CFA Institute

Other recent work documents systematic biases in LLM-based financial analysis, including foreign bias in cross-border predictions (Cao, Wang, and Xiang, 2025) and sector and size biases in investment recommendations (Choi, Lopez-Lira, and Lee, 2025). Building on this emerging literature, four potential channels are especially relevant for investment practitioners: 1. Size bias: Large firms receive more analyst coverage and media attention, therefore LLMs have more textual information about them, which can translate into more confident and often more optimistic forecasts. Smaller firms, by contrast, may be treated conservatively simply because less information exists in the training data. 2. Sector bias: Technology and financial stocks dominate business news and online discussions. If AI models internalize this optimism, they may systematically assign higher expected returns or more favorable recommendations to these sectors, regardless of valuation or cycle risk. 3. Volume bias: Highly liquid stocks generate more trading commentary, news flow, and price discussion. AI models may implicitly prefer these names because they appear more frequently in training data. 4. Attention bias: Stocks with strong social media presence or high search activity tend to attract disproportionate investor attention. AI models trained on internet content may inherit this hype effect, reinforcing popularity rather than fundamentals. These biases matter because they can distort both idea generation and risk allocation. If AI tools overweight familiar names, investors may unknowingly reduce diversification and overlook under-researched opportunities. source

The voluntary CFA Institute Inclusion Code (UK) provides investment industry organizations with a structured framework for supporting executive leaders, investment professionals, Inclusion specialists, and HR professionals in crafting impactful and measurable strategies to enable better people management. source

Cognitive abilities describe how humans collect, process, and interpret information such as attention, memory, pattern recognition, logical reasoning, and quantitative analysis. Non-cognitive abilities include traits such as motivation, perseverance, communication, ethical judgment, and the capacity to act under uncertainty. The framework below categorizes these capabilities across two dimensions: cognitive versus non-cognitive, and basic versus advanced. Basic cognitive capabilities (QIII: third quadrant), such as memorization, structured record-keeping, and routine calculation, have long been automated. Their automation marked the first wave of technological compression. Advanced cognitive capabilities (QII), including high-dimensional modeling, statistical inference, and complex analytical verification, are increasingly within the reach of AI systems. As these tools scale across firms, analytical differentiation narrows. By contrast, advanced non-cognitive capabilities (QI), such as setting goals under uncertainty, exercising ethical judgment, and creating or obtaining first-order information, remain less amenable to standardization. These capabilities influence how organizations interpret ambiguous signals, coordinate decisions, and allocate capital when data is incomplete. The implication is organizational rather than purely technical. When analytical tools become widely accessible, sustainable advantage depends less on computational sophistication and more on how firms structure teams, cultivate judgment, and design decision processes that integrate technology with human insight. source

source

Fangfang Wang, Florina Silaghi, Steven Ongena, and Miguel García-Cestona CDS spreads rise sharply after ESG downgrades — most notably within the social pillar and among financially constrained firms — while upgrades show little effect. Positive ESG sentiment and transparency can mitigate these adverse credit impacts. source

Institutional allocators rely on managed futures strategies for diversification and drawdown control, yet often misunderstand how risk is actually taken inside these allocations. They frequently lack clarity on which trend horizons drive performance, how similar managers truly are to one another and to benchmarks, and how differences in horizon mix shape behavior during periods of market stress. By decomposing CTA managed futures returns into a small set of distinct trend horizons (fast, medium, and slow), this post shows that much of the variation across managers and benchmarks reflects differences in horizon mix rather than fundamentally different strategies. Framing managed futures allocations in this way allows investors to better diagnose overlap, benchmark more precisely, and assess whether their exposure is aligned with its intended role in the portfolio. The analysis that follows is necessarily technical, introducing a horizon-based framework that decomposes CTA returns into a limited set of systematic building blocks. While the mechanics are described in detail, the objective is practical: to provide a clearer, more transparent way to interpret managed futures behavior and to link observed outcomes to explicit, governable risk choices. source

What are the best first use cases?Start where state, action, and reward are clear and the feedback cycle is short: adaptive trade execution, dynamic portfolio rebalancing, and cost-aware option hedging. These map cleanly to RL/POMDPs, have measurable baselines (e.g., time-weighted average price/volume-weighted average price [TWAP/VWAP], discrete delta), and abundant historical data for offline training. Can I train only on historical data, or do I need live exploration?You can (and usually should) start with offline RL using your fills, prices, and positions. Then validate in a high-fidelity simulator with costs/impact/latency, run shadow mode alongside your existing process, and promote gradually with guardrails (caps, kill-switch, rollback). How do I build risk and costs into the objective?Make risk and costs part of the goal. Define the reward as the money you make after subtracting trading fees/price impact and a penalty for risk. In words:Reward = Profit − Costs − λ × Risk (risk can be tail risk, such as CVaR, drawdown, or mean–variance). Use distributional RL to capture rare big losses (“the tails”). And set hard limits — on exposure, turnover, and market participation — both while training and when the system runs live. IRL versus imitation learning — when do I use which?Use IRL to infer the underlying objective from behavior (managers, clients, “the market”) when you want portability and the ability to surpass demonstrations. Use imitation to quickly mimic actions when you don’t need a reward function. Ranked data? Consider T-REX. Probabilistic, flexible rewards? MaxEnt/Bayesian (GPIRL). What metrics should I monitor to know the policy is working?At minimum, track implementation shortfall (IS) for execution quality, risk-adjusted return after costs (e.g., Sharpe or mean–variance utility) for performance, and CVaR/drawdown for tails. Add drift detectors (feature, policy, regime) and compare to baselines (TWAP/VWAP, risk parity, discrete delta). How do I make the RL/IRL policy compliant and explainable?Log state → action → outcome with immutable audit trails; publish a “policy card” (objective, constraints, data lineage, promotion criteria); add explainability (feature attribution, counterfactuals), runtime guardrails (exposure/participation/loss caps), challenger policies, and human-in-the-loop approvals. These actions turn the model into an accountable decision system, not a black box. source

Since its inception in 1945, the Financial Analysts Journal (FAJ) has advanced some of the investment profession’s most influential ideas by providing an outlet for innovative thinkers. We trace the FAJ’s history by identifying the most prolific contributors and innovations featured over its first 80 years and in each of nine financial eras. Using the comprehensive database and rigorous methodology that we developed, this article provides rankings of the top authors and the most frequent words in titles and examines the context in which these words were used to identify seminal ideas and the authors behind them. source