Thematic investing has long had an execution problem. Ideas are plentiful, spanning AI infrastructure, nuclear energy, rare earths, and geopolitics, but turning them into investable, rules-based benchmarks has traditionally required extensive research. We explore our partnership with Noonum, and how AI can support with thematic investing.
By Emir Kirdan, Product Manager, MerQube — August 4, 2026
Key Takeaways
Noonum is an AI-powered engine built for producing thematic baskets. MerQube partnered with Noonum, whose knowledge engine maps 150K+ public companies and 1M+ securities across 287 exchanges from 141 countries to turn investment themes into rules-based benchmarks.
Reproducible results. The same query on the same date always returns the same result. Company rankings follow the evidence in corporate filings, so passing market attention does not move them.
It catches what an LLM overlooks. Across five themes, Noonum surfaced names LLMs missed including micro-caps, foreign listings, and even Saudi Aramco for the Strait of Hormuz.
Thematic investing has long had an execution problem. Ideas are plentiful, spanning AI infrastructure, nuclear energy, rare earths, and geopolitics, but turning them into investable, rules-based benchmarks has traditionally required extensive research. That is why we partnered with Noonum.
What Is Noonum?
Noonum is an AI-powered knowledge engine built specifically for thematic investment analysis. Most AI investment tools are built on large language models, which are adept at synthesizing text but probabilistic by nature. Ask an LLM which companies are exposed to a given theme and it will return a reasonable list. What it will not offer is any assurance that the answer would be the same if asked again or if another person asks the same question, or that the result could be backtested.
Noonum takes a different approach, pairing large language models with a proprietary knowledge engine, what the company calls the "Web of World Commerce." This engine maps more than 23 million private companies, 150K+ public companies (153,388), and 1M+ securities (1,029,604) across 287 exchanges from 141 countries. The result is deterministic relative to time: the same query, run against the same historical date, returns the same answer, producing the kind of reproducibility an institutional benchmark requires. That combination of structured knowledge engine and LLM reasoning is what gives Noonum both interpretability and reproducibility, two qualities that matter most for benchmark construction.
The Signals Behind a Conviction Score
Two scoring concepts underpin this approach. Linguistic beta measures how closely a company's actual activities, including its disclosures, earnings calls, patents, and news, align with the specific language of a theme; it is best understood as a thematic relevance score grounded in evidence rather than intuition. Market buzz, by contrast, measures a company's relative share of voice in that market. The two are combined into a single conviction score through a weighted harmonic mean, which is conjunctive by design: the weaker of the two signals governs the result, so a company must be both well-evidenced and genuinely prominent to rank highly. Market capitalization never enters the calculation, as linguistic beta tilts negatively against size while buzz tilts positively, by comparable magnitudes, so balancing the two neutralizes size structurally rather than through an explicit correction. This score ranks how prominent and well-evidenced a company's exposure actually is, rather than how frequently it is discussed or how large the company happens to be.
How MerQube Uses Noonum
The workflow runs in four steps:
- Describe a strategy in plain English, using Noonum's prompt refinement tool to tighten a vague description into workable language.
- Noonum generates a theoretical basket of stocks tailored to the investment objective.
- Apply investability screens for market capitalization, exchange, or sector.
- Rank the resulting universe by conviction or its component signals.
- Export the result, whether as a spreadsheet, a factsheet, or simply the objective text.Because the model is deterministic, sharing the objective itself is often the simplest option, since it allows anyone to reconstruct the same portfolio independently.
We primarily use this to build preliminary baskets that let us test a new methodology, a rebalancing condition, or an inverse strategy before committing meaningful research time. It is also a useful way to confirm that a universe looks as expected, or to check whether part of a supply chain has been overlooked. Noonum's X-ray tool complements this by comparing one portfolio's thematic exposure against another's across sectors and industries.
Case Study: Noonum vs. a Traditional LLM
To see how Noonum differs from a general-purpose model in practice, I ran five investment objectives through both Noonum and a widely used LLM: AI power infrastructure, connections to President Trump, nuclear energy, the Strait of Hormuz, and agentic AI. I then compared each model's top 15 holdings by conviction. Overlap varied considerably by theme: nuclear produced the closest agreement, with 8 of 15 names shared; agentic AI and the two energy-adjacent themes fell in the middle, at roughly 5 to 7 of 15; and the Trump objective produced no overlap at all.
