Accuracy record Live
How our numbers have done, scored live from the moment each window settles: the SettleMath model's leans and every signal, wins and losses, our first model next to the market's own price, our BTC price next to the official settlement value, and the settlement range. Nothing on this page is a backtest or a simulation, and most callers never publish theirs.
Where our Bitcoin price comes from
Main source: CF Benchmarks' Bitcoin Real Time Index, the official index these markets settle on, read once a second through the exchange's data feed. After every settlement we check it: in 1,423 markets, what we read matched the official settlement value to the cent in 1,414 of them, and never by more than $0.01.
Backup: our own estimate (our merged order books from three of the exchanges in the index), only if the index can't be read for a moment: over the last 7 days it was typically $1.97 from the official value, 95% of the time within $6.55. The model and the signals still use that estimate for now; it carries its own error into the numbers, so near the strike they never claim more certainty than it can give.
Nov 17 2026: Chance the result flips: from Bitcoin alone, beside the market's. The flip chance (the card and the chart's line) now comes from the model on Bitcoin alone: its distance to the strike, how much it moves, the time left and the seconds of the final minute already in. Until now the model it used also read the market's own price, which pulled the flip chance onto the market's number. The card now shows the market's chance beside it, so the two can disagree. The SettleMath model's other numbers are unchanged. Every change, dated
Record since · 2,453 windows scored · updated , after the latest window settled. This page updates by itself.
Our price called the side
99.7%
of the last 7 days' markets. Typically $1.97 from the official value; 95% within $6.55
Our price's error, measured
0.0049%
of Bitcoin's price: σe = $3.20 on about $65,095. The model carries it, so near the strike it never claims more certainty than our price can give.
Called by our price alone
97%
of markets ended more than $6.55 from the strike, clear for our price to call; the rest the Live page marks too close to call.
Signals, scored
Every built-in signal is logged the second it fires. Next to how often the side it pointed to won is what the market was already pricing for that side then: a signal is only telling you something new if its side wins more often than that price. Each record is also split by when it fired and by how busy Bitcoin was, so you can see the times it does better or worse; small groups swing a lot, so read them together with their counts. What each signal is, and the last 7 days as a picture to share: the Signals page. Information, not advice.
SettleMath model: buy signal (Oct 28 to Nov 14, retired)
| When it fired | Signals | Pointed side won | Market price then | Model then |
|---|---|---|---|---|
| All | 35 | 42.9% | 40.5% | 48.7% |
| 5 min or more left | 1 | 100.0% | 57.5% | 64.8% |
| 1 to 5 min left | 16 | 50.0% | 49.5% | 59.2% |
| under a minute left | 18 | 33.3% | 31.5% | 38.3% |
| By the hour it fired (Central time) | ||||
| Overnight, 12 to 6 AM Central | 3 | 33.3% | 9.3% | 16.7% |
| Morning, 6 AM to noon | 8 | 62.5% | 49.8% | 57.2% |
| Afternoon, noon to 6 PM | 10 | 40.0% | 43.8% | 52.3% |
| Evening, 6 PM to midnight | 14 | 35.7% | 39.4% | 48.1% |
| By how busy BTC was when it fired | ||||
| BTC quieter than usual | 18 | 50.0% | 50.9% | 60.5% |
| BTC busier than usual | 17 | 35.3% | 29.4% | 36.1% |
First one .
The SettleMath model (its buy signal retired Nov 14, 2026)
Over its 81 leans, the side the model leaned to won 38% of the time where the price said 45%: it did not beat the market. Its buy signal has stopped and it is off the leaderboard. The Lab’s top 3 on the leaderboard are the built-in signals now. Its full record stays here.
Its record (tap to open)
The SettleMath model starts from the market's own price and moves off it only where our recorded data says the price has been wrong: every indicator on the Live page, the other markets, the exchanges that move first and, in the last minute, the arithmetic of the seconds already averaged. It is refitted every hour on the last 14 days and scored here only on markets that came after. Each window, the first second it sits 3+ points from the price is logged with the result. The model is ahead when the side it leaned to wins more often than the price said; every lean counts, wins and losses.
It now leans only where it has earned it (the luck test, added Oct 30 Central time): at each hourly refit, a stretch of the window is used only if, predicting markets it was never fitted on (a walk-forward over the 14 days), it beat the price by more than luck explains (2 standard errors, over 100+ markets); the last minute's settlement arithmetic is tested on its own the same way, on the markets since it was added. Everywhere else the model is the market's price, so it may lean rarely or not at all; that is the honest answer when the readings say nothing the price doesn't. Its earlier leans stay in the record below.
| Which windows | Leaned | Its side won | The price said | The model said |
|---|---|---|---|---|
| All | 81 | 38.3% (31 of 81) | 45.1% | 49.9% |
| By how far from the price | ||||
| 3 to 6 points from the price | 68 | 36.8% (25 of 68) | 43.8% | 47.2% |
| 6 to 10 points | 6 | 16.7% (1 of 6) | 50.2% | 57.9% |
| 10 points or more | 7 | 71.4% (5 of 7) | 53.4% | 68.6% |
| By time left | ||||
| 5 min or more left | 44 | 31.8% (14 of 44) | 44.2% | 47.2% |
| 1 to 5 min left | 24 | 54.2% (13 of 24) | 53.2% | 60.7% |
| under a minute left | 13 | 30.8% (4 of 13) | 33.6% | 38.7% |
| By version | ||||
| Before the luck test: leaned on any edge | 81 | 38.3% (31 of 81) | 45.1% | 49.9% |
| Since the luck test: leans only past luck | 0 | – | – | – |
First one . Its buy signal (its chance of a side 3+ points above that side's asking price after the fee) is scored with the other signals above. Information, not advice; past results do not guarantee future results.
