Everyone Has the Same AI. Where Does the Advantage Come From?
The same model can sit inside very different investment processes. A closer look at data, judgment, evaluation and implementation.
Imagine two investment teams using exactly the same AI model. They can still supply different information, ask different questions, reject different answers and take different positions. A shared tool does not make the rest of the investment process identical.
The question is which differences improve decisions, rather than merely making the workflow more elaborate.
The information around the model
A financial figure needs context: the reporting period, currency, accounting definition, publication time and any subsequent restatement. A model given the latest revised history is answering a different question from one restricted to information available before a trade.
These distinctions become important when evaluating an apparent edge. Was the information genuinely available? Was the company matched to the correct security? Were delisted firms included? More data is not automatically better evidence.
Cao and colleagues provide a useful historical case. In their study, analysts’ relative forecasting performance improved following the availability of alternative data, particularly at brokerages with stronger AI capabilities, measured using hiring data. [1] This is evidence about a particular setting, not a randomised demonstration that buying proprietary data improves every investment process.
The question determines the answer’s value
“Is this a good company?” differs from “What might the market have misunderstood?” Both differ from “Would this position improve our portfolio at the current price?” An eloquent answer to the first need not resolve the others.
For example, a hypothetical team might use a model to summarise an earnings release. Another might compare the release with earlier guidance and investigate which changes matter to a specific investment thesis. The second workflow asks a different question; that alone does not prove it will earn more.
To claim an advantage, define the decision and test whether the additional work helps. Otherwise, “better prompts” and “deeper insights” remain descriptions of a process, not evidence of its value.
A good signal can be used badly
Barber, Lin and Odean demonstrate why evaluation choices matter. Using US retail trades from 2010–2019, they reconcile two apparently conflicting findings: retail order imbalance predicts returns across equally weighted stocks, yet actual retail purchases concentrate in attention-grabbing stocks that subsequently underperform. [2]
This is not a study of AI. Its relevance is narrower: an attractive predictive relationship and the results of using it are different objects. Weighting and implementation can change the answer.
For an AI-assisted strategy, evaluate the actual proposed decisions, including timing, position sizes, costs and constraints. A model’s accuracy score cannot answer those questions on its own.
Test the surrounding process
Start with a baseline. Hold the model fixed and change one element: the input data, the research question or the decision rule. Use information genuinely available at the time, evaluate unseen cases and account for the cost of the change.
Keep what demonstrably helps; discard complexity that does not. The same discipline applies when a new model arrives. An upgrade can improve one task while adding cost or creating different errors elsewhere.
Shared AI access is neither proof that everyone has the same advantage nor proof that an advantage exists. The useful question is what the complete process does differently—and whether that difference survives a fair comparison.
Sources
- Cao et al. (2021 working-paper version). From Man vs. Machine to Man + Machine: The Art and AI of Stock Analyses. NBER 28800.
- Barber, Lin and Odean (2024). Resolving a Paradox: Retail Trades Positively Predict Returns but Are Not Profitable. Journal of Financial and Quantitative Analysis, 59, 2547–2581.