Keeping score on the numbers this site shows and the markets they come from: how often a prediction market’s favorite actually wins, how often a stock moves more than its options priced for earnings, how our own Fair Value model is doing live, and how often our AI commentary checker sends back an unsupported figure. Every number is recomputed daily from records we keep; nothing is picked by hand, and the misses stay in.
We record how every market we track settles and its price before trading closed. Across 10,868 outcomes in 3,215 events that traded at least $10,000, outcomes priced at 70% or more a week before close happened 91% of the time, against an average price of 89%. Long shots priced at 30% or less happened 8% of the time, against 9%. The furthest any price band strayed from its result: outcomes priced 40–50% happened 43% of the time, against 45% priced.
Priced vs happened, a week before close
Each dot is a band of prices; on the dashed line, outcomes happen exactly as often as priced. Bigger dots hold more markets, and the bars show the 95% range for that many.
By price band
Brier 0.124
Priced
Outcomes
Avg price
Happened
95% range
0–10%
3,477
2.9%
1.8%
1%–2%
10–20%
870
14.5%
14.9%
13%–17%
20–30%
898
24.6%
26.1%
23%–29%
30–40%
748
34.6%
35.2%
32%–39%
40–50%
951
45.4%
43.3%
40%–46%
50–60%
1,275
52.6%
50.7%
48%–53%
60–70%
524
64.6%
63.7%
60%–68%
70–80%
441
74.5%
75.3%
71%–79%
80–90%
453
84.8%
85.9%
82%–89%
90–100%
1,231
95.8%
97.7%
97%–98%
Category
Outcomes
Favorites (70%+) happened
Brier
Sports
4,353
79% vs 82%
0.193
Politics & Elections
3,002
95% vs 91%
0.066
Culture & Media
850
90% vs 92%
0.080
Crypto
807
89% vs 89%
0.095
Markets & Companies
748
97% vs 92%
0.075
Economy & Fed
387
89% vs 91%
0.103
Tech & AI
346
87% vs 88%
0.098
World
322
89% vs 84%
0.089
Brier score: the average squared gap between price and result, 0 for perfect foresight, 0.25 for always saying 50%. Lower is better.
The same, a day before close (32,175 outcomes)
Priced vs happened, a day before close
Each dot is a band of prices; on the dashed line, outcomes happen exactly as often as priced. Bigger dots hold more markets, and the bars show the 95% range for that many.
By price band
Brier 0.154
Priced
Outcomes
Avg price
Happened
95% range
0–10%
8,626
2.5%
2.7%
2%–3%
10–20%
2,602
14.5%
17.6%
16%–19%
20–30%
2,692
24.7%
28.0%
26%–30%
30–40%
2,681
34.7%
37.3%
35%–39%
40–50%
4,381
45.1%
44.3%
43%–46%
50–60%
4,198
54.0%
52.2%
51%–54%
60–70%
2,017
64.1%
60.2%
58%–62%
70–80%
1,242
74.3%
66.7%
64%–69%
80–90%
1,084
84.5%
79.0%
76%–81%
90–100%
2,652
96.8%
96.2%
95%–97%
Category
Outcomes
Favorites (70%+) happened
Brier
Sports
22,053
75% vs 83%
0.192
Politics & Elections
3,230
98% vs 94%
0.037
Culture & Media
1,731
88% vs 92%
0.081
Weather & Climate
1,458
88% vs 90%
0.133
Crypto
1,303
91% vs 92%
0.061
Markets & Companies
1,018
87% vs 94%
0.103
Economy & Fed
530
97% vs 95%
0.059
Tech & AI
465
99% vs 93%
0.035
World
387
96% vs 95%
0.055
Brier score: the average squared gap between price and result, 0 for perfect foresight, 0.25 for always saying 50%. Lower is better.
Before each report, the at-the-money straddle prices how far the stock is expected to move. For companies worth $2 billion or more, the stock moved more than that on 34% of 7,181 reports since 2021, and the typical move was 0.69 times what was priced. If moves followed a bell curve and straddles were fairly priced, a stock would finish outside about 42% of the time; earnings moves have fatter tails, which lowers that, so read the gap as a sign straddles run somewhat rich rather than a precise figure.
The Fair Value screener publishes long and short picks every trading day, and each day’s picks are frozen the night they are published. The current model (v11.3) has published 29 daily cohorts since 2026-10-02.
Over 21 trading days, its long picks have averaged −27.0% against the S&P 500, beating it 0% of the time across 25 matured pick-days on 5 names, so a few stocks decide the figure.
Returns are gross of costs, through 2026-11-11. Earlier model versions are kept on the full record page and never pooled with this one.
Our market commentary is written by an AI model and checked before it is published: every figure has to match the data the model was given, and what it says about a company’s results has to match the reported numbers. A section that fails goes back to the model once; anything still unsupported is cut. In the last 30 days, 121 commentary runs were checked: 115 sections went back for a rewrite, 51 sentences were cut, and 37 runs passed unchanged.
Counted since 2026-11-02, when the checker started recording: 121 runs, 115 sections rewritten, 51 sentences cut.
How these scores are kept
Prediction markets
When a prediction market we track settles, we record how it resolved and keep its daily price history. A market’s price is read as the chance it gave that outcome. For every yes-or-no market that settled yes or no and traded at least $10,000, we take its last daily price at least seven days, and separately one day, before trading closed, group outcomes by that price and count how many happened. Fifteen-minute up-or-down contracts, parlays and multi-game combinations are left out. Each outcome in a multi-outcome event counts once, so an event with eight candidates contributes eight.
A price a day before close can already reflect a result that is partly known, such as a vote count under way, so the week-before figures are the cleaner test. Records start when we began capturing settlements, with earlier events added where we still had their history, so the sample grows every day.
Earnings moves
Before each earnings report we record the price of the at-the-money straddle on the first expiration after the report, as a percentage of the stock price, and afterwards the stock’s close-to-close move across the report. Any trading day between the report and that expiration is ordinary volatility the straddle charges for but the report move never gets, which would make straddles look overpriced, so only reports whose options expire within two days are scored. Stocks under $5 are left out.
Our own models
The Fair Value screener’s picks are recorded the night they are published and never changed. Results are measured against the S&P 500 and are gross of trading costs. When the model is retrained, the new version starts its own record.
Past results say nothing certain about future ones. This page is for information only and is not investment advice.
