Every forecast site tells you what it can see coming. This page is the other list — the things nobody can predict, with the research receipts — and the narrower set of claims that survive on the honest side of the line.
Rankings and direction are forecastable. Levels and timing are not. Sixty years of forecasting research, from Philip 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.
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 the 2023 season, NFL teams scored a touchdown on about 21% of drives league-wide, the Bills on just over 30%, and a team gets roughly 10 to 11 drives a game. Run that math and the Bills scoring at least once per game comes out near 98% — across five games, the "at least one touchdown" bet is about as close to certainty as sports offers. 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, a hard wall that forty years of better satellites and bigger computers has only nudged — forecasting skill has improved by roughly one day per decade.
Same universe, same physics, opposite bets. The difference isn't how smart the bettor is. The first bet 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, where a wobble today compounds into a different world by 2029. Markets contain both kinds of bet, tangled together — and almost everyone selling forecasts hopes you never learn to tell them apart.
Before believing any market claim — ours included — run it through these:
| Question | Forecastable side | Chaotic side |
|---|---|---|
| Is it an aggregate or a point? | Sum of many independent chances (a season of drives, an industry of firms) | One path, one date (a single game, a single stock, a single day) |
| Is it a ranking or a level? | "A beats B," "at least one," "top five vs bottom five" | "The index closes at exactly X" |
| Is the force slow or fast? | Demographics, capex cycles, adoption curves, regulation | Headlines, positioning, one man's mood on a Tuesday |
| Do wobbles wash out or compound? | Mean-reverting fundamentals | Path-dependent cascades, leverage unwinding into leverage |
Every output our engine produces was designed to sit in the left column. That's not modesty. It's the only place the evidence says skill exists.
These aren't cherry-picked embarrassments; they're the measured averages of entire professions. Analyst price targets miss by about 45% on average. Professional directional market calls, graded across thousands of them, land around 46.9% — a coin flip beats the experts. Economists as a group missed 148 of 153 recessions looking just one year ahead. And in Tetlock's famous twenty-year study of 284 experts and roughly 82,000 forecasts, accuracy on world events beat chance about one year out, decayed steadily, and hit what he called dart-throwing-chimpanzee level three to five years out.
Now the half nobody quotes: inside that short window, skill is real and large. The Good Judgment Project beat professional intelligence analysts with access to classified information, and its best forecasters looking 300 days out beat regular forecasters looking 100 days out. Method matters, aggregation matters, and discipline extends the window — it just never repeals it.
Here's the part of the research that built this site. When you stop asking "what will the market do" and start asking "what will this do relative to that," the picture flips. Industry-versus-industry ranking signals have shown persistent, statistically serious spreads across 212 years of market history. Around three quarters of the difference in how industries get valued traces to their relative profitability over the following fifteen years — the slow stuff, in other words, is most of the story. Starting valuations can't tell you what next year holds (the correlation is close to zero at twelve months), but they rank decade-ahead returns almost monotonically, cheapest cohort to richest.
Long horizons kill event prediction and reward structural ranking. Which is precisely backwards from how most people use forecasts — confident about next quarter, shrugging about the decade. The evidence says: shrug about next quarter. The decade is where the signal lives, if you only ask it ranking questions.
Nobody can call the day, and the proof is brutal: the August 2007 quant meltdown began with "virtually no signs of turmoil" in any major index — a 12-standard-deviation loss arriving out of a quiet sky. Bubbles have been formally shown to be observable after the fact but unpredictable before it. But fragility — which exposures break hardest when something hits — is measurable in advance, and the record here is startling. One published concentration gauge spiked before 100% of the worst 1% of monthly drawdowns in its sample. Sharp price run-ups lift the two-year odds of a crash from a baseline near 20% to between 53% and 80%. Slow credit buildups flag coming crises with real statistical skill. The IMF itself adopted a fragility-ranking heuristic — not to predict shocks, but to rank what the next shock breaks.
You cannot predict the earthquake. You can absolutely grade the buildings. That single sentence separates the forecasting that works from the forecasting that sells.
One more finding from the history, and it's the one that should genuinely worry you. Internet traffic grew relentlessly through the 2000s — and Cisco, the company selling the very plumbing of that growth, fell 86% from its peak. Solar deployment boomed through 2012 — while Solyndra, SunEdison, and Q-Cells died building it. British railway construction kept expanding after 1845 — while the railway stock index fell two-thirds from its mania peak. Across every technology wave on record, the theme being right did not save the stocks, because value migrates — usually away from the producers of the new thing and toward its adopters and their customers.
This is why we model industries, not tickers, and why a strong industry score on our matrix is never a claim about any single company inside it. The wave can be real and your surfboard can still snap.
Everything the engine outputs is conditional on one explicit bet: that AI lands with industrial-revolution-scale force across the economy — labor, materials, drugs, media, logistics, the works — while the hard physical limits stay hard (power plants take years, fusion isn't scheduled, chips still need fabs). If you think AI is mostly a productivity gadget, our forecasts are not for you, and no engine math will rescue the disagreement. The honest structure of any forecast from us is two-part: the odds our premise holds, times the odds our ordering is right given the premise. We only defend the second part. You own the first.
And because a premise you can't falsify is a religion: ours breaks if AI revenue stalls as a share of the economy for years, if the deployment gap between model capability and business adoption stops closing, or if the buildout's spending never converts to margin anywhere outside the vendors selling shovels. We track all three in public.
AI Stock Market Impacts maps AI disruption across 31 industries and 204 cross-industry effects — rankings, direction, and fragility, recalibrated continuously since February 2026, with every parameter change logged and traceable. No targets, no timing, no theater.
See the 31-industry matrixThis page is the research half of our honesty. For the model-specific half — the eight real-world shocks that would actually break our own rankings, and the public signals that would catch it first — read What Could Break Our Read.
If you want to know when the rankings move — which industries changed places and what evidence moved them — get the "what moved" note. It goes out when the model changes, not on a content calendar.