader appeal.
How agentic AI could reshape asset allocation
Andrew Ang's proposed agentic framework assigns specialized agents to assess regimes and assets, construct and review portfolios, and prepare a CIO recommendation for human oversight.
What actually drives an AI-built portfolio?
In the paper's example, the CIO agent selects a materially underweight-equity portfolio, apparently favoring risk-based and heuristic methods amid uncertain, stagflationary macro signals.
The problem with backtesting AI investment models
The speakers identify look-ahead bias, data mining, inconsistent LLM outputs, and monoculture risk as major obstacles to trustworthy AI investment backtests.
Could the same AI framework work for trend following?
Nick sees potential to adapt parts of the agent workflow to trend research, especially comparing portfolio-construction methods, while noting that the framework's debate and voting process may not transfer directly.
Using AI to manage multiple investment objectives
Agentic tools may make it easier to balance multiple and sometimes conflicting portfolio goals, but investment expertise is still needed to make outcomes sensible and usable.
When will we see portfolios run by AI agents?
AI agents could expand institutional decision-making capacity and improve auditability, but stochastic outputs and fiduciary responsibilities mean humans remain accountable for portfolio decisions.
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