Asset management has spent the past decade accumulating data faster than it can interrogate it. Fund allocation and due diligence, in particular, remain surprisingly manual processes: analysts pore over factsheets, holdings disclosures and manager commentary to assess whether a fund’s positioning still makes sense. The data is there, but the challenge lies in translating thousands of individual security-level views into a coherent judgement on a fund itself., In other words, it is an aggregation problem. AI-driven scoring frameworks are beginning to close that gap, and for anyone considering where machine learning can add genuine value in investment workflows, it is worth understanding how they work.
From stock signals to fund-level insight
The approach is conceptually simple, even if the underlying modelling is not. Stock-level predictions, generated through learning-to-rank models trained across global equities, are combined with look-through holdings data for each fund. These models are built to learn how to order a group of stocks from most to least attractive. They do not need to be right about the precise magnitude of a return; they only need to identify the relevant order.. It is the same class of model that underpins the search engine results,, which face an almost identical problem: given thousands of candidate pages, which ones belong nearer the top?
Fund scoring, at its core, can be seen as a ranking problem. The models do not forecast exactly how much a stock will rise or fall; they forecast which holdings deserve more confidence than others That makes the output a close fit for how analysts actually use it: as a consistent ranking that helps them compare holdings quickly and focus their attention where it matters most.Each position a fund holds is scored individually; those scores are weighted by their allocation within the portfolio and then aggregated into a single, forward-looking fund score. A concentrated growth fund with a handful of high-conviction positions and a diversified income fund will arrive at their scores through very different weighting profiles, but the output is directly comparable: a rank that sits within a broader universe of funds, updated as frequently as the underlying holdings data allows. That makes the payoff concrete.Analysts can compare funds consistently without rebuilding the analysis from scratch each time holdings change.
Why model choice matters
Choosing the right type of AI is just as important as the modelling approach itself. While generative AI has attracted significant attention across financial services, it is not necessarily the most appropriate technology for systematic fund allocation or scoring. As noted in Morningstar’s 2026 report on AI in Active Fund Management: The State of Adoption in 2026, large language models are nondeterministic — the same prompt can produce a different answer tomorrow — and prone to look-ahead bias, since they are trained on internet-scale data that may already “know” how a historical period played out. That combination makes it difficult to build a reliable backtest for an LLM-driven process, or to prove, after the fact, that it actually added value.
Learning-to-rank models of the kind used in fund scoring do not suffer from that problem in the same way. The same inputs produce the same outputs, and the approach can be tested and audited across historical periods, as analysts have always demanded of quantitative signals. As AI adoption increases, a technique’s ability to demonstrate its track record might sound like a technicality, but it marks the fundamental divide between a claim and a verifiable demonstration.
A scalable second opinion
This produces a second opinion that scales. Analysts can use fund scores to challenge or validate their existing conviction on a holding: a fund ranked well outside consensus, or one whose rank has moved sharply, becomes a prompt for further scrutiny rather than a verdict in itself.
KPMG’s 2026 Global AI in Finance Report, which surveyed over 1,000 senior finance leaders across 20 countries, reinforces this point. It found that AI’s biggest benefits are emerging in judgment-heavy tasks, not just automating transactions. 70% reported better decision-making, 71% saw faster decisions, and 64% reported more accurate forecasts. Applied to fund scoring, that means AI can help analysts review more decisions, more quickly, and across a wider universe of holdings. The same scoring method can also support systematic strategies, ranking funds or ETFs and building portfolios from the top and bottom of those rankings, depending on the strategy’s rules.
Where the practical value lies
Signal generation at the individual security level has attracted the most investment and attention over the past few years. But the practical value increasingly lies in aggregating those signals coherently at the portfolio or fund level, with proper attention to weighting and coverage.
The bigger opportunity to treat AI outputs as a decision layer for portfolio construction, not just as stock-level signals. Once individual security scores are aggregated into a consistent, forward-looking ranking of funds or ETFs, the model becomes a practical tool for allocation rather than security selection alone. Investment teams can use these rankings to identify the most compelling opportunities within a broad universe, monitor changes in conviction over time, or build systematic strategies that rotate capital towards the highest-ranked funds.
From security selection to allocation
That is the shift worth watching: applying the same models one level up the investment hierarchy, from individual securities to investment vehicles. Used in this way, AI does not replace analyst judgement, but gives investment teams a consistent, scalable and testable decision layer that can sit naturally inside existing allocation processes.
