How a top 20 US hedge fund turned universe screening into evidence-backed idea generation with Pascal.

A large share of the research effort inside any investment firm goes into maintaining existing coverage. That means the companies the firm already owns or follows, the earnings cycles, the model updates and the steady flow of disclosures that come with every position.

One of the top 20 hedge funds in the US came to Pascal with the other half of the job, which is discovery. For a fundamentally driven fund, the single biggest problem is screening the investment universe, and the universe is very large, something like 2,000 listed companies in one region alone and several thousand more once the fund looks across geographies. The fund's edge depends on finding the right names before the rest of the market does, and on being able to defend every one of those names in front of an investment committee.

The fund had a well-resourced research team and access to every major data terminal and screening tool on the market. The constraint was the shape of the questions the team actually wanted to ask.

The screens the fund wanted to run combine quantitative conditions with qualitative ones. This is so because the signal sits partly in financial metrics and partly inside conference calls, management commentary, news and regulatory filings. An analyst cannot possibly read everything, which means the quality of the ideas that make it onto the desk is constrained by the quality of the screening process that sits upstream.

A query no screener could answer

Consider a query like finding companies trading below a certain P/E that have announced capacity expansion, recently brought in professional management, and whose CEO sounds more optimistic on the latest earnings call than on the one before.

The first condition lives inside structured financial data (numerical information like revenue, earnings, and valuation ratios). The remaining conditions live almost entirely inside unstructured documents, and no traditional screener can answer that query. Running it manually means reading filings, transcripts and news one company at a time across two thousand names, which is why, before AI, the question mostly went unasked.

Three further problems

Screening was the headline problem, though the conversations with the fund's portfolio managers surfaced three more that sat right next to it.

Institutional memory

The fund had looked at a great many companies over the years and passed on a large number of them. The reasoning behind those decisions lived in investment memos, analyst notes and email threads, which meant that when a name resurfaced, the team often rebuilt its view from scratch without knowing what the fund had already concluded the last time around.

Thesis drift

Once a position was on the book, the analyst owning it had to keep watching for anything that weakened the original thesis. In a long/short book with a meaningful number of positions on both sides, a change in management tone, a quiet revision to guidance or a new risk factor in a filing could easily go unnoticed for weeks.

Trust and compliance

The fund had trialled general-purpose AI tools and found the answers hard to verify. For a regulated US fund, there was also a firm requirement that its research, its queries and its internal documents stay within its own environment, and that every query could be audited after the fact.

So the job Pascal ended up doing for the fund is idea generation, with the surrounding work of memory and monitoring built on the same foundation.

Pascal was deployed inside the fund's own cloud environment and connected to the fund's internal research repository, its licensed data feeds and the public domain. The platform reads public filings, news, broker research, conference call transcripts, podcasts and whatever other sources the fund cares to point it at. The investor, meanwhile, describes to the platform the kind of opportunity they are looking for. The system returns a small set of evidence-backed investment ideas, each tied directly to the filings, transcripts and news that led to their discovery. The portfolio manager still decides whether the idea is worth pursuing, but the work of searching for it has largely disappeared.

Breadth combined with depth

The way we think about the mechanism is breadth combined with depth. Breadth is the number of data sources being read at once, with filings sitting alongside news, broker research and management commentary. Depth is how far down the system reads within each source, sometimes to a single line buried inside a conference call transcript from five years ago.

Interestingly, both of these capabilities already existed in isolation before. An analyst could achieve extraordinary depth, but only across the thirty or forty companies they actively followed. Similarly, a traditional screener could achieve enormous breadth, but only across structured financial data. Holding both at once across an entire investment universe is the new capability the fund is betting on.

Evidence-backed ideas

Every investment idea the platform surfaces carries the chain of evidence that led to it. The analyst can open the exact filing paragraph, the conference call transcript, the news article or the announcement that contributed to the result and verify each claim for themselves.

This matters because investment teams judge ideas by whether they survive scrutiny. In our experience, most ideas will hold up under inspection, but there will be some that won't, and that is acceptable. What really matters here is making verification fast enough that an analyst can discard weak ones in minutes, without spending hours chasing them.

Screens that keep running

Once a portfolio manager has a screen they trust, it keeps running in the background. The fund's recurring screens are checked against every new filing, transcript and news item across the universe as it arrives. When a company newly meets the conditions, the relevant analyst receives the idea along with the evidence that triggered it.

In practice the screen behaves like a standing instruction to a junior analyst who reads everything overnight.

The fund's own memory

Pascal connected the fund's internal research, including investment memos, analyst notes and past committee discussions, into the same context layer as the public data. When a screen surfaces a company the fund has looked at before, the analyst sees what the team concluded at the time, who worked on it and why it passed. One can then ask the more useful question, which is whether anything has changed since that decision was made.

Thesis monitoring on the book

For every position on the book, the analyst's original thesis is captured as a set of conditions the system watches. New filings, transcripts and news are read against that thesis as they land, and the analyst is alerted when something weakens it, whether that is a change in management tone, a guidance revision or a new risk disclosure. The same evidence trail applies, so the analyst can see the exact line that triggered the alert and decide quickly whether it matters.

Built for a regulated environment

The deployment runs within the fund's own infrastructure, so research, queries and internal documents stay inside the environment the fund controls. Access follows the fund's existing permissions, meaning an analyst only sees the internal material they are already entitled to see. Every query and every output is logged, which gives the fund's compliance team a full audit trail of what was asked, what sources were used and what came back.