Where Will Investors Find Alpha in the Age of AI? | Alistair Smallwood, Head of Applied AI at Primer

Summary

The discussion examines how AI could reshape fundamental equity research and where human investors may retain an edge. Alistair Smallwood argues that AI can scale document analysis, connect otherwise separate company and industry signals, and generate a new “AI consensus” for investors to compare with market expectations. However, he sees AI primarily as a research aid—not a replacement for human judgment, portfolio decisions, or risk oversight.

Investing and market implications

• AI can systematize parts of mosaic research. Models can sift through company disclosures and other public information to identify relationships among customers, suppliers, peers, and macroeconomic developments. These connections may suggest changes in a company’s outlook before they appear in consensus forecasts.
• Faster access to common information may compress short-term alpha. As more investors and agents process the same information, recurring, near-term earnings signals may become less distinctive. Smallwood expects greater importance for understanding regime changes, turnarounds, leadership changes, and multi-quarter or multi-year business prospects.
• AI may itself create market mispricings. Models can display systematic biases. If many investors or autonomous agents respond similarly, their reactions could amplify short-term price swings, including overreactions, underreactions, flash-crash-like moves, or crowded positioning.
• A model’s forecast could become a new expectations benchmark. Investors could compare AI-generated estimates with analyst consensus to assess what is already priced in. A company that meets traditional consensus could still disappoint if AI-driven expectations are higher.
• Forecast accuracy is not the same as investment performance. Getting quarterly figures right does not necessarily identify the best stocks or predict the share-price reaction. Returns also depend on valuation, future expectations, business quality, management, market narratives, and the potential for multiple expansion or contraction.
• Longer investment horizons may regain importance. The speakers discuss a possible shift toward assets and strategies with longer duration and more tolerance for volatility as short-term signals become more crowded. Smallwood presents this as a cautious view, not a firm forecast.
• Alternative data requires careful interpretation. Data panels can have sample, geographic, or payment-method biases. Superficially strong signals may not represent a company’s actual customer base or demand.
• Human oversight remains important. Both speakers stress liquidity, position concentration, changing market conditions, and the possibility of second-order responses—for example, a company taking action to avoid a dividend cut and triggering a short squeeze.

AI research framework

Smallwood outlines several components for scaling mosaic theory and related research:

1. Obtain accurate primary data at scale. Reliably extract information from filings, announcements, presentations, and other company documents, with manageable latency and cost.
2. Give the system investment-specific skills and context. Explain what mosaic research is trying to find, supply industry knowledge, and teach the system to assess economic links, management commentary, inventories, and potential effects on other businesses. The aim is not simply to find high statistical correlation; some useful signals are qualitative.
3. Evaluate and challenge candidate signals. Use a scoring process, confirmation checks, and competing agents. One described test is to ask whether adding mosaic evidence changes the forecast for a company’s upcoming quarter compared with using information from that company alone.
4. Use outcomes to improve the process. Compare forecasts with subsequent results, investigate misses, and refine the system’s understanding of the business relationships involved.
5. Test skills systematically. Smallwood recommends A/B testing and systematic evaluation rather than choosing a prompt or research output based on a few examples or intuition.
6. Keep humans accountable. Treat AI outputs as inputs for further research, test them pragmatically, and retain human responsibility for investment decisions and risk.

Key figures and examples

• Smallwood said his investment career included eight years on the sell side and two years at a fund before he moved to Primer around 12 months before the interview.
• Early research systems used models with approximately 4,000-token context windows. He contrasted this with later models capable of processing roughly one million tokens, while cautioning that model handling of charts and tables was still not consistently reliable.
• The data infrastructure described covered primary documents for 15,000 companies globally over at least 15 years. One mosaic-research example involved searching 10,000–12,000 documents.
• In a test of 550 U.S. earnings reports, five language models characterized 85% as extremely positive, although 30% of the companies missed expectations.
• Primer’s primary agent was said to have been closer to actual company performance than consensus in 68% of cases in the previous reporting season. For the last 25 reports after improvements, the figure was 85%. The agent was closer to the midpoint of new guidance than consensus 70% of the time.
• A separate benchmark contained about 220 multi-step questions. The systems scored around 50%, while the speakers said about 75% of the answer keys contained contradictions.
• Smallwood cited a human-investor reference range of roughly 50–60% accuracy in earnings estimates and about 48% accuracy in predicting the stock-price move on reporting day.
• In a hypothetical example, Claude assigned about an 80% probability to Nike cutting its dividend within nine months. Smallwood explicitly said this was the model’s view, not his. The speakers discussed how crowded short positions could make such a signal risky, including the possibility of a short squeeze.
• Another illustration described alternative-data signals implying CTS Eventim revenue might double year over year, versus its typical growth of about 20%. The data panel had just 50,000 people and was distorted by Taylor Swift ticket sales. The speaker said the implied ticket volumes would have required roughly 24 concerts in venues twice the size of those booked in Germany. He said the fund opened a short position, while noting that outcomes were not always successful.
• Smallwood said Primer launched its customer product in April and was doubling monthly at the time of the discussion. He expected investment workflows to shift toward AI-system research within roughly 18 months, while broader industry adoption could take a few years.
• 80 basis points of alpha after slippage was given as an illustrative example of a testable recurring quantitative pattern, not as a claimed strategy result.
• One example of possible structural growth described a business unit generating 10% of profit that might grow 10-fold over five years if it fit a strong theme and had capable management.

Instruments, companies, sectors, and strategies mentioned

• Stocks and companies: Nvidia, Nike, CTS Eventim, Humana, UnitedHealth, American Express, and eHealth.
• Sectors and themes: Semiconductors and generative AI; healthcare and Medicare Advantage; consumer spending; British small- and mid-cap equities.
• Strategies and instruments: Pan-European long/short market-neutral, long-only, quantitative and factor investing, multi-manager funds, low-net and low-volatility strategies, short selling, dividend analysis, alternative data, earnings forecasts, and discounted cash-flow (DCF) valuation.
• No specific ETF, bond, commodity, or cryptocurrency recommendation was discussed.

Cautions and disclosures

• Smallwood described the Nike dividend-cut estimate as Claude’s opinion, not his own. Brett also said the Nike discussion was not investment advice.
• The speakers cautioned that AI forecasts, backtests, benchmarks, and alternative-data signals can be misleading or biased, and that accurate forecasts do not guarantee profitable trades.
• They emphasized the risks of crowded positions, volatility, short squeezes, model changes, and strategies failing during periods of poor performance.
• No general investment recommendation or portfolio allocation was provided.

Presenters and sources

Presenters: Brett Caughran and Alistair Smallwood, Head of Applied AI at Primer.

Sources and references mentioned: Primer’s research paper on scaling mosaic theory; Primer’s openstocks.com AI-forecasting resource; Rich Sutton’s essay The Bitter Lesson; and discussion or examples involving UBS, Credit Suisse, Bloomberg, Yipit, AQR, Bridgewater, Ken Griffin, and Claude.