We tested our model against the market's price
We built a model of this market from the Bitcoin price, its volatility and the settlement rules, and scored it against the market's own price on the same seconds. It ties: the market already prices what our model knows, so the site no longer shows a model number. What is left where our arithmetic can still differ from the price is the final minute, below.
The Brier score is the average squared miss of a probability against what happened: 0 is perfect, and always saying 50% scores 0.250. Rows are the model's numbers as our collector stored them live: every 5 seconds, every second in the last 2 minutes, and only seconds when the market had a quote, so both are scored on the same rows. Since 10:15 PM CT Oct 22 2026, when the model began reading the market's quote (see the changes list), so this is the model as it works now; the first version's numbers ignored the market and are not mixed in.
| Time left | Windows | Rows | Model | Market | Model − market |
|---|---|---|---|---|---|
| 10-15 min | 2,453 | 144,555 | 0.217 | 0.217 | +0.0002 · no clear difference |
| 5-10 min | 2,453 | 147,153 | 0.162 | 0.162 | +0.0001 · no clear difference |
| 2-5 min | 2,453 | 86,923 | 0.102 | 0.102 | +0.0002 · no clear difference |
| 1-2 min | 2,227 | 110,570 | 0.067 | 0.066 | +0.0001 · no clear difference |
| 30-60 s | 1,374 | 28,219 | 0.052 | 0.052 | −0.0000 · no clear difference |
| 0-30 s | 606 | 8,137 | 0.033 | 0.033 | +0.0000 · no clear difference |
A lower Brier score here means a probability that matched past results more closely. It is a measurement of the past, not a trading result, and says nothing about profit after the spread and the fee.
Calibration: when a price said X%, how often was it Yes?
Rows grouped by the probability they gave (0-10%, 10-20% …). A well-calibrated number sits on the diagonal: of the rows that said 70-80%, about 75% ended Yes.
Our model The market Perfect calibration
| Said | Our model | The market | ||||
|---|---|---|---|---|---|---|
| Rows | Average | Yes | Rows | Average | Yes | |
| 0-10% | 1,684 | 7.1% | 2.3% | 1,265 | 7.7% | 1.4% |
| 10-20% | 6,170 | 15.9% | 15.4% | 6,365 | 15.8% | 13.7% |
| 20-30% | 13,361 | 25.5% | 24.2% | 13,791 | 25.4% | 24.5% |
| 30-40% | 21,693 | 35.3% | 32.6% | 21,815 | 35.3% | 32.6% |
| 40-50% | 30,938 | 45.2% | 43.2% | 30,279 | 45.2% | 43.3% |
| 50-60% | 30,519 | 54.7% | 53.2% | 30,331 | 54.7% | 53.0% |
| 60-70% | 20,234 | 64.6% | 62.8% | 20,339 | 64.7% | 62.5% |
| 70-80% | 12,186 | 74.5% | 70.9% | 12,550 | 74.6% | 70.8% |
| 80-90% | 5,995 | 84.2% | 82.6% | 6,314 | 84.3% | 83.8% |
| 90-100% | 1,775 | 93.2% | 93.9% | 1,506 | 92.8% | 94.0% |
Our model The market Perfect calibration
| Said | Our model | The market | ||||
|---|---|---|---|---|---|---|
| Rows | Average | Yes | Rows | Average | Yes | |
| 0-10% | 17,594 | 5.4% | 4.6% | 17,404 | 5.5% | 4.7% |
| 10-20% | 15,673 | 15.2% | 14.5% | 16,377 | 15.0% | 14.3% |
| 20-30% | 15,654 | 24.9% | 24.4% | 15,296 | 24.9% | 24.6% |
| 30-40% | 13,875 | 35.0% | 33.2% | 13,800 | 34.9% | 33.2% |
| 40-50% | 13,213 | 45.1% | 44.6% | 13,033 | 45.1% | 44.4% |
| 50-60% | 13,227 | 54.9% | 52.8% | 12,985 | 54.9% | 53.1% |
| 60-70% | 13,501 | 65.1% | 61.2% | 13,221 | 65.1% | 60.5% |
| 70-80% | 14,240 | 75.0% | 72.5% | 13,905 | 75.0% | 71.5% |
| 80-90% | 13,786 | 84.7% | 83.1% | 14,681 | 85.0% | 83.3% |
| 90-100% | 16,390 | 94.9% | 95.6% | 16,451 | 94.8% | 95.6% |
Our model The market Perfect calibration
| Said | Our model | The market | ||||
|---|---|---|---|---|---|---|
| Rows | Average | Yes | Rows | Average | Yes | |
| 0-10% | 24,983 | 3.1% | 3.2% | 25,114 | 3.1% | 3.2% |
| 10-20% | 7,245 | 14.8% | 17.7% | 7,243 | 14.7% | 18.0% |
| 20-30% | 5,397 | 24.9% | 24.7% | 5,328 | 24.9% | 24.6% |
| 30-40% | 4,732 | 35.0% | 35.8% | 4,668 | 35.0% | 35.3% |
| 40-50% | 4,399 | 44.9% | 41.5% | 4,366 | 44.9% | 41.8% |
| 50-60% | 4,026 | 54.9% | 52.3% | 3,980 | 54.9% | 52.3% |
| 60-70% | 4,150 | 65.0% | 64.2% | 4,120 | 65.0% | 63.8% |
| 70-80% | 4,539 | 75.1% | 75.5% | 4,442 | 75.1% | 75.2% |
| 80-90% | 5,847 | 85.1% | 86.7% | 5,838 | 85.3% | 86.2% |
| 90-100% | 21,605 | 97.2% | 97.7% | 21,824 | 97.1% | 97.6% |
Our model The market Perfect calibration
| Said | Our model | The market | ||||
