Primer says its LLM-based system can execute mosaic theory at scale, combining weak public signals into conviction about a company across filings for 15,000 companies.
Alistair Smallwood of Primer argues short-term repeatable alpha is compressing, so investors must lengthen horizons toward regime shifts, turnarounds and management changes where AI is underrepresented.
Primer says its agent beat consensus on reported numbers 85% of the time across 25 reports, yet Brett Caughran notes human investors manage only 48% accuracy on print-day price moves.
AI Compresses Short-Term Alpha and Pushes Investors Toward Longer Horizons
AI is turning fundamental equity research into mosaic analysis at scale, with models combining many weak public signals and forecasting earnings better than consensus. That compresses repeatable short-term edge, so human alpha moves to longer horizons, regime shifts and management judgement. Models still need rigorous evaluation, and accountability for decisions stays with people.
Takeaways
LLMs can now execute mosaic theory at scale. Primer's paper shows models combining weak public signals into conviction about a company. The system uses filings for 15,000 companies over at least fifteen years, skills files that encode analyst reasoning, and an evaluation engine testing whether each signal changes next quarter's growth forecast.
Build on state of the art harnesses. Citing Rich Sutton's Bitter Lesson, Smallwood says compute wins over time. Primer's custom work would be overtaken within three to five months. Million token windows have replaced the 4,000 token limits that once forced a custom filings pipeline.
Short-term repeatable alpha is compressing. Smallwood says investors must lengthen horizons toward regime shifts, turnarounds and management changes, where AI is underrepresented. Caughran adds that fundamental investing has been gutted over fifteen years by factors and the move to multi-managers, which leaves distortions to exploit.
AI earnings forecasts often beat consensus. Primer's agent was closer to reported numbers than consensus 68 percent of the time last season, and 85 percent across 25 reports after improvements. It was closer to guidance midpoints 70 percent of the time. Human analysts often manage high 50s to low 60s.
Model outputs need careful evaluation. Primer ran 550 earnings calls through five foundation models. 85 percent were judged super positive although 30 percent missed earnings. A benchmark of 220 questions had contradictory answer keys in about 75 percent of cases.
Systems should generate leads, not trades. Smallwood wants lines of inquiry rather than trade signals, since management meetings and channel checks carry context models miss. Accountability stays with the human owner. A credit card panel of 50,000 Europeans, skewed by Taylor Swift tickets, once implied CTS Eventim revenue would double.
Demand is rising but skills are scarce. Primer launched its customer product in April and has doubled every month since. Large funds want sovereignty over their research systems. Smallwood sees a skills gap between technology teams and fund managers, and says AI skill takes about ten thousand hours.
What comes next
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Smallwood expects many fundamental teams to look more like quantitative teams within five to ten years.
He favours self learning loops as the way skills files stay durable.
Caughran suggests AI generated consensus could become a new data stream, with risks of crowding.
Caughran argues the vertical AI layer serving finance will be as large as the coding layer.
Both expect finance to take a few years to transition, and a part two is planned.