Dump the CEO, Thirty Years Later

Short version

• Robin Hanson's "dump the CEO" market never got built. We investigate whether AI forecasting can unlock it.
• We use EPS three years out as the indicator and ask whether dumping the CEO raises it versus the status quo. Then we compare to the share price version.
• The differentials between the two decisions are consistently small relative to the uncertainty, and stable between runs, which is necessary for decision forecasts to be useful.
• We get different results from our earlier Fire the CEO piece and explore how formulating the question matters.

Hanson's question

Three decades ago, Robin Hanson proposed a conditional prediction market to measure the impact of firing the CEO: one market on the share price if the board keeps the chief executive, one if it dumps them, and a rule that the board acts when the dump price sits far enough above the keep price. The goal was better governance. It does not exist partly because of challenges with liquidity, incentives, and causality.

A key promise of accurate, even superhuman, AI forecasting is that the idea is now tractable. Just ask an AI forecaster.

At FutureSearch we recently introduced decision forecasts for situations like the one Hanson envisions. As anyone who has thought hard about conditional forecasts knows, Hanson included, a naive conditional market can give a very different answer from the causal impact participants have in mind. We designed decision forecasts to reflect causal beliefs, at the cost of constraints that narrow the allowed use cases.

Decision forecasts work well for grantmaking because the constraints are satisfied: (1) the grantmaker fully controls the decision, (2) the application supplies the background context, (3) the decision is made soon, and (4) it is irreversible, since you cannot easily take the money back. Under those conditions the forecast is effectively causal. An accurate AI forecaster can truly help guide the decision.

The bigger dream is to extend this to a broader class of deciders, like a board deciding whether to dump the CEO.

Complications for a board

This sounds straightforward at first. You sit on the board, and you ask an AI forecaster whether dumping the CEO would raise the stock price.

One challenge is that the decision is largely reversible. Keep the CEO at this meeting and you can dump them at the next one, or sooner. A forecaster convinced that dumping the CEO will have a big causal impact might forecast little difference between branches, if they expect the CEO to be gone by the following meeting anyway. While this is a challenge, we don't observe it happening in our examples below.

Another is that a board has private information we do not. We can forecast from the outside, but constraint (2) is only partly satisfied. Our product lets a board supply that context and get a better forecast. We just cannot do that from the outside. A real board can.

A third is that a board controls only some things. It can fire a CEO, start a search, and make an offer. It cannot control whether the offer is accepted, or when. It is reasonable to think forecasting a specific appointment is more useful, and that is the approach taken in Fire the CEO, which priced named replacements at ten companies.

A further complication is what the market already thinks or knows. If it expected the board to dump the CEO, an announcement might barely move the stock. If the change is a surprise, the market might assume the board knows something bad it cannot disclose, and the stock might fall. Or it might read the change as good governance.

Hanson asked about stock price. Another formulation is to ask about business fundamentals after a sufficient period, perhaps three years. That is what we do first below, using GAAP earnings per share three years out as a proxy for fundamental value, under both board decisions. Then we compare to the share price version.

What we ran

We used the Fire the CEO piece as the foundation, with a few changes. We dropped Meta, because Zuckerberg holds a voting majority and no board can dump him, and Apple, which had just changed CEOs. We used a keep or dump decision, close to Hanson's, rather than a ladder of named candidates. And we forecast EPS three years out instead of share value one year out.

Every company got the same two alternatives, word for word:

The Board of Directors decides to retain the CEO for now

The Board of Directors decides to replace the CEO

Retain means the board keeps the CEO for now. Replace means it decides now to remove the CEO within three months and searches for a successor meanwhile, with no particular successor assumed. The outcome is the sum of GAAP diluted EPS over the four most recent fiscal quarters reported by August 31, 2029. The details are in a footnote.¹

All eight companies went into one decision forecast at high effort. Each run took about twenty minutes and cost about thirty dollars.

Forecasting results

In every case the forecaster expects dumping the CEO to lower EPS three years out. The differentials are small relative to the confidence interval, every gap but Tesla's at 5 to 12 percent of the retain branch's p10 to p90 band, so we read the forecasts as saying there is no huge change either way, which matches our priors. Tesla is the exception: no earnings effect, but a narrower distribution under replacement, because a professional successor would cut the capital expenditure binge and forfeit the CEO award's stock compensation, raising the floor, while losing the autonomy upside, lowering the ceiling. Replace bands are wider at the bottom and slightly higher at the top almost everywhere, the shape you would expect if a transition mostly adds downside with a small chance of a better operator.

Is the differential stable?

As anyone who uses LLMs knows, you get a slightly different answer every time you ask the same question. So when you care about the differential between two cases, the same forecaster should answer both at the same time, from one view of the situation. That is how a decision forecast works: both branches are researched jointly inside one row. To test how much this buys, we ran the identical eight rows again.

The absolute levels move by up to half a dollar. The differential moves less and points the same way in fifteen of sixteen company-runs; Tesla, with no effect to begin with, is the exception. Seven of eight gaps reproduce within thirty cents. The fact that the causal impact of the board's decision is a stable output builds our trust in its recommendations.

What does the framing do to the level?

As another check, we asked for the unconditional EPS forecast: the same eight questions as ordinary forecasts, with no alternatives. If the decision framing were neutral, this should land near the retain branch.

Medians, USD per share.

There is clearly more noise in the absolute EPS forecasts, but they do not move a lot when compared to the confidence interval. Again, we find this reassuring.

Share price instead of earnings

Hanson asked about price, so we asked too. Same eight companies, same alternatives, same context, with the outcome switched to the adjusted closing price on August 31, 2029, on the standard total-return convention: every dividend and distribution reinvested, adjusted only for splits.

Replace is again below retain everywhere, but where the earnings discounts were uniform, the price discounts are concentrated. Tesla loses a quarter of its median value on a forced Musk exit; 3M and Boeing lose about a tenth. Tesla had no CEO effect on earnings and a large one on price: the forecaster is pricing Musk into the multiple, not the income statement.

These results differ from Fire the CEO, which found replacements that added value at several of these companies: its alternatives appointed a willing, named successor within weeks, while ours fires first and searches after, and the forecaster prices that difference at several points of share value.

Take-aways

We explored whether AI forecasting might achieve what Hanson hoped for from his dump-the-CEO markets. For a board deciding on the CEO, the key is the differential impact of the decision. With our decision forecaster that differential is stable across repeated runs and consistent across companies, and, importantly, it reflects the causal impact of the decision rather than the potentially confounded conditional-market version. We find this encouraging.

We also showed that question formulation matters. We got different results from our Fire the CEO version of the questions, and by using EPS instead of stock price we got a markedly different answer for Tesla, because of the valuation multiple the market attributes to Musk.

We hope to keep exploring and expanding the use cases where AI decision forecasting can give decision makers more value on consequential topics.

Try decision forecasts yourself on our app by starting a query with the keyword "Decision".

The forecasts behind this piece, with their full rationales: EPS, run 1, EPS, run 2, EPS without the decision framing, and share price.