Traditional financial and market information now sits alongside news, sentiment, analyst estimates and a growing universe of alternative datasets, creating an increasingly rich information environment for investment firms.
But access to more data doesn’t necessarily translate into better investment decisions. Firms still have to establish which information has genuine predictive value, turn disparate datasets into usable investment signals, and give portfolio managers enough transparency to understand why those signals matter.
For Hassan Salamony, Co-Founder and COO of 3AI, the starting point is distinguishing persistent signals from historical relationships that may disappear as market conditions change.
“The biggest challenge is establishing that a dataset has persistent significance for future alpha, rather than simply looking interesting historically,” he tells Market & AltData Insight. “In markets, it is relatively easy to find relationships that work for a particular period, regime or subset of stocks. What matters is whether the information remains predictive through time, across a broad equity universe and through different market environments.”
From raw data to investment signals
Making that distinction requires more than testing a dataset against historical returns. Point-in-time integrity is critical to avoiding look-ahead and survivorship bias, while long histories are needed to establish whether relationships persist across different market regimes.
There is also considerable work involved before a dataset can be used in an investment model. Providers differ in their histories, frequencies, definitions, coverage and release timings. Data needs to be cleaned and normalised, missing observations addressed and information mapped consistently to individual securities.3AI’s approach incorporates point-in-time treatment, currency and temporal normalisation, feature engineering and imputation before information reaches its forecasting models. Its training framework also uses walk-forward validation on unseen future periods and includes failed and delisted companies.
But Salamony argues that the bigger challenge comes when different sources of information are brought together. “A dataset rarely has a fixed relationship with returns. Its usefulness can depend on the company, sector, market regime and the other information you observe alongside it,” he says. “Financial relationships are often non-linear and conditional, so the opportunity is not simply to collect more datasets but to engineer meaningful features from them and understand how those features interact.”
Separating signal from noise
This puts feature engineering at the heart of the process. 3AI engineers hundreds of company-level factors from financial, market, analyst and other data, covering areas such as earnings quality, valuation, financial change, shareholder treatment, sentiment, momentum and credit risk.
A feature may capture an economically meaningful characteristic of a company without necessarily being predictive in isolation. Its value may emerge only in combination with other information or under particular conditions.
“Raw datasets are rarely the signal themselves,” says Salamony. “The real work is in designing and combining features that extract the economically relevant information in a form that is stable and comparable across a very broad stock universe.”
Large-scale empirical testing and out-of-sample validation can then be used to identify which combinations retain predictive value rather than simply fitting a particular historical period.
Making AI explainable
As AI-generated signals move further into institutional investment processes, predictive performance alone isn’t enough. Portfolio managers, risk teams and regulators increasingly need to understand how a model arrived at its conclusions.
“Explainability is critical for trust and credibility,” notes Salamony. “Investment professionals are not going to rely on a forecast simply because a model says a stock should outperform. They need to understand what is driving that view, whether the reasoning is economically intuitive and whether it is consistent with the information they themselves are seeing.”
3AI provides alpha attribution down to the engineered features and underlying data observations contributing to a forecast, allowing users to examine factors such as valuation, sentiment, financial change, technical indicators and business-cycle effects. Its methodology also incorporates SHAP-based explainability and traceability through the forecasting process.
That transparency can serve two purposes. A model may systematically combine information that an investment manager already follows, providing additional confidence in its output. Alternatively, it may identify information or relationships outside the manager’s existing process.
Explainability is also becoming a product-governance issue. Salamony says regulators and oversight bodies reviewing AI-powered investment products increasingly want assurance that signals can be traced back through the model to the economic factors and observations behind them.
From distributing data to distributing intelligence
The longer-term impact of AI could extend beyond individual investment processes to reshape the alternative data industry’s value chain.
Today, investment firms often duplicate much of the same work: acquiring datasets, cleaning them, mapping them to securities, engineering features and then determining whether they contain persistent predictive information. Salamony expects more of that foundational analysis to be performed centrally by data and intelligence providers. “Over the longer term, I think AI will change the alternative-data industry from primarily a data-distribution model into much more of an intelligence-distribution model,” he says.
For systematic investors, that could mean consuming validated, investment-ready features and signals without repeatedly rebuilding the underlying data-engineering and initial research stack. For discretionary managers, AI could distil much larger quantities of information into explainable stock-level insights, leaving managers to focus more of their time on applying investment judgement.
That doesn’t imply replacing proprietary investment processes with a standardised answer. Rather, firms could begin their own research further up the information chain, applying their expertise and investment philosophy to data that has already undergone much of the costly preparatory analysis.
As Salamony puts it: “Ultimately, I think the value chain moves up a level: from selling data, to selling intelligence about what in that data actually matters.”
Subscribe to our newsletter
