Momo AnalyticsResearch · February 24, 2026
Finance and AI- where's the value?
Most conversations about AI in finance start with judgment: will it pick better stocks, size better positions, call the top of a cycle. That’s the exciting question, so it’s the one that gets asked first. It’s also, so far, the wrong place to look. The controlled studies that actually measure AI performance against a baseline point somewhere less exciting and much better supported: engineering work with a clear, checkable answer.
Coding is the strongest, cleanest result in the entire literature. GitHub ran a randomized controlled trial with 95 professional developers building an HTTP server in JavaScript. The group using Copilot finished 55.8 percent faster, a result with a tight enough confidence interval and low enough p-value that it isn’t noise. That’s a lab result, but it held up inside a real bank: a six-week study at ANZ measured developers completing production tasks 42.36 percent faster with Copilot, with the biggest gains going to less experienced developers. Accenture’s internal rollout found smaller but still real numbers on live teams- 8.69 percent more pull requests per developer, 11 percent higher merge rates, and 84 percent more successful builds. Three different studies, three different methodologies, the same direction and a similar magnitude.
Forecasting and judgment tell a different story. ForecastBench, a continuously updated benchmark that scores real-world prediction questions, put trained human superforecasters at a Brier score around 0.096- lower is better. Frontier models like Claude 3.5 Sonnet and GPT-4 Turbo scored between roughly 0.121 and 0.159 depending on the setup, landing close to the general public’s score of about 0.121. That’s a meaningful gap: models are matching an average informed person, not the experts who make a living being calibrated. A separate study specifically testing judgmental forecasting- the kind that requires adapting to a shifting situation rather than pattern-matching a stable base rate- found LLM forecasts significantly less accurate than human forecasters in exactly those adaptive, uncertain scenarios.
It is more than just “engineering good, judgment bad.” A University of Chicago Booth study found GPT-4 actually beat human analysts at predicting earnings changes from financial statements, 60 percent accuracy against a 53 to 57 percent range for humans. That looks like a contradiction until you look at what the task actually was: a structured prediction problem with defined inputs, years of historical patterns to draw on, and a checkable right answer. It’s an engineering task wearing a judgment costume. The real dividing line isn’t “coding versus forecasting.” It’s whether the problem has a stable structure and a feedback loop, or whether it requires adapting to a situation that hasn’t happened before. Earnings prediction from known financial-statement features has the former. Sizing a position mid-regime-shift, or deciding whether this time actually is different, has the latter- and that’s exactly where the human edge shows up in the data.
Humans are typically the bottleneck and not the model. A comparative study of human, AI, and hybrid financial advice found the AI and hybrid groups more than doubled the portfolio improvement of the human-only group- but the driver wasn’t that human advisors gave worse advice. It was that the human condition had a significantly higher rate of clients rejecting good advice outright. That’s not a capability gap. That’s an adoption and trust problem, which is itself an engineering problem: how the recommendation gets surfaced, explained, and integrated into someone’s actual workflow determines whether a good model output ever turns into a decision.
None of this shrinks the value of judgment, however. CFA Institute’s own workflow research on AI adoption across investment roles found something counterintuitive: as firms automate more of the data-gathering and modeling stages, demand for the judgment stage doesn’t fall. It grows. That tracks with what the forecasting data shows- a junior analyst’s day used to be justified partly by the hours spent pulling and cleaning data before getting to an opinion. AI collapses those hours. What’s left, position sizing under regime uncertainty, knowing when to override the model, is the part with no engineering shortcut yet, and it’s now a bigger share of what the job actually is.
So where do we spend the money? Money spent automating financial-services engineering- data pipelines, reconciliation, code review, report generation- is buying a well-documented, replicated 40 to 55 percent productivity gain today. Money spent trying to automate discretionary investment judgment is buying a system that calibrates like an informed member of the public, not like the expert it’s meant to replace. That gap will close over time as forecasting benchmarks improve. It hasn’t closed yet, and the firms treating the two categories the same are misallocating the budget.
Sources: Peng et al., “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot” (2023); Chatterjee et al., ANZ Bank Copilot study; Accenture-GitHub Copilot enterprise trial results; ForecastBench (Karger et al., 2025); Schoenegger et al. on judgmental forecasting under uncertainty (2024); Kim, Muhn & Nikolaev, University of Chicago Booth, on LLM earnings prediction accuracy; comparative human/AI/hybrid financial advice study, ScienceDirect (2026); CFA Institute, “Creating Value from Big Data in the Investment Management Process” and related workflow research.