|---|---|---|---|---|---|---|
| Rows | Average | Yes | Rows | Average | Yes | |
| 0-10% | 42,985 | 1.9% | 1.6% | 43,084 | 1.9% | 1.7% |
| 10-20% | 6,010 | 14.5% | 18.0% | 5,936 | 14.4% | 17.8% |
| 20-30% | 4,045 | 24.9% | 30.8% | 4,004 | 24.9% | 31.0% |
| 30-40% | 3,528 | 34.9% | 35.5% | 3,525 | 34.9% | 35.2% |
| 40-50% | 3,189 | 45.0% | 45.5% | 3,196 | 44.9% | 45.7% |
| 50-60% | 2,943 | 54.9% | 61.5% | 2,899 | 54.9% | 61.5% |
| 60-70% | 3,244 | 65.1% | 69.2% | 3,252 | 65.2% | 69.1% |
| 70-80% | 3,989 | 75.1% | 77.1% | 3,970 | 75.1% | 76.9% |
| 80-90% | 5,739 | 85.3% | 88.1% | 5,670 | 85.4% | 88.0% |
| 90-100% | 34,898 | 97.9% | 98.1% | 35,034 | 97.9% | 98.1% |
Our model The market Perfect calibration
| Said | Our model | The market | ||||
|---|---|---|---|---|---|---|
| Rows | Average | Yes | Rows | Average | Yes | |
| 0-10% | 11,415 | 1.7% | 1.8% | 11,419 | 1.8% | 1.8% |
| 10-20% | 1,343 | 14.6% | 19.3% | 1,339 | 14.5% | 19.3% |
| 20-30% | 739 | 24.7% | 33.3% | 735 | 24.7% | 33.3% |
| 30-40% | 518 | 34.6% | 41.1% | 522 | 34.5% | 40.8% |
| 40-50% | 464 | 45.2% | 59.5% | 459 | 45.1% | 59.0% |
| 50-60% | 450 | 54.9% | 63.8% | 452 | 54.8% | 64.2% |
| 60-70% | 488 | 65.3% | 76.0% | 483 | 65.3% | 76.2% |
| 70-80% | 821 | 75.3% | 79.5% | 819 | 75.3% | 79.5% |
| 80-90% | 1,283 | 85.5% | 86.6% | 1,284 | 85.5% | 86.3% |
| 90-100% | 10,698 | 98.1% | 98.4% | 10,707 | 98.1% | 98.5% |
Our model The market Perfect calibration
| Said | Our model | The market | ||||
|---|---|---|---|---|---|---|
| Rows | Average | Yes | Rows | Average | Yes | |
| 0-10% | 3,396 | 1.3% | 1.4% | 3,396 | 1.3% | 1.4% |
| 10-20% | 235 | 14.4% | 14.0% | 233 | 14.3% | 14.2% |
| 20-30% | 115 | 24.4% | 12.2% | 116 | 24.3% | 12.1% |
| 30-40% | 89 | 34.6% | 24.7% | 89 | 34.5% | 24.7% |
| 40-50% | 86 | 45.2% | 45.3% | 86 | 45.1% | 44.2% |
| 50-60% | 83 | 55.0% | 62.7% | 84 | 55.0% | 63.1% |
| 60-70% | 124 | 66.0% | 66.9% | 122 | 65.9% | 67.2% |
| 70-80% | 154 | 75.4% | 76.6% | 152 | 75.3% | 76.3% |
| 80-90% | 251 | 86.1% | 82.5% | 249 | 86.0% | 83.1% |
| 90-100% | 3,604 | 98.4% | 99.1% | 3,610 | 98.5% | 99.0% |
Which volatility scored better
The model runs on realized volatility (how much our BTC price moved in the last 10 minutes). It also computes the same number with the volatility Bitcoin options imply (the DVOL index): the "With implied volatility instead" line under the Live page's gauge, stored live every second like the main number. Whichever scores better over a month becomes the default.
| Time left | Rows | Realized | Implied | Realized − implied |
|---|---|---|---|---|
| 10-15 min | 144,554 | 0.217 | 0.218 | −0.0004 · no clear difference |
| 5-10 min | 147,153 | 0.162 | 0.163 | −0.0008 · no clear difference |
| 2-5 min | 86,923 | 0.102 | 0.103 | −0.0005 · no clear difference |
| 1-2 min | 110,570 | 0.067 | 0.067 | −0.0004 · no clear difference |
| 30-60 s | 28,219 | 0.052 | 0.052 | −0.0002 · no clear difference |
| 0-30 s | 8,137 | 0.033 | 0.033 | −0.0000 · realized lower |
The settlement range, scored
The yellow range on the Live page says where the official settlement value should land. If it is honest, the official value lands inside the likely range about half the time and inside the wide range about 90% of the time. Well above means the range is too wide; well below, too narrow.
| When | Windows | Inside the 50% range | Inside the 90% range |
|---|---|---|---|
| 10 min left | 2,453 | 56.2% | 91.1% |
| 5 min left | 2,453 | 57.5% | 91.6% |
| 2 min left | 2,453 | 58.8% | 91.8% |
| 1 min left | 2,453 | 56.4% | 89.8% |
| 30 s left | 2,453 | 56.2% | 90.5% |
Each indicator, scored
The 15 indicators behind the strength score (and the watched ones - liquidations, VWAP, open interest, options skew, who led the move, history vs the market, big trades on the market, streaks, orders pulled from the market's book and the ETH market moving first - before they get points), one by one: at 10, 5, 2 and 1 minute before the close, which way each pointed, how often that side won, and what the market was already pricing for that side. An indicator only adds something when its side wins more often than that price. The strength score's own rows show whether more points meant more wins. These records decide which indicators earn more points. Information, not advice.
