Machine learning from a “Universe” of signals: The role of feature engineering

GAICIL Insights

Throw every number at a powerful algorithm and let it find the money.¹ That is the big-data dream, and this paper is its reality check. Machine learning from a "Universe" of signals: The role of feature engineering, by Bin Li, Alberto G. Rossi, Xuemin (Sterling) Yan and Lingling Zheng, tests machine learning in asset pricing, an enthusiastic adopter of the technology. The authors build real-time machine learning strategies using a broad universe of fundamental signals, avoiding the curated predictor sets used in much prior research.

Here is the uncomfortable test for anyone buying an AI strategy: ask what the model was told to look at and who decided. If the answer is everything, chosen by nobody, expect less than the backtest promises.

The results humble the hype without dismissing it. Out-of-sample performance is economically meaningful and statistically significant. It is also considerably weaker than performance documented by studies using curated signals; long-short returns and Sharpe ratios are considerably lower. The gap is driven by differences in input data, with algorithm implementation playing no comparable role. Strikingly, strategies based on a simple recursive ranking of each signal's past performance achieve substantially better out-of-sample results. Analyses using past-return signals reach qualitatively similar conclusions. A spreadsheet-simple ranking beats the universe-fed machine.

The interpretation centers on structure. Feeding the universe of predictors to machine learning produces performance inferior to the same methods with curated predictors and worse than a simple ranking that incorporates feature engineering. Imposing an appropriate structure, or inductive bias, is essential for machine learning to perform well in cross-sectional prediction. The authors cite the maxim that feature engineering is the key in machine learning applications. Data without design is noise with a large budget.

The finding speaks to a central question in artificial intelligence: how far can data and computation substitute for domain knowledge? In this domain, not fully. Powerful algorithms given vast raw inputs perform less well than simpler methods guided by thoughtful structure. Human judgment about which features matter, and how to build them, remains a key ingredient. Expertise returns as an input, the one the scaling story tends to leave out.

The real-time design keeps everyone honest. Studies that select signals with knowledge of their historical performance may overstate what investors could have achieved. Using signals available in real time gives a more realistic estimate of machine learning's value. The bottom line is constructive: machine learning helps real-time investors, and feature engineering plays a vital role in substantially elevating those benefits. Some published alpha may be hindsight in disguise.

The lesson travels far beyond finance. From medicine to economic forecasting, the temptation is to supply algorithms with every available variable and let them find the signal. This paper suggests the choice and construction of inputs can matter as much as the algorithm. Inductive biases that encode knowledge about the problem help algorithms generalize. Organizations deploying AI should invest in domain expertise and feature design alongside computational capacity. The analyst who knows which ratios matter is still worth hiring.

Market efficiency enters too. If machine learning applied to broad signal sets yields weaker returns than earlier studies suggested, opportunities for sophisticated investors may be more limited than headline results implied. Evidence like this helps calibrate expectations about how AI will change financial markets, deflating both utopian and dystopian forecasts.

For regulators, the study cautions against overstating AI's power in finance. Claims about algorithmic performance should face real-time, out-of-sample tests of the kind presented here. Transparency about inputs and methodology helps distinguish genuine predictive power from artifacts of data selection. Marketing materials for AI funds would look different under that standard.

Read as a whole, AI works best combined with human insight. Machine learning extracts value from data, and the value grows when algorithms receive the right structure. In finance and beyond, the most effective AI systems will combine computational power with carefully engineered features grounded in knowledge of the domain. This paper supplies rigorous evidence for that principle in one of the most data-rich settings available.