Theprocess by which individuals select stocks has long been a central question in finance and is often examined through theoretical models or analyses of trading data. Traditional financial models such as the Efficient Market Hypothesis or classical asset-pricing theories make strong assumptions about rational behaviour and homogeneous information processing. Yet, real human investors often behave in ways that deviate from these theoretical frameworks. The groundbreaking 2025 study by Hwang, Noh, and Shin, "How Investors Pick Stocks: Global Evidence from 1,540 AI-Driven Field Interviews," takes a radically different and more direct approach.

By conducting and analyzing a massive number of AI-driven interviews with actual investors across ten countries, the researchers move beyond what investors do to understand “why and how they make decisions.” The study's foundation is its pioneering use of large-scale, AI-driven field interviews. Engaging with real investors, including a significant proportion of millionaires, across a diverse set of ten countries provides a dataset of exceptional depth and global relevance.

AI-led interviews offer several advantages. First, they hold the interviewer quality constant. Second, they remove unwanted variation due to personal chemistry (or the lack thereof) between the interviewer and the interviewees. Most importantly, the AI agents are cheaper than human interviewers and enhance scalability. Traditional surveys and experiments have limitations: low response rates, interviewer bias, and reliance on predefined questions. These are typically limited to small samples because interviewers must be trained and dispatched individually to conduct conversations. In contrast, AI-led interviews can be initiated simply by sharing a link to a chat room.

The AI-driven interview structure is as follows:

Upon answering some background information questions, investors are directed to an AI-based interactive web-app that asked in their preferred language what led them to purchase specific stocks in the past. The AI agent then poses 10–15 follow-up questions to probe and clarify their motivations. Recent studies note that Artificial intelligence (AI) agents driven by Generative Pre-trained Transformers (GPT) can conduct interviews of quality comparable to those of human experts.

After the interviews, GPT was used to conduct a textual analysis of the transcripts. For each transcript, the analysis produces an argument map that traces the investor’s stated reasoning from signals, beliefs, and preferences through to the final buy decision. These maps are then compared across investors to identify recurrent mechanisms that consistently appear in practice.

This methodological innovation yields a rich, empirically grounded taxonomy of investor behaviour, challenging conventional theories and revealing a landscape of profound heterogeneity. The study's core contribution lies in its identification of thirteen recurrent decision-making mechanisms, its documentation of substantial diversity in how these mechanisms are applied, and its consequential argument for theoretical refinements in asset pricing.

These mechanisms capture almost the full range of preferences, and belief formation processes arrived from the sample. The thirteen mechanisms emerge entirely from the transcripts in a bottom-up fashion, without the imposition of any external theoretical priors. The mechanisms listed with the proportion of responses in parentheses for investors buying the stocks are as follows:

Fundamental Strength (40.5 per cent): when audited financial information indicates robust current performance, which, to investors, also signals high expected future returns.

Growth Innovation (37.6 per cent):when they see substantial growth potential tied

Familiarity / Brand Effect (28.4 per cent): Brand familiarity and positive product experiences generate intuitive comfort. Blue Chip Comfort (28.1 per cent): favor large, stable firms whose established operating history and low perceived risk create a sense of safety.

Authority / Follow (18.6 per cent):delegate stock selection to trusted experts or platforms, relying on source credibility rather than independent analysis.

Momentum (17.9 per cent):stocks with strong recent returns and high trading volume, as these stocks’ prices are expected to continue to drift upwards.

Confluence (14.6 per cent):act only when multiple independent signals from different sources give a buy recommendation.

Dividends (14.0 per cent):emphasise dividends over capital gains and purchase stocks whose dividend yields and payout stability satisfy their income requirements.

Social Copy (13.6 per cent):picks of trusted peers or family members, relying on relational trust rather than analytical verification.

Valuation / Mispricing (13.6 per cent):When valuation metrics indicate that a stock trades below its intrinsic value.

Buy the Dip (12.5 per cent):interpret temporary price declines as buying opportunities, as they believe in mean reversion.

ESG / Values (10.3 per cent):firms that meet their ethical or ESG standards.

Technical Analysis (6.8 per cent):when technical patterns generate predefined entry signals independent of fundamental considerations.

Only 0.8 per cent of interview transcripts remain unexplained, in the sense that none of the above mechanisms is a meaningful contributor to the purchasing decision. The study makes three broad observations and contributions to the finance literature. Rather than assuming investors follow a common rational model, research and teaching should adopt frameworks that account for multiple co-existing decision mechanisms; individual and cultural differences in financial cognition; and the blending of analytical and intuitive reasoning. First is a comprehensive and empirically grounded taxonomy of how actual investors feel and think about stock selection.

Second, the taxonomy could help refine existing finance theory. Several of the recurrent mechanisms that were identified aligned with existing theoretical frameworks. For instance, although Momentum is related to the extrapolation framework, a defining feature of the study is that investors believe recent price trends will continue only when accompanied by substantial trading volume. This line of thinking is currently absent from mainstream extrapolation models.

Similarly, while the ‘risk framework’ defines a ’safe’ stock as one with low covariance with adverse states of the world, investors in the sample invoking the Blue-Chip Comfort mechanism do not allude to concepts tied to covariance. Instead, safety is measured by whether the company is likely to survive in the future or by the stock’s stand-alone volatility. Neither of these two safety definitions appears in standard formulations of the risk framework.

The study also found that a non-negligible share of transcripts prominently feature mechanisms that do not align, or align only weakly, with mainstream asset pricing theory, such as Confluence or Dividends. For instance, traditional asset-pricing theory focuses on total returns and treats the two sources of total returns, capital gains and dividends, as perfectly substitutable, building on the dividend irrelevance. In contrast, investors invoking the Dividends mechanism consider dividends separately and only purchase stocks whose dividend yield meets their minimum income requirement.

The third principal contribution is the systematic comparison of how investors select stocks across countries. We observe substantial cross-country heterogeneity, even though the demographic composition of participants is relatively homogeneous across countries. For instance, Fundamental Strength is the most frequent mechanism among Indian investors (60 per cent) but only the fourth most frequent among US investors (30 per cent). Likewise, in Japan, Familiarity/Brand Affect ranks first, appearing prominently in nearly half of the transcripts (43.6 per cent), whereas in Singapore it ranks only sixth (19.3 per cent).

Further analysis from the data emerges three clusters endogenously. Investors in India resemble those in Singapore and South Korea. Investors in Australia, Canada, France, Germany, the UK, and the US form a largely similar cluster. Investors in Japan constitute a distinct group of their own. These observations should help the literature gauge to what degree results based on data from one country may extend to other countries.

This research opens several avenues for future inquiry, providing an indispensable foundation for a more behaviourally realistic and socially informed science, one rooted in the documented voices of investors themselves. The findings underscore the need for heterogeneous agent models, behavioural insights, and AI as a methodological tool in finance research. This study, thus, marks a major step toward bridging theory and the nuanced reality of global investor behaviour.

(The author is a partner at “Wealocity Analytics”, a SEBI-registered Research Analyst firm, and could be reached at [email protected])

Study Outcome

•13 bottom-up mechanisms explain nearly all stock purchases

•Findings challenge rational models, highlighting investor heterogeneity worldwide

•AI methodology enables scalable, unbiased, multilingual investor interviews

•Cross-country differences show culture shapes stock-selection behaviour globally