| Indicator | 10 min left | 5 min left | 2 min left | 1 min left |
|---|---|---|---|---|
| RSI (14, one-minute) | 70.5% of 940, market 70% | 83.5% of 946, market 84% | 92.9% of 950, market 90% | 95.4% of 934, market 88% |
| Bollinger %B | 69.9% of 1,151, market 70% | 84.2% of 1,158, market 84% | 91.3% of 1,172, market 89% | 92.9% of 1,160, market 85% |
| MACD histogram | 67.6% of 1,341, market 67% | 76.1% of 1,337, market 76% | 76.9% of 1,344, market 73% | 75.6% of 1,338, market 65% |
| 1-minute change | 56.7% of 1,238, market 57% | 58.2% of 1,233, market 57% | 59.7% of 1,235, market 57% | 57.1% of 1,211, market 53% |
| 5-minute change | 72.7% of 1,300, market 72% | 70.0% of 1,270, market 70% | 71.0% of 1,286, market 67% | 70.5% of 1,280, market 62% |
| Trend: EMA 9 vs 21 | 61.3% of 1,353, market 61% | 73.7% of 1,360, market 75% | 82.7% of 1,361, market 81% | 86.6% of 1,362, market 75% |
| Taker-bought volume (11 exchanges) | 59.2% of 1,144, market 59% | 60.8% of 1,129, market 60% | 61.4% of 1,135, market 60% | 60.4% of 1,102, market 57% |
| Exchange books, bid side | 48.5% of 668, market 51% | 54.1% of 676, market 52% | 53.3% of 672, market 55% | 49.5% of 669, market 52% |
| Futures premium change | 51.1% of 804, market 50% | 50.7% of 850, market 51% | 47.9% of 823, market 50% | 46.4% of 828, market 46% |
| Whale orders, net 5 min | 62.2% of 1,096, market 62% | 65.1% of 1,093, market 64% | 64.9% of 1,067, market 62% | 63.9% of 1,064, market 60% |
| Market book, Yes side | 44.4% of 1,129, market 46% | 32.0% of 1,191, market 33% | 21.4% of 1,115, market 22% | 25.0% of 711, market 26% |
| Market price change, 60 s | 60.4% of 1,213, market 61% | 68.5% of 1,064, market 67% | 78.4% of 791, market 77% | 83.2% of 583, market 83% |
| Crowd by trades: taking Yes | 26.5% of 994, market 28% | 20.0% of 714, market 19% | 24.4% of 435, market 28% | 13.2% of 529, market 32% |
| The other 15-minute market vs this one | 51.8% of 245, market 46% | 51.8% of 170, market 44% | 50.3% of 143, market 45% | 51.8% of 114, market 51% |
| Coinbase premium change, 5 min | 52.5% of 606, market 50% | 51.2% of 633, market 51% | 51.8% of 655, market 52% | 48.3% of 654, market 49% |
| Liquidations, net 5 min (watched: no points yet) | 77.8% of 167, market 76% | 87.2% of 148, market 90% | 97.1% of 139, market 93% | 97.9% of 142, market 89% |
| BTC vs this window's VWAP (watched: no points yet) | 66.6% of 1,194, market 67% | 75.0% of 1,249, market 74% | 80.4% of 1,285, market 77% | 80.2% of 1,286, market 69% |
| New positions behind the 5-min move (watched) | 68.2% of 132, market 71% | 70.1% of 137, market 67% | 64.3% of 129, market 61% | 59.2% of 130, market 48% |
| Options skew, 15 min change (watched) | 63.3% of 768, market 62% | 68.8% of 773, market 68% | 72.5% of 726, market 68% | 72.3% of 762, market 62% |
| Spot-led vs futures-led move (watched) | 66.1% of 56, market 59% | 56.0% of 50, market 54% | 64.8% of 54, market 63% | 63.6% of 66, market 57% |
| History vs the market's price (watched) | 53.6% of 28, market 53% | 65.9% of 44, market 64% | 67.7% of 164, market 69% | 23.1% of 13, market 22% |
| Big trades on the market, 2 min (watched) | 73.4% of 1,027, market 72% | 82.3% of 1,251, market 83% | 90.7% of 1,338, market 90% | 95.1% of 1,337, market 90% |
| Streak: the windows before (watched) | 48.2% of 1,420, market 49% | 48.2% of 1,420, market 50% | 48.2% of 1,420, market 49% | 48.2% of 1,420, market 50% |
| Orders pulled from the market's book (watched) | 50.9% of 882, market 50% | 54.7% of 730, market 53% | 70.9% of 666, market 72% | 76.1% of 419, market 76% |
| The ETH market moving first (watched) | 48.4% of 854, market 48% | 42.8% of 760, market 44% | 39.3% of 720, market 40% | 34.5% of 449, market 35% |
| Crowd by contracts: taking Yes (watched) | 27.9% of 1,109, market 30% | 34.6% of 746, market 34% | 69.3% of 704, market 74% | 59.3% of 784, market 79% |
| Big money vs small trades (watched) | 77.3% of 944, market 75% | 83.6% of 1,237, market 85% | 91.4% of 1,324, market 91% | 95.5% of 1,343, market 91% |
| Market price velocity, 10 s (watched) | 56.4% of 828, market 56% | 57.7% of 688, market 58% | 64.7% of 487, market 65% | 69.5% of 364, market 70% |
| BTC's place in its 5-min range (watched) | 75.4% of 731, market 74% | 72.8% of 751, market 72% | 74.9% of 793, market 71% | 72.3% of 792, market 63% |
| Move accelerating, 1 min vs 5 min (watched) | 47.6% of 1,318, market 47% | 48.3% of 1,301, market 47% | 50.1% of 1,291, market 50% | 46.0% of 1,290, market 48% |
| Fast exchanges' move vs ours, 5 s (watched) | 47.0% of 640, market 49% | 51.9% of 595, market 52% | 47.5% of 573, market 46% | 46.1% of 557, market 51% |
| Other coins' 15-min markets moving first (watched) | 44.4% of 828, market 43% | 38.9% of 821, market 41% | 35.9% of 814, market 36% | 26.2% of 592, market 26% |
| Final-minute arithmetic minus the price (watched) | 40.7% of 317, market 44% | 30.7% of 329, market 30% | 30.4% of 263, market 27% | 37.9% of 190, market 37% |
| Pull of the nearest round $500 (watched) | 48.2% of 226, market 50% | 49.6% of 228, market 51% | 50.2% of 245, market 50% | 49.6% of 238, market 50% |
| Volume surge behind the 1-min move (watched) | 80.0% of 20, market 72% | 80.6% of 31, market 85% | 80.0% of 25, market 72% | 78.4% of 37, market 67% |
| Strength 70 or more points | 71.7% of 60, market 74% | 85.2% of 27, market 87% | 90.6% of 32, market 91% | 100.0% of 28, market 94% |
| Strength 40 to 69 points | 74.8% of 441, market 73% | 82.6% of 449, market 82% | 87.8% of 476, market 86% | 88.1% of 445, market 79% |
| Strength 1 to 39 points | 58.2% of 905, market 60% | 66.9% of 934, market 65% | 68.4% of 898, market 64% | 67.2% of 937, market 60% |
The last 14 days of windows (the indicator rows are kept 14 days); the first row counted is from .
