AI Stock Prediction Accuracy: Workshop's Graded Record
Workshop is an AI forecaster that registers yes/no calls before the answer exists, then lets code grade them against a plain baseline. Era 4, the current exam, has registered 32 calls: 6 settled and 26 still open. Graded against the baseline so far: 4 beat it, 0 tied and 2 lost (time-weighted, the measure the frozen rules use). Era 4 has 6 of 100 storylines settled, the sample the frozen rules require before any edge can be claimed. The gate is not reached, so no edge is claimed.
The record, era by era
Workshop keeps its record in eras. Each era has its own frozen rules, written before its first question and hashed so the bar cannot move. When the rules change, a new era starts and the old one is closed with its own numbers. Results from different eras are never added together: each ran under its own frozen rules and is a separate test.
| Era | Window | What was tested | Storylines settled | Calls graded | Beat / tied / lost | Mean edge per call | Status |
|---|---|---|---|---|---|---|---|
| Era 2 | 2026-07-21 to 2026-08-01 | Dated yes/no questions on public data (government releases and company earnings), registered before the answer and graded by code. | 43 of 100 | 43 | 18 / 15 / 10 | +0.0135 | Closed. Gate unreached — no edge claim. |
| Era 3 | 2026-08-01 to 2026-11-05 | A new exam under new frozen rules: the same kinds of dated questions, re-registered on a monthly loop, with the forecasting model pinned in the spec. | 56 of 100 | 62 | 34 / 5 / 23 | -0.0009 | Closed. Gate unreached — no edge claim. |
| Era 4 | 2026-11-05 to now | Will an S&P 100 company report EPS below the consensus rounded to the cent? Will a government release print above its bar on first release? | 6 of 100 | 6 | 4 / 0 / 2 | +0.0126 | Open. The live exam. |
The mean edge here is averaged per call. The gate test itself averages per storyline, so the figures on every open question can differ slightly. Before the frozen eras began in July 2026, Workshop made directional market calls, graded differently (hit rate against a simple momentum rule); that older record is on the Scoreboard.
Edge is the baseline's Brier score minus Workshop's, time-weighted over the days a call was open. Positive means Workshop beat the baseline, zero means it matched it, negative means it lost. A forecaster that simply copies the baseline scores exactly zero. Beat, tied and lost count individual settled calls; a call within 0.0005 of zero counts as tied. A storyline is a cluster of related questions, which is why the gate counts storylines rather than calls.
What is on the board now
20 earnings calls on S&P 100 companies and 12 government data calls are on the board.
| Settled call | Workshop said | Baseline said | Result |
|---|---|---|---|
| Lowe's Earnings (Nov 18): Lowe's Met or Beat the $2.89 Estimate. We Said 7% | 7% | 11% | Met or beat it |
| Target Earnings (Nov 18): Target Met or Beat the $2.07 Estimate. We Said 12% | 12% | 12% | Met or beat it |
| Disney Earnings (Nov 17): Disney Met or Beat the $1.69 Estimate. We Said 10% | 10% | 11% | Met or beat it |
| Home Depot Earnings (Nov 17): Home Depot Missed the $3.95 Estimate. We Said 42% | 42% | 28% | Missed the estimate |
| Nvidia Earnings (Nov 17): Nvidia Met or Beat the $2.53 Estimate. We Said 5% | 5% | 11% | Met or beat it |
| Charter Earnings (Nov 6): Charter Met or Beat the $9.24 Estimate. We Said 45% | 45% | 28% | Met or beat it |
Open now, soonest first:
Browse everything in Markets and Economy.
How to read these numbers
A Brier score is the squared gap between a probability and what happened, averaged over calls; the Brier score explainer works one through by hand. A baseline is the answer you would give knowing only how often this kind of thing has happened before. It is computed by code from past cases before each call is registered, as described in what a base rate is.
The point of the baseline is that raw accuracy flatters. A forecaster can look right most of the time on questions that almost always resolve the same way. Grading against the baseline asks the harder question: did this forecast beat what the plain history already said? That is also the defence against the base rate fallacy.
What this record does not show
What this record does not show:
- An edge. Under the frozen rules an edge can be claimed only after the gate (100 storylines) is reached and one pre-registered test, a one-sided Wilcoxon signed-rank test at the 5% level, is passed. Until then the numbers above are point estimates, not a claim.
- Certainty from small samples. A handful of settled calls can look good or bad by luck.
- Opinion. Columns and culture picks are labelled as opinion. Anything code cannot settle, such as rumor odds, is marked unscored and never counted. No language model grades anything.
- A trading record. These are dated yes/no calls, not trades, and nothing here is investment advice.
The closed eras say the same thing about themselves: Era 2: gate unreached — no edge claim; Era 3: gate unreached — no edge claim.
Common questions
How accurate are AI stock predictions?
There is no single answer, and most published figures are claims rather than graded records. Workshop's own record is above. Era 4, the current exam, has registered 32 calls: 6 settled and 26 still open. Graded against the baseline so far: 4 beat it, 0 tied and 2 lost (time-weighted, the measure the frozen rules use). Accuracy alone is not the test; the test is whether the forecast beats the plain base rate. Era 4 has 6 of 100 storylines settled, the sample the frozen rules require before any edge can be claimed. The gate is not reached, so no edge is claimed. See the Brier score explainer for how that is scored.
Can AI predict the stock market?
This page does not claim that it can. Workshop tests something narrower: whether an AI forecaster can beat a plain base rate on dated yes/no questions, such as whether a company will report earnings per share below the analysts' estimate. Era 4 has 6 of 100 storylines settled, the sample the frozen rules require before any edge can be claimed. The gate is not reached, so no edge is claimed.
How is Workshop graded?
Each call is registered before the answer can be known, with the public source that will settle it. Code grades it: earnings to the cent against the reported figure, government data on its first release. The forecast is scored with the Brier score against a baseline, and copying the baseline scores zero. The frozen rules cannot be changed mid-exam.