The audio archive that is bigger than Spotify’s library 

Quantitative investing is often pictured as a black box that runs itself. Sadhvi Gupta is Executive Director, Quantitative Investment Strategies at Goldman Sachs Asset Management and client portfolio manager of the Yarra Global Small Companies Fund. She describes a process where people design every signal. Technology supplies the speed and scale.

One dataset illustrates this blend of human skill and assessment empowered through the processing power. Gupta’s team holds recordings of every earnings call from every US company over the past 10 years. “At the time the model was trained, the data set was bigger than Spotify’s global library,” she says.

Most analyst teams would struggle to work through that archive manually. The technology compares the tone, the speed and the awkward pauses of a chief executive or chief financial officer against previous quarters. Gupta says that comparison gives a better read on whether a stock goes on to do better or worse.

Alongside the raw processing power, there is human psychology in how to interpret this rich data set. The team studied how FBI investigators identify people who are being circumspect with the truth, and trained the technology to listen for the same cues.

Four pillars of quantitative investing

The Portfolio stays close to the benchmark at a sector and country level, with deviations of no more than 2 or 3 per cent.

Most of the risk budget goes into bottom-up stock selection, organised around four pillars. Quality and price are the more fundamental pair and make up about 40 per cent of decision making. Sentiment, and the themes and trends driving markets, account for the other 60 per cent.

Those pillars are expressed through around 250 signals. Some use traditional data and some use alternative data. Some use AI and some do not. They span trend-following and contrarian ideas across short and long horizons.

The quality pillar draws on data such as credit card transactions, website traffic and geolocation. It gives an early sense of where people spend their time and money.

One recent signal tracks app downloads and daily users. If many people download the app of a company selling CCTV cameras, Gupta explains, it probably means they have just installed a camera. They need to see the footage.

Valuation goes well beyond price to earnings or price to book. The measure of intrinsic value may be rental occupancy for a real estate company, the patents it owns for a biotech, or the quality of the loan book for a bank.

Sentiment asks who is bidding up a stock. It looks at whether the buyer is an informed institution, a retail investor or passive money following an index inclusion.

Large language models read sell side research, earnings call transcripts, patent filings and regulatory reports.

Themes and trends, the team’s more sophisticated version of momentum, looks for stocks connected to a market driver that have not yet seen the price appreciation. Gupta says it has been the pillar contributing most to performance over the long term.

The human layer

Every alpha factor is defined by a person, down to how raw data is cleaned and where its biases lie.

“It’s humans that really guide this technology on what to do, and then the technology does it much faster than we could.”

Risk management is the second area of oversight. A risk model cannot read geopolitics, so portfolio managers test how exposed the portfolio is to shocks such as a 10 per cent rise in oil prices.

“It’s not just are you overweight or underweight energy names, but there are so many stocks in so many different industries that would be impacted by an increase or a decrease in oil price,” Gupta says.

The skills behind this work have shifted. A decade or more ago the team primarily hired economics, finance and mathematics graduates who could also code, while recruits now increasingly come from AI, machine learning and computer engineering.

Every factor must still rest on economic intuition and behavioural finance rather than on a back test alone. “It really has to be something that we understand and that we think is a fundamental driver of stock returns,” she says.

Research and portfolio management work in one combined team, so the people who build a signal answer for how it performs, and the firm pays them accordingly. The team formally reviews the model around every six months, making incremental improvements each year.

Gupta notes the firm has invested quantitatively since 1989, starting with a handful of signals built on financial data that analysts could run in Excel.

Crowding

With more managers now describing themselves as quantitative, crowding is the question clients raise most often. The team tests for it from the outset through alpha decay, measuring how long traders can act on a signal’s view before the market prices it in.

“If we start to see that window shorten, then that to us means that other people are also betting on a similar idea,” Gupta explains. Factors that become too commonplace are reduced or removed from the model.

Gupta acknowledges crowding as one of the bigger risks facing the quantitative community. In August 2007, a group of quantitative managers holding similar positions sold at the same time, and losses spread quickly across strategies that had looked diversified. Measuring decay helps, but it depends on spotting the crowd before it moves.

“Quant” is becoming a more common label. Gupta suggests the question worth asking of any manager is whether the technology and data really drive differentiated returns.

For advisers and their clients weighing quantitative investing, that answer may prove more revealing than any label. It may also prove more revealing than any list of signals.