The official index we show, checked every window
Since Nov 2, 2026 the chart and the Live page show the official index these markets settle on (CF Benchmarks' Bitcoin Real Time Index), read once a second through the exchange's data feed. To check that what we read is the real thing, after every settlement we take the index's own average over the final minute, from the seconds exactly as we read them, and compare it with the official settlement value. If what we show is right, the two match to the cent.
| The index's average, seconds close-60 to close-1 | |
|---|---|
| Windows checked | 1,423 |
| Matched to the cent | 1414 of 1423 |
| Median distance | $0.00 |
| 95% of windows within | $0.00 |
| Largest distance | $0.01 |
The settlement is the average of the index's 60 one-second values before the close, from close-60 to close-1 (the reading one second later, close-59 to the close, has not matched). A window counts only when we read all 60 of its seconds. Our own older estimate, which the model still uses, is scored just below.
The same check for ETH
Since Nov 3, 2026 the ETH Live page shows the official ETH index (CF Benchmarks' Ether-Dollar Real Time Index), read once a second. After every ETH settlement we compare its own final-minute average with the official value, the same way.
| The ETH index's average, seconds close-60 to close-1 | |
|---|---|
| ETH windows checked | 1,300 |
| Matched to the cent | 1289 of 1300 |
| Median distance | $0.00 |
| Largest distance | $0.01 |
Our own BTC estimate vs the official settlement value
The market settles on CF Benchmarks' index. Before Nov 2, 2026 the pages showed our own estimate of it, and the model, the settlement math and the signals still use it until they are moved to the index: our composite of three of the exchanges in it, so it never matches to the cent. This is how far apart they have been at the close.
| Last 7 days | Since the start | |
|---|---|---|
| Windows | 664 | 2,854 |
| Median distance | $1.97 | $2.74 |
| 95% of windows within | $6.55 | $9.53 |
| Largest distance | $10.43 | $97.38 |
| On average, ours minus official | +$0.51 | −$1.53 |
| Root-mean-square | $3.20 | $5.41 |
| Median, other timing reading | $2.15 | $2.74 |
| 95%, other timing reading | $6.69 | $9.45 |
| Merged order books: windows | 664 | 1,940 |
| Merged order books: median distance | $2.74 | $2.82 |
| Ours on the same windows: median distance | $1.97 | $2.70 |
| Merged order books: 95% within | $8.38 | $8.30 |
| Same side of the strike as the result | 662 of 664 | 2,807 of 2,854 |
An average offset over $1.00 either way would be subtracted by the model; the offset subtracted now is $0.00.
The merged order books (from Oct 28 2026). The official index is not an average of exchange prices: it merges the exchanges' order books every 200 ms and takes a depth-weighted mid of the merged book. We build the same thing from the five index exchanges we read and record it beside our estimate for every market (the "merged order books" rows). Since Oct 31 it has been the price the pages fall back to in a second the official index can't be read (since Nov 2 the pages show the index itself). The model and the signals still use the estimate scored above, until they are moved over and checked.
How the model works
The result is the average of the 60 one-second index values before the close. Let S be our BTC price now, K the strike, τ the seconds to the close, σ the volatility per second (realized, over the last 10 minutes), and σe how far our final-minute average has typically been from the official value (above).
- m = settlement seconds still to come (at most 60), n = 60 − m seconds already locked, A = their average, d = seconds until the settlement minute starts.
- Expected settlement value: E = (n·A + m·S) / 60.
- Its uncertainty: SD = (m/60) · σ · S · √(d + (m+1)(2m+1)/(6m)).
- Fair value for Yes: P = Φ((E − K) / √(SD² + σe²)), Φ being the standard normal distribution. While the strike is our own estimate (a window's first seconds at most), σe counts twice.
- Since Oct 22 2026, the market's own price: its traders see the official index, which our model doesn't use yet, so the gap δ between the official average and ours is read off the market's quote too. The quote implies δm = SD · Φ⁻¹(mid) − (E − K), known to within vm = SD² · ((h / φ(Φ⁻¹(mid)))² + 0.05²) (h = half the spread, at least 1¢); combined with δ's own spread σe²: w = σe² / (σe² + vm), δ = w · δm, v = σe² · vm / (σe² + vm), and P = Φ((E + δ − K) / √(SD² + v)). Early in a window SD is large and this changes little (1-2 points on average); in the last minute it does most of the work. On 244 windows of Oct 20-22 (fitted on half, checked on the other half), with 0-10 s left the model alone scored a Brier of 0.145, the market 0.0012 and this 0.0014. The predicted settlement range is centred on E + δ with spread √(SD² + v).
- The move needed to flip: BTC must average (60·K − n·A) / m over the remaining m seconds.
Simplifications, stated plainly: no drift, the same σ through the window, normally distributed moves (fat tails would show up in the calibration above, which is why it is published), and our price is not the official index (that is σe; a steady offset over $1.00 across 7 days would be subtracted from E).
