Proxy Contest Analysis: How AI Models Evaluate Activist Campaigns

Proxy contests — where an activist investor nominates their own slate of directors and campaigns against incumbent management — are some of the most consequential votes a retail investor can cast. They're also genuinely hard to evaluate. Until recently, the kind of analysis that helps you cut through the noise was only available to institutional investors through paid advisory firms like ISS and Glass Lewis. That's what Proxyanalyst is trying to fix.

Why proxy contests are tricky

Unlike a routine Say on Pay vote, where the question is pretty straightforward (did pay align with performance?), a proxy contest asks you to weigh two competing stories. The activist says the company is underperforming and that new board seats will unlock value. Management says the activist's plan is short-sighted and the current strategy just needs more time. Both sides file detailed materials with the SEC, each making their strongest case. Reading and making sense of both sides takes hours — and you need to know what to look for.

That's exactly the kind of problem large language models are well-suited for.

How the analysis works

When a new proxy contest shows up in SEC filings, Proxyanalyst runs the full filing text through four AI models: Claude (Anthropic), Grok (xAI), GPT-4 (OpenAI), and Gemini (Google). Each model gets the same prompt — review both sides, evaluate the merits of each case, and return an analysis with a confidence score. The models work independently with no sharing between them, so each assessment is that model's own read of the filings.

After all four come back, Claude does a consensus review. It looks at where the models agree, where they diverge, and why. The result is a five-part display on the site: four independent views plus a consensus that highlights the key points of alignment and disagreement.

What each model is looking at

Each model evaluates the same set of questions:

• How credible is the activist's thesis, and what's their track record?
• How has management actually performed, and is their strategic response convincing?
• Who are the board nominees on each side, and are they the right people?
• Is the stock underperformance attributable to the current board or to factors outside their control?
• Would activist board representation actually produce the changes being promised?

Each analysis ends with a position — Support Activist, Support Management, or Split Ballot — plus a confidence score. When models disagree, that disagreement is itself informative. A split often reflects genuine ambiguity in a contest where both sides have real arguments.

A couple of real examples

The first contest added to the site was Lululemon ($LULU). All five outputs — the four individual assessments and the consensus — landed in the same direction, with confidence scores clustered tightly together. When four models trained by different organizations, reading the same filings independently, reach the same conclusion, that's a meaningful signal.

Contrast that with Pacira BioSciences ($PCRX), where the models genuinely split. Claude leaned toward management conditionally, Grok favored the activist, and OpenAI and Gemini landed somewhere in between. The consensus reflected that honestly. For a retail investor, knowing that sophisticated AI models disagree is actually useful — it tells you this is a close call where your own read of the key variables matters.

The bottom line

I'm not trying to tell anyone how to vote. The goal is to make sure retail investors have access to the same quality of analysis that institutional investors get, so you can make your own informed decision. Whether you end up siding with the activist or sticking with management, you should be doing it with a clear picture of both sides' arguments — not just the summary card that shows up in the mail.

New contests get picked up automatically as SEC filings come in. You can follow along at the Proxy Contests page, where each active contest shows the full multi-model analysis, confidence scores, and once the vote has taken place, the actual outcome.