Much of this divergence traces back to source material. The LLM relies on patterns embedded in its training data rather than current filings, so it tends to favor companies receiving broader market attention rather than those with the strongest documented thematic exposure. Noonum's methodology, by contrast, combines evidence-based thematic relevance with a size-neutralized market-buzz signal to rank companies by documented exposure rather than by visibility. Across all 75 picks in the exercise, the LLM consistently converged on the same tier of company: predominantly large-cap, largely U.S.-domiciled, and obviously relevant at first glance. Noonum's baskets were less uniform, spanning micro-cap names, foreign listings, and financing entities that would not typically appear on a conventional list of top holdings. These selections reflect what a company's filings state rather than what is written about it elsewhere. The figures bear this out (See Chart 1). The pattern suggests the LLM's bias has less to do with company size than with familiarity, which tends to track media coverage more closely than market capitalization.
Market-Cap Distribution
Noonum vs. LLM, out of 75 picks across five objectives
| Metric | Noonum | LLM |
|---|---|---|
| Micro-cap holdings | 5 | 2 |
| Large-cap holdings | 28 | 33 |
| Mega-cap holdings | 14 | 17 |
Source: MerQube
This distinction is most evident at the extremes. For AI power infrastructure, Noonum interpreted the objective as an entire supply chain, encompassing compute, power, cooling, and the financing behind it, which is how a firm like Blackstone entered the basket. The LLM interpreted the objective more narrowly, as a bet on power generation alone: 13 of its 15 picks were conventional utilities, and it omitted Schneider Electric entirely, likely because the company's role in AI data center infrastructure is better recognized in Europe than in the U.S. media on which the LLM was trained. For nuclear energy, the LLM selected BHP Group, a mining conglomerate for which uranium represents a small share of revenue, suggesting difficulty assessing how much of a diversified company's business genuinely falls within the theme. For agentic AI, the LLM ranked Nvidia 11th, treating GPU infrastructure as downstream of the pipeline rather than the supply-chain component the objective explicitly requested.
The Trump objective proved the most instructive, since the two models did not merely disagree on ranking but on the underlying question itself. Noonum searched for policy beneficiaries and demonstrated proximity to the administration, including firms positioned to benefit from financial deregulation, while the LLM reached for the "Trump trade" shorthand already established in financial media: drilling, defense, crypto, and private prisons. One of the more striking results appeared here: Noonum surfaced Kura Sushi USA (KRUS), a conveyor-belt sushi chain that appears in Trump's own financial disclosures as a personal investment of $1-5 million this year.
It is an unusual pick, illustrating how Noonum's document-driven approach can surface relationships that might otherwise be overlooked; as with any institutional research process, we incorporate human review to determine whether those relationships are relevant to the specific investment objective. That same literal approach is what led Noonum to rank Saudi Aramco first for exposure to the Strait of Hormuz, a pick the LLM omitted entirely despite Aramco arguably being the company with the greatest physical exposure to the Strait during the 2026 disruptions.
Why This Matters for Index Construction
Noonum excels at one of the more time-intensive aspects of thematic investing: identifying genuine thematic exposure from financial disclosures and grounding broader market narratives in evidence drawn directly from source documents. The refinement tool adds similar value, turning an underdeveloped idea into something well-defined before meaningful research time is invested. Noonum has also indicated plans to incorporate transcripts of corporate meetings into its information universe, which should extend this advantage further.
Noonum is designed to integrate naturally into a broader investment research workflow, complementing fundamental analysis, portfolio construction, and index design. Constraints such as domicile or numeric thresholds are applied through its filtering framework rather than natural-language prompts, which keeps them explicit, reproducible, and easy to audit. We integrate Noonum-generated universes with our internal portfolio analytics and index construction platform, where we evaluate volatility overlays, methodology variations, and implementation details before launch. Because the X-ray tool is built to compare thematic exposure rather than historical performance, we typically pair it with our own analytics when evaluating strategies.
Noonum provides a faster, more evidence-grounded foundation for thematic research, reproducible enough to build upon and distinct enough from conventional LLM output to add meaningful insight. In practice, the two approaches are complementary: an LLM reflects how a theme is perceived in broader discourse, while Noonum grounds that theme in documented corporate activity. Understanding where those two views diverge is often the most valuable insight of all.