Changes to how the numbers are made
Retired signals - every version of the Lock (retired Nov 10) and both versions of Indicators agree (until Oct 24) - no longer appear on the pages. Every one they logged, with its side, the market's and the model's price when it fired and the result: download the retired signals' records (CSV).
Indicators, kept or retired automatically
Every indicator is judged on its own running record (each market's four moments counted as the one result they share), and only when its readings pass 600, 1,200, 2,400 and so on, doubling, so checking often can't make luck look like a verdict. One that wins clearly more often than the market's price said (3 standard errors) is kept with the top points; one that wins clearly less is kept and read the other way; from 2,400 readings one within 0.5 standard errors of the price is retired. Kept ones are judged again every 4,800 readings after their verdict and go back on trial if they stop holding up. Points (since Nov 13 2026), from the record in the table above (the last 14 days at 10, 5, 2 and 1 minute left), for every indicator, watched ones too: each one whose side won more often than the price said by more than two standard errors earns that difference, a little less the less sure it is (times 1 minus 1 over z squared); the rest none; shared out of 100. The score says which way the indicators lean, not that following them pays. Download every indicator's record (CSV).
- · Fast exchanges ahead of ours retired: neither way helped; off the Live table and the scanner (after 2,403 readings; won -0.2 points against what the prices said)
- · Whale orders retired: neither way helped; off the Live table and the scanner (after 4,800 readings; won -0.2 points against what the prices said)
- · Other coins' markets moving first retired: neither way helped; off the Live table and the scanner (after 2,400 readings; won -0.2 points against what the prices said)
- · Streak retired: neither way helped; off the Live table and the scanner (after 4,800 readings; won -0.4 points against what the prices said)
- · Coinbase premium retired: neither way helped; off the Live table and the scanner (after 2,400 readings; won +0.2 points against what the prices said)
- · Orders pulled retired: neither way helped; off the Live table and the scanner (after 2,401 readings; won +0.2 points against what the prices said)
- · The ETH market moving first retired: neither way helped; off the Live table and the scanner (after 2,403 readings; won -0.4 points against what the prices said)
- · The other 15-minute market proven: kept, with the top points (after 600 readings; won +7.0 points against what the prices said)
- · RSI retired: neither way helped; off the Live table and the scanner (after 2,401 readings; won -0.3 points against what the prices said)
- · Bollinger %B retired: neither way helped; off the Live table and the scanner (after 2,401 readings; won +0.2 points against what the prices said)
- · 1-minute change retired: neither way helped; off the Live table and the scanner (after 2,400 readings; won +0.2 points against what the prices said)
- · Taker-bought volume retired: neither way helped; off the Live table and the scanner (after 2,400 readings; won +0.2 points against what the prices said)
- · MACD retired: neither way helped; off the Live table and the scanner (after 2,400 readings; won -0.2 points against what the prices said)
- · Trend: EMA 9 vs 21 retired: neither way helped; off the Live table and the scanner (after 2,402 readings; won +0.1 points against what the prices said)
- · Streak back on trial (the review's rule changed on Oct 26: every indicator judged again)
- · Big trades back on trial (the review's rule changed on Oct 26: every indicator judged again)
- · 5-minute change back on trial (the review's rule changed on Oct 26: every indicator judged again)
- · MACD back on trial (the review's rule changed on Oct 26: every indicator judged again)
- · Big trades retired: neither way helped; off the Live table and the scanner (after 1,200 readings; won +0.7 points against what the prices said)
- · 5-minute change retired: neither way helped; off the Live table and the scanner (after 1,217 readings; won +1.8 points against what the prices said)
- · MACD retired: neither way helped; off the Live table and the scanner (after 1,278 readings; won +1.7 points against what the prices said)
- · Streak retired: neither way helped; off the Live table and the scanner (after 1,200 readings; won -1.6 points against what the prices said)
- · How a number is made · Chance the result flips: from Bitcoin alone, beside the market's. The flip chance (the card and the chart's line) now comes from the model on Bitcoin alone: its distance to the strike, how much it moves, the time left and the seconds of the final minute already in. Until now the model it used also read the market's own price, which pulled the flip chance onto the market's number. The card now shows the market's chance beside it, so the two can disagree. The SettleMath model's other numbers are unchanged.
- · How a number is made · The bot: no 5-minutes-left rule, and empty places filled. From now on the bot also takes a fire with under 5 minutes left (the last 10 seconds aside). And while fewer than 3 signals pass its bar (30+ fires since found, ahead by a standard error), the next signals with 10+ fires since found and ahead on them fill the empty places, marked still being tested on /bot: weaker proof, more trades. An app before 6.9 keeps getting only buys with 5 minutes or more left; 6.9 has its own setting (box 11). Earlier trades stay as booked.
- · How a number is made · The bot: no price limits and no trend guard. From now on the top-3 bot, the ways to trade, the ETH bot and the app take a signal's first fire at any price, with 5 minutes or more left: the 25¢ floor and the 70¢ top price are gone, and so is the trend guard (it stayed out when the indicators, the crowd and the last minute had turned against the side). It trades far more often, cheap long shots and dear favourites included, so its swings are bigger. Every earlier trade stays as it was booked; a signal's record counts each fire under the limits it had then. The cool-down after a losing run stays. The leaderboard still judges signals on their fires at 85¢ or less. Apps: SettleMath Bot 6.9 has its own price range (a setup from before keeps 25-70¢ until its member changes it); an older app keeps getting only the buys from 25¢ to 70¢.
- · How a number is made · Top 5 and the Top 3 row stop at 70¢ too; the baskets keep 85¢. The ways to trade that make one buy a market, Top 5 and the Top 3 row beside them, now take fires up to 70¢ like the bot itself, for new trades only (every earlier trade stays as booked). Their own paper trades since Nov 14, split at 70¢: Top 3 made +6.5% a dollar under 70¢ and lost 6.9% at 70¢ and up; Top 5 lost 2.5% under and 4.5% over. The baskets went the other way (the Top 10 basket lost 11.5% under 70¢ and was flat over), so they keep 85¢. Two and a half days of trades is a short test: a reason for the change, not a promise.
