An exclusive look inside AI’s investment laboratory

Today, ChatGPT can screen a company’s annual report in minutes or give you a concise summary of a megatrend in seconds. But in most cases, the human is still choosing the investment opportunities and making the investment decisions.

A new generation of AI, however, is beginning to work differently.

Recently, we were given an exclusive opportunity to for a first-hand look at the technology behind GOAT’s strategy, Australia's first AI-powered international equity ETF.

The most important thing to understand is that GOAT is not asking a chatbot which stocks to buy, nor is it a fund that only invests in companies leveraged to the AI megatrend. What GOAT is doing is using generative reinforcement learning to evaluate approximately 16,000 investment signals across company fundamentals, technical indicators and macroeconomic data. Those signals are distilled into a single score for every company in the investment universe, with the 150 highest-scoring companies selected for the portfolio.

The output is the transparent, trackable index, GOAT tracks - the Akros Enhanced World ex Australia Index (GOAT Index).

The tour gave us a rare look inside the technology itself: the ideas it generates, the library of market data that helps it run and the process that turns thousands of possibilities into a handful of investment decisions.

Let’s start at the very beginning

The Akros engineers describe the starting point as an ‘idea factory’.

Every week, the platform generates thousands of potential strategies. Some explore familiar ideas, such as whether highly profitable companies tend to outperform or whether recent price trends matter. Others examine different combinations of company fundamentals, technical signals and macroeconomic data, searching far more broadly than a traditional research process could realistically allow.

Unlike a conventional analyst, the platform isn't searching for confirmation nor is it generating lists of companies to buy. It's searching for possibilities, asking “what if” at a scale that no human research team could match.

The library is open

Generating strategy ideas is only the first step. The real work begins when those ideas enter the platform's historical market library.

The library draws together more than terabytes of historical market data – equivalent to holding tens of millions of books’ worth of information. The data includes company financial statements, market prices, macroeconomic indicators and technical signals, creating a record of financial history stretching back more than a century.

Building the library was one of the biggest engineering challenges behind the platform and the index. The data had to be reconstructed exactly as it appeared at each point in time, preserving company financial statements before they were later corrected, while ensuring businesses that went bankrupt or were acquired remained part of the historical record rather than disappearing from it.

If the platform is testing an investment idea using data from 2015, it only sees the information that was available in 2015. The model doesn't know what happened next, meaning the benefit of hindsight is removed. This helps prevent the platform from developing strategies that only look successful because they already know the answer. Instead, every investment idea must prove it could have worked in real time, using only the information investors had back then.

The assembly line

Every candidate strategy then enters what the Akros engineers describe as an assembly line. The assembly line is divided into four stages: generate, score, learn and validate.

1. Generate: The platform creates thousands of candidate strategies, each representing a different way of selecting and weighting stocks.
2. Score: Every strategy is run through the historical market library and given a report card. The platform measures not only returns, but how consistently they were achieved and whether the strategy held up across different market conditions.
3. Learn: The engine studies those report cards and looks for patterns. It uses what worked, and what didn't, to guide where it searches next, with each cycle informing the next.
4. Validate: The strongest candidates are tested using data they have never seen before. Only the ideas that continue to hold up make it through.

Why the engine distrusts easy answers

One of the biggest misconceptions about AI-powered investing is that the technology is making investment decisions in real time.

But AI is fast at research, not investing. This is why nothing that is revealed during research automatically touches real money.

If a strategy survives, the AI's job is done. Humans at Akros will then review the validated strategy before it is published as a transparent set of rules describing which stocks to hold, how much of each to own and when those holdings should change. The portfolio is then constructed by following those rules.

The bigger challenge, according to the Akros engineers, is stopping the AI from fooling itself. To prevent this, the best candidate strategies are tested against data they have never seen before, to see how they would have performed in different market environments. This way, the machine eliminates the ideas that only appeared to work because they were lucky.

The biggest surprises from the construction process

Two things from the platform's development surprised even the engineers behind it.

Firstly, without ever being taught decades of academic finance research, the engine independently rediscovered many of the investment ideas already familiar to investors. Concepts such as quality, momentum and value emerged naturally from its own search process. It was a reassuring result because it meant that the machine had independently arrived at many of the same conclusions as generations of investment researchers.

Alongside those familiar ideas, the engine also began identifying candidate strategies that don't appear in the published research at all. For the engineers, it reinforced something they had suspected from the beginning: human research is limited by the ideas humans think to test. An AI system has no such limitation.

From strategy to portfolio

At the end of each cycle, the engine produces a constituent list describing which companies should be held or removed, and how much weight each should carry. Every holding is accompanied by a paper trail explaining why it belongs there, allowing every decision to be traced back to a transparent set of rules rather than simply accepting that "the AI said so". The lessons from each cycle then feed back into the engine, informing where it searches next.

This is the process underpinning the VanEck Dynamic International Equity ETF (GOAT), Australia's first AI-powered international equity ETF. It offers a glimpse of how investment research is evolving, and where it may be headed next.

Key risks

An investment in the ETF carries risks associated with: ASX trading time differences, financial markets generally, individual company management, industry sectors, foreign currency, country or sector concentration, political, regulatory and tax risks, fund operations and tracking an index. See the PDS and TMD for more details.

GOAT is likely to be appropriate for a consumer who is seeking capital growth, is intending to use the product as a major, core, minor or satellite allocation within a portfolio, has an investment timeframe of at least 5 years, and has a high risk/return profile.