AI is becoming increasingly sophisticated at identifying patterns and generating investment signals from disparate sources of financial and market data. But for systematic investors, those signals still have to be converted into something much more concrete: a portfolio that determines which assets to hold, in what proportions, and with what degree of conviction.
Doing that becomes more complicated when an investment strategy draws simultaneously on different types of information. Fundamental, macroeconomic, technical and unstructured data can each point in different directions, while their relevance can change as market conditions change. Building an investable strategy therefore requires not only generating signals, but determining how they should interact and how much confidence should be placed in them.This is the problem that QuantumStreet AI has been working on since its origins as a research project at UC Berkeley. Cofounder and President Art Amador says one of the observations behind that work was that different investment disciplines tended to examine their own domains independently, despite the relationships between the information they were analysing.
“We were one of the first to really think about how you can combine all these different types of data because they are connected, and you can find interesting patterns across these different datasets,” he tells Market & Alt Data Insight.
Bringing the signals together
QuantumStreet was founded by Amador alongside colleagues with backgrounds in institutional investment management and AI and machine learning. Early research focused on combining fundamental, technical and macro data with information extracted from news using natural language processing (NLP).
The project attracted the attention of IBM, whose Watson technology QuantumStreet was using for its NLP work, and subsequently developed into a commercial business. Interest came in particular from quantitative investment strategy (QIS) teams at global banks looking to incorporate AI into systematic index strategies.
Amador describes the resulting investment architecture in three broad layers. The first generates signals from different categories of information. The second integrates those signals into an expected-return forecast. The third uses the forecast, together with the model’s confidence in it, as an input into portfolio construction.
QuantumStreet now works with QIS teams at banks including HSBC, Deutsche Bank and BNP Paribas, with more than $8 billion invested in indices it has developed, according to Amador.
Knowing what to leave out
The process also puts a different perspective on the value of large datasets. QuantumStreet’s early work with IBM Watson involved processing around one million news articles a day. Amador says the experience demonstrated that the objective shouldn’t simply be to maximise the amount of information fed into a model.
“It’s not necessarily about how much information you can process. It’s about how you can transform that information into a signal that ultimately translates into an investment strategy. You don’t necessarily need to process a million articles; you need to process the right subset of that information.”
The question then becomes which information belongs in that subset. For unstructured data in particular, relevance is only part of the equation. The credibility and persistence of the source can also affect the weight that should be attached to the information.
“We’ve thought a lot about how you ensure trust in data,” says Amador. “We spent a lot of time developing trust scores for unstructured data sources. You can think about it in a similar way to how a search engine works: we care about which data sources are effectively pointing towards the source we’re using.”
QuantumStreet draws on purchased structured and unstructured datasets alongside information gathered from the web, covering global equities, fixed income and most major commodity markets. Assessing the provenance of those inputs becomes another component of the investment methodology rather than a separate data-management exercise.
Conviction as an input
One of the more interesting elements of the approach is the use of confidence alongside the forecast itself. A model might generate an expected return for a security, but that forecast doesn’t necessarily deserve the same weight on every occasion. Incorporating confidence provides another input into determining how strongly the portfolio should act on what the model is predicting.
This becomes particularly relevant when multiple models are being combined. Macro conditions may favour one investment factor while fundamental or technical signals suggest another. Rather than relying on a single analytical lens, the portfolio construction process can take account of what each model is saying and how strongly it is saying it. It also creates a requirement to understand where a forecast came from.
Explaining the portfolio
For institutional investors, that becomes important once AI-generated forecasts are responsible for determining actual exposures. QuantumStreet allows users to decompose a forecast to see the contribution from macro, fundamental, technical and news signals and drill down further into individual factors. Clients can also examine forecast confidence and model accuracy.
The level of visibility expected by institutional clients has changed considerably since the company’s early years, says Amador. “In the beginning, we erred on the side of not being as transparent to protect IP, so it felt much more like a black box. Now, in order to win an index mandate from a risk committee or investment committee, you need to provide a certain level of transparency.”
That creates a balancing act for developers of proprietary investment models. Investment firms don’t necessarily need access to every element of the underlying intellectual property, but they do need enough information to understand why capital is being allocated in a particular way and to subject the methodology to appropriate oversight.
New technology, same investment problem
The technologies available to tackle these problems continue to evolve. QuantumStreet’s early work was heavily focused on NLP; the arrival of large language models has expanded the tools available for extracting information from unstructured sources. The company is also exploring potential applications of quantum computing in areas including portfolio optimisation and time-series prediction.
Those developments may change the sophistication and speed of investment models, but they don’t remove the fundamental task of portfolio construction. For systematic investors, the value of AI ultimately depends on whether its outputs can be incorporated into a coherent investment methodology: bringing together different signals, deciding how much confidence to place in them and converting those views into positions that can be understood and governed.
Subscribe to our newsletter