- · New · The app: order types, the last 10 results, and the signals on the chart as they fire. SettleMath Bot 6.9's Trade by hand ticket has order types: Limit (a buy waits on the exchange's book until it fills or the market closes), Stop, Stop limit, Trailing stop, a trailing buy and If… (buy when seconds left, a price or Bitcoin against the price to beat say so), and take profit / stop loss set on a buy the moment it fills. The exchange has no stops of its own, so the app watches those itself while it's open and sends each one once, as a now-or-never order inside your max slippage; a sell never takes the bot's contracts, and each ends with its market. A Waiting orders list cancels any of them. The last 10 markets' results sit above the chart. And 6.9 draws what fires on its Live market chart, on both the Up price and the Bitcoin views: the bot's buy (a bigger arrow marked BOT), each leaderboard signal's first fire inside the bot's limits (any price with 5 minutes or more left), your own signals and the strategy you picked, each where and when it fired, pointing up for Up and down for Down. A line under the two prices says what fired in this market. The chip above the chart hides them for a clean chart. They are only drawn: the app still buys only what you switched on.
- · New · Ethereum's own signals, and the app trades Bitcoin and Ethereum. The ETH leaderboard now has signals found on Ethereum itself: every few hours the Lab's scan runs on ETH's own stored markets, and each find that holds up on its newer markets (not luck, ahead walking forward, better than buying its side every market) joins as "found on Ethereum". Its record there counts only markets after it was found, so it can join the ETH bot's top 3 only once those reach 30 fires and stay ahead. SettleMath Bot 6.6 has a Bitcoin switch and an Ethereum switch on the Signals tab: with both on it can buy one Bitcoin and one Ethereum market every 15 minutes, and the size and loss notes count both. The Ethereum switch opens once the ETH bot's own record from its start passes the bot's bar.
- · How a number is made · The bot: no buys above 70¢. The top-3 bot now skips any fire above 70¢ (it was 85¢), for its new trades only: every earlier trade stays as it was. Since Nov 10 the bot page has tracked the bot's buys at 70¢ and up beside the rest, on trades booked after the idea came up: the dear ones lost money while the cheaper ones made it, because a loss at 80¢ costs the whole stake for a 20¢ win. It is a short test, so it is a reason for the change, not a promise. In the SettleMath Bot app, the Top 3 bot, the other signals and your own signals follow the same limit (their records are now shown under it too, a limit chosen after seeing some of those markets); the other ways to trade keep 85¢ for now.
- · New · The app: the Live market is now a full trading screen. SettleMath Bot 6.5 turns Live market into a trading screen across the whole Bot tab: the price to beat next to Bitcoin's price now and how far apart they are, a countdown, the Up price as one-minute candles (or Bitcoin against the price to beat), the order book as a ladder with the spread, every trade in the market as it happens, and an order ticket that shows how many contracts, the average price from the book, the cost with the exchange's fee, what it pays and its break-even before you send. Your position shows its average cost and profit or loss now. One-click orders can be turned on in the ticket; orders still never go past the price limit shown, and trades by hand stay yours, not the bot's.
- · New · The app: a live market view with buying and selling by hand, and a Hide details button. SettleMath Bot 6.4 can switch the Your account card to Live market: the 15-minute market trading now, its prices, how many contracts people are waiting to buy on each side, and your position, with buttons to buy Up or Down or sell what you hold, each after a check with the exact numbers (Practice mode only shows what it would do). Trades by hand are yours, not the bot's: they don't count toward its limits or its numbers. A Hide details button at the top hides the explaining text; warnings about your bet size stay.
- · New · Testing: flip after a bad day. From a member's idea: once the bot's day is down a set amount, every buy for the rest of that day goes to the other side. The bot page now tracks it at several amounts over the bot's own trades, and SettleMath Bot 6.4 can trade it (pick it on the Signals tab; it uses the amount that has done best so far). It rests on only a few red days and that amount was picked by looking at them, so there isn't enough data to say it is profitable: a day that dips and then recovers can turn into a big red one. Use it at your own risk; it stays off unless you pick it.
- · New · The app: pick how the bot's buys are traded; every way side by side on the bot page. SettleMath Bot 6.3 adds a choice on its Signals tab. Top 3 is how the bot trades, and what every setup stays on until you pick something else. Top 5 buys the first of the top 5 signals to fire each market. The Top 10 basket buys every top-10 signal's first fire with 30% of your amount on each, up to 10 buys a market, so one market can take up to 3 times your amount, and it can buy Up and Down in the same market. Top 5 and the basket are paper tests on the same leaderboard and limits, a few days old: hypothetical records, much shorter than the bot's own, and past results don't promise future ones. The app's size and loss notes count the most one market can take. The bot page now shows every way on one chart under the bot's own, with each day's dollars up or down, and two newer paper baskets were added: one side a market, and only the signals that bet either way.
- · New · Post rewards: we may reward your posts about SettleMath. Members with an active plan can send in their own posts about SettleMath (YouTube, Instagram, TikTok or X) from the account page. Put your post link in the post; a week later we look at how many people watched it and may reward it: up to $1 per 1,000 views on YouTube, $0.75 on Instagram, $0.50 on TikTok and $0.25 on X, and up to $100 for one post. It is a thank-you we choose to give, not a job or a contract, and the rules are on the Post rewards page. Your referral link works as before, for telling someone yourself. The account page also asks, once, where you first heard about SettleMath.
