Will AI Replace Investment Analysts—or Change Their Jobs?
Winning a forecasting test is not the same as replacing a job. What the research supports, and what it leaves unanswered.
An investment analyst does more than read financial statements. The job includes deciding which questions matter, checking accounting adjustments, investigating a business, valuing uncertain outcomes and explaining a recommendation. A model can improve one of those tasks without taking over the whole job.
What the studies actually test
Cao and colleagues built a machine-learning analyst to forecast stock prices using financial, textual and economic information. Their historical forecasting tests covered 2001–2016. Adding human analysts’ forecasts to the machine’s inputs improved the combined model’s performance. Human contributions were particularly useful where institutional knowledge mattered. [1]
That is evidence of complementarity, but the setup matters: the hybrid was a statistical model given analysts’ forecasts, not an experiment in which employees used a chatbot. It does not establish how many analysts a modern investment team needs.
Kim, Muhn and Nikolaev report that GPT-4 could predict the direction of earnings changes from standardised, anonymous financial statements, outperforming their analyst comparison on that task. Their reported result concerns a specific input and forecast—not management assessment, valuation, client communication or an entire research process. [2]
Tasks are not jobs
Consider an earnings release. Extracting the reported figures is one task. Checking whether an acquisition changed their comparability is another. Deciding whether an apparently temporary margin decline reflects a lasting competitive problem is another again.
A useful adoption test separates those tasks. Measure whether the tool extracts figures correctly, preserves units, identifies its sources and flags missing information. Evaluate forecasts against a relevant baseline. Then examine whether the resulting research actually changes a decision for the better.
Saving preparation time, improving forecast accuracy and improving investment returns are different outcomes. Success at the first does not establish the other two. Nor does a polished explanation establish that the underlying analysis is correct.
Human oversight must do real work
Keeping a person in the workflow is not, by itself, a quality standard. Someone must independently check the facts and assumptions. A reviewer who simply accepts a fluent answer has not provided that check.
For an investment team, a practical trial would give the AI-assisted and existing processes the same information and deadline. Compare factual errors, forecast quality, time spent correcting outputs and the decisions reached. Include difficult cases, not just clean statements and familiar companies.
Responsibility also needs an owner: who approves an adjustment, resolves contradictory sources and signs off the recommendation? A tool’s confidence is not a substitute for those decisions.
The narrower answer is more useful
The research supports investigating combinations of machine processing and human knowledge. It does not provide a profession-wide employment forecast. Whether a particular task can be automated—and whether doing so improves a team’s work—must be tested in its actual setting.
Instead of asking whether AI can replace “the analyst”, ask which work it can perform reliably, which checks remain necessary and what evidence would justify changing the workflow. That produces an actionable decision rather than a prediction about everyone’s career.
Sources
- Cao et al. (2021 working-paper version). From Man vs. Machine to Man + Machine: The Art and AI of Stock Analyses. NBER 28800.
- Kim, Muhn and Nikolaev (2024). Financial Statement Analysis with Large Language Models. Preprint; the reported task and result are described in its abstract.