Understanding the present before it is priced: where an investment agent’s edge comes from

Prices pay whoever processes information

The argument starts in 1980. Sanford Grossman and Joseph Stiglitz showed that prices cannot reflect all available information: if they did, no one would be paid for the cost of gathering it, no one would gather it, and prices would stop reflecting it. In equilibrium prices reflect what informed traders know only in part, and the gap is what compensates them [1].

So the useful question about an investment agent is not whether markets can be predicted. Every position is an implicit forecast, and an agent that writes a probability before each outcome is forecasting. The question is where prices are slow to absorb what is published, and who can afford to process what they are slow on.

Attention, not intelligence, is what runs out

Four lines of research locate the slowness, and each locates it in attention. When more firms announce earnings on the same day, the immediate price reaction to each surprise is weaker and the drift after it stronger [2]. Announcements made on a Friday, when investors are less attentive, draw a 15% weaker immediate response and a 70% stronger delayed one [3].

Links between companies are absorbed late too. The relation between a customer and its supplier is public, yet news about the customer reaches the supplier’s price with a lag, so the customer’s returns predict the supplier’s [4].

And so are quiet changes. Companies that change the language of their annual and quarterly reports do worse afterwards, and changes to the annual report predict future earnings, profitability, news and even bankruptcies. Yet the price does not move when the report is filed: the returns accrue later, when the information arrives through news, events or earnings, and the authors read this as investors not attending to the changes [5].

None of these needs a better model of the market. Each needs someone to read what was published, on the day, and connect it to what it touches. Andrew Ang and his co-authors put the institutional version plainly: “The most binding constraint in institutional asset management is not data availability or model sophistication, but the finite bandwidth of human decision-makers” [15]. Norway’s sovereign wealth fund now screens each new company in its portfolio with a language model within 24 hours of investing, and says that in several cases it sold before the wider market reacted [16].

Breadth outweighs accuracy

Richard Grinold’s fundamental law of active management turns the argument into arithmetic. A strategy’s value depends on how well it plays, its skill, and how often it plays, its breadth: the ratio rises with the square of skill and directly with breadth [6]. In the form now standard, the information ratio is roughly skill times the square root of breadth. Skill is hard to raise; breadth is what a machine that reads everything can raise.

The table below is the formula’s arithmetic, not anyone’s result. Half the accuracy over sixteen times the independent judgements doubles the ratio.

Two caveats keep the arithmetic honest. Breadth counts independent judgements: a hundred readings of one theme are one bet in disguise. And the law in its full form adds a transfer coefficient, the degree to which the forecasts reach the portfolio’s weights: it is 1.0 without constraints, and the authors find it between 0.3 and 0.8 in practice [7]. An agent’s discipline lives there: sizing each judgement by its edge and its risk, inside limits written in advance, rather than letting the best idea of the day take the book.

The speed that counts is the speed of understanding

Speed to the order is a race already decided. High-frequency traders trade in the direction of permanent price changes and against transitory errors, their order flow predicts prices over horizons of seconds, and it is correlated with public information such as macroeconomic announcements [8]. An agent that reads a headline and sends an order competes there, and loses.

Reading headlines with a language model is not a lasting edge either. Alejandro Lopez-Lira and Yuehua Tang found that models’ readings of news predict the drift that follows, most for small companies and bad news, and that the returns of such strategies decline as language models spread [9]. In prediction markets, mostly on sports, some models beat the market days before a question resolves, but as resolution nears the market takes in breaking news faster than any of them [17].

What remains is the middle ground of the research above: information too deep for a fast machine and too broad for a team. A note to the accounts, a changed sentence in a risk section, a supplier’s filing that moves a customer’s thesis. There the market takes days and weeks, and an agent’s minutes are the difference. The speed that counts is the time from publication to a verified judgement linked to every position it touches.

Speed without verification is a liability

A fast judgement on a wrong figure is worse than a slow one, and language models bring their own ways to be wrong. They remember the future: a measure of look-ahead in their forecasts is clearly positive inside their training period and collapses to near zero after its cutoff, and the forecasts are most accurate exactly where it is high [10].

They also agree with each other. At the start of a war, T. Rowe Price saw all the language models recommend the same strategy, to buy the dip [11]. An agent that reads like everyone else trades like everyone else, and speed only gets it into the crowd first. And a review on the CFA Institute’s blog lists what breaks in multi-agent systems in production, error cascades and groupthink among them, and warns that such systems are least reliable when markets are most stressed [12].

So the rigour is what makes the speed usable. Each figure comes from a document, with its location, or from a tool; each claim is checked against its source; what cannot be established is escalated, and the agent abstains. Most of what is published changes no thesis, and the right action is usually none.

What it means for whoever builds an agent

Start from the portfolio, not from the forecast. The profession’s process runs from planning to execution to feedback, and planning is written down in an investment policy statement: the return objective and risk tolerance over a horizon, and the constraints [13]. A system of many agents widens the search over portfolios without settling what it is searching for [14], so the objective has to come from the institution, as a charter the agent cannot change. One published design of a portfolio run by agents puts its whole pipeline under the investment policy statement, written by humans [15].

Where we stand

Quantaup is a technology company in its research and development phase. We build a system of investment agents that works to each portfolio’s charter, use it internally today, and are opening a research partnership to a limited number of licensed institutions. This report states no figure about our own work: what would settle its argument is a record of sealed judgements on an institution’s own questions, not a report.