# What We Refuse to Predict (And What Can Actually Be Forecast)

> By Scott Covert · July 14, 2026 · AI Stock Market Impacts
> Source: https://aistockmarketimpacts.com/what-we-refuse-to-predict.html
> No price targets, no buy/sell calls, no forecast stated as certainty — on this page or any other.

**The verdict, first: rankings and direction are forecastable. Levels and timing are not.** Sixty years of forecasting research, from Tetlock's tournaments to two centuries of market data, keeps landing on the same split: relative claims ("this industry reprices harder than that one") show real, measurable skill, while point claims ("the S&P hits X in June") fail at rates that should embarrass the people selling them. Our engine only makes the first kind of claim. This page is the standing list of everything in the second kind — published, so you can hold us to it.

## Two bets, and everything you need to know sits between them

You'd be an idiot not to bet that the Buffalo Bills score at least one touchdown somewhere in their first five games next season. In 2023, NFL teams scored a touchdown on about **21% of drives** league-wide (the Bills on just over 30%), and a team gets roughly 10–11 drives a game — so "at least one TD per game" computes to about **98%**. You would also be an idiot to bet on the exact temperature in Timbuktu 3.31 years from today: weather models lose their skill entirely about two weeks out, and forty years of better satellites have improved that wall by roughly **one day per decade**.

Same universe, opposite bets. The first is a **sum of many chances**, asks only "at least one," and rides a slow stable statistic. The second is a **single point on a chaotic path**. Markets contain both kinds of bet, tangled together.

## The four questions that sort every prediction

1. **Aggregate or point?** A season of drives / an industry of firms — vs one game, one stock, one day.
2. **Ranking or level?** "A beats B," "at least one" — vs "the index closes at exactly X."
3. **Slow force or fast?** Demographics, capex cycles, adoption curves — vs headlines and positioning.
4. **Do wobbles wash out or compound?** Mean-reverting fundamentals — vs leverage unwinding into leverage.

Left side: forecastable. Right side: chaotic. Every engine output was designed to sit on the left.

## The measured track record of people paid to predict

- Analyst price targets: about a **45% average miss**
- Professional directional market calls: about **46.9%** — a coin flip beats the experts
- Recessions: **148 of 153 missed** looking one year ahead
- Tetlock (284 experts, ~82,000 forecasts): accuracy beats chance ~1 year out, hits dart-throwing-chimpanzee level **3–5 years out**
- The other half: inside ~18 months, skill is real — the Good Judgment Project beat intelligence analysts with classified access, and its best forecasters at 300 days beat regulars at 100. Method extends the window; nothing repeals it.

## What survives at long horizons: relative claims

- Industry-vs-industry ranking signals persist across **212 years** of market history
- Roughly three quarters of the difference in industry valuations traces to **relative profitability over the following 15 years**
- Starting valuations say almost nothing about next year (near-zero correlation) but rank decade-ahead returns **almost monotonically**

Long horizons kill event prediction and reward structural ranking — the opposite of how most people hold their forecasts.

## Can anyone see a crash coming?

Nobody can call the day: the August 2007 quant meltdown began with "virtually no signs of turmoil" in any major index, and bubbles have been formally shown to be observable ex post but unpredictable ex ante. But **fragility is measurable in advance**: one published concentration gauge spiked before 100% of the worst 1% of monthly drawdowns in its sample; sharp price run-ups lift two-year crash odds from ~20% to **53–80%**; slow credit buildups flag coming crises with real statistical skill; the IMF adopted a fragility-ranking heuristic to rank what the next shock breaks.

**You cannot predict the earthquake. You can absolutely grade the buildings.**

## The trap: right theme, dead stock

Internet traffic grew relentlessly — Cisco fell **86%** from its peak. Solar deployment boomed — Solyndra, SunEdison, and Q-Cells died building it. British railway construction kept expanding after 1845 — the railway index fell two-thirds from its mania peak. The theme being right does not save the stocks: value migrates away from producers, toward adopters. This is why we model **industries, not tickers**.

## Our premise, stated so you can reject it

Everything the engine outputs is conditional on one explicit bet: AI lands with industrial-revolution-scale force while the hard physical limits stay hard (power plants take years, fusion isn't scheduled, chips still need fabs). The honest structure of any forecast from us: **the odds our premise holds × the odds our ordering is right given the premise.** We defend only the second part.

Falsifiers (a premise you can't falsify is a religion): AI revenue stalls as a share of the economy for years; the deployment gap stops closing; buildout spending never converts to margin outside the shovel-sellers.

## The Refusal List — what you will never see from us

- **No dated calls.** Nobody has demonstrated that skill anywhere in the measured record.
- **No price targets.** A 45% average miss is astrology with a Bloomberg terminal.
- **No single-stock predictions.** Cisco fell 86% while its theme came true.
- **No certainty theater.** A forecast without its premise and confidence class is marketing.
- **No crash calls, no all-clears.** Fragility rankings, never timing.

## What we do claim — and how hard to hold it

- **Relative rankings of 30 industries** across five timeframes, conditional on the stated premise
- **Direction of structural repricing** — firm at 1–2 years, structure-not-events at 5–10
- **Fragility flags** — which crowded, levered, capex-heavy exposures amplify a shock — as conditionals, never timing
- **Declining confidence by horizon, on purpose** — 10-year numbers are sorting devices, not prophecies

## How to read any forecast, including ours

1. **Find the premise.** If the forecaster won't state what the forecast assumes, close the tab.
2. **Sort the claim with the four questions.** Left column: listen. Right column: entertainment.
3. **Treat every number as a rank, never a level.** The useful sentence is "A over B, and here's the mechanism" — not the decimal places.

---

*Sources include Tetlock & Gardner (Superforecasting, 2015), Mellers et al. (Psychological Science, 2014), Khandani & Lo (2007), Greenwood, Shleifer & You (2019), Schularick & Taylor (2012), the CAPE literature, and the 2026 evidence review behind our Predictability Charter — 68 findings, every citation independently verified before publication.*

*From AI Stock Market Impacts — rankings, direction, and fragility across 30 industries and 179 cross-industry effects. https://aistockmarketimpacts.com*