- · New · The app: your day next to the bot's, a loss limit that fits your bet, and your suggested bet. SettleMath Bot 6.0 puts your day next to the bot's at your amount a trade and says where any difference came from (markets it didn't trade, your fills, your own signals). Next to the daily loss limit it shows how often that limit would have stopped the bot on its recent days: on the bot's record, a tight limit sold the dips its days usually recover from. A new setup's limit now starts at about 7.5 times the bet. Beside box 6 it shows your suggested bet per window for your Crypto balance (never more than 1.5% of it). Above that it warns you, asks you to tick "I'm sure" and checks once more at Start, with the bot's worst recent day in dollars at your size. It remembers where its window was (a second monitor, maximized or full screen included), opens centered with nothing cut off, uses its own messages and sounds instead of Windows' ones, and can send Windows notifications when it pauses or stops. One SettleMath key runs one copy of the app at a time.
- · New · The bot's top 3 keep their exact conditions private; one key runs one app. While a signal is one of the three the bot follows, Pro sees its name, record and place and can still simulate it in the Lab, but not its exact conditions; they show again once it leaves the top 3. A SettleMath key now runs one copy of the app (or one program) at a time. The Terms say the app and the Lab's signals are for the subscriber's own use. The home page shows the bot the same way as the bot page: pick your amount a window and its results show at that size.
- · How a number is made · The SettleMath model's buy signal retired; the Lab's top 3 are the built-in signals. Over its leans, the side the model leaned to won 38% of the time where the market's price said 45%: it did not beat the price, so its buy signal has stopped and it is off the leaderboard. The built-in signals are now the Lab's top 3 on the leaderboard, plus any that get vetted. Its full record stays on the Accuracy page. The model's numbers stay on the Live page for now, since Lab signals can read them.
- · New · Your amount a window first on the bot page; the simulator reaches every stored market. The bot page asks how much you would trade per window before it shows any results, then shows every amount at that size: its profit, week, month, normal range and trades. The Lab's simulator now runs over every stored market (about 30 days, the leaderboard's own reach; markets older than 14 days are checked every 5th second), not only the last 14 days. Signals from your Lab scan, or put together in the simulator, can go straight onto the members' board.
- · How a number is made · Vetting fixed: judged on a signal's whole record since it was found. No signal had ever been vetted, and a study of 1,914 Lab signals over 30 days showed why: the bar read only the last 500 markets and counted fires over 85¢ the bot never buys, so it passed almost nothing, and what it did pass then lost 8.2% a dollar. Now a signal is vetted on its whole record since it was found, counting only the fires the bot could trade: 100 or more, ahead by 2 standard errors. In the study, signals that reached this bar made +4.3% a dollar afterwards against +2.6% for every signal, and +4.2% against +0.7% when judged only on the later half of the month: one month, many near-identical signals, and a bar chosen on that month, so it's a sign, not a promise. A signal that only breaks even still reaches it within a month about 1 time in 7, so some vetted ones will be luck; it's checked again after every market. Vetted signals carry a badge on the leaderboard; the bot's own rules, its order and how much it stakes are unchanged.
- · New · Amounts at your bet size, and every past market on its own page. A switch at the top of the bot page shows its profit, its trades, its day-by-day boxes and its week and month at the amount a trade you pick ($10 by default), so its $80-$150 trades read in your own numbers. The home page has a box that shows what the bot did at your bet size over its last finished days, worst day and lowest point included. Every settled market of the last 30 days now has its own page under /markets with its strike, settlement value and result. The Status page says what's happening right now, and the site's jargon has plain-word tips on hover or tap.
- · New · Why the bot leans one way, and its normal range. The bot page, its feed beside the Live chart and the app now say why the bot is betting the way it is: how far Bitcoin moved over the last 24 hours, how many of the last 96 markets settled Down, and how many of its buys were Down. The bot page shades its normal range (the middle 80% of its finished days) behind its last 30 days, so a red day can be read as normal or unusual.
- · How a number is made · What to expect from the bot, and a suggested bet size. The bot page now leads with this week and this month (each also shown at $10 a trade) and a "What to expect" box from its whole record: how many days were red, the average day, its worst day and how soon that was made back, and red weeks. The bankroll box suggests an amount a trade from the bot's own record: the biggest share of a bankroll that lost half in a 30-day stretch in 2% or fewer of the stretches drawn from its recent days (about 1% when this was written; it is worked out again every 6 hours), replacing a looser rule that allowed several times more. A signal still tagged "could still be luck" no longer takes a top-3 place unless it has been vetted.
- · New · Strategies side by side, on paper. Four ways to trade the same leaderboard now run on every market with the bot's own rules, no money: the top 3 (the bot as it trades), the top 5, a top-10 basket with a smaller stake on each, and a master that follows whichever of them made the most in the last 48 hours. Each one's today, this week and all time is in the Lab's paper tests (linked from the bot page), so the best one can be picked on its live record.
- · How a number is made · The leaderboard looks for Up signals too. When 70% or more of the listed signals' bets are Down, the leaderboard's own scan leans its search toward Up signals and takes Up and either-way finds past 25 listed, up to 40. Every bar a signal must pass is the same.
- · How a number is made · New readings, and indicator points earned from the record. Two new readings, recorded every second and scored on the Accuracy page: the pull of the nearest round $500 and a volume surge behind the last minute's move; and the hour of the day and the weekend, as new Lab inputs. Each indicator's points in the strength gauge now come from its own record in the Lab's indicator table: one clearly ahead of the price earns points by how far and how surely; one clearly behind it earns them read the other way round; the rest earn none (the two new readings, until Nov 27, earn a share by how much of a full record they have, once ahead by one standard error). The top-3 bot now takes a signal that was greyed out and listed again only once it is vetted, and a benched one stays benched at least 6 hours. The leaderboard keeps at least 25 signals: a dip only greys a signal, and it is taken off for good after 2 days behind.
- · New · The Up bot, and the bot's buys on the Pro chart. The bot copies the leaderboard's top 3 signals, and while Bitcoin keeps falling those are Down signals, so its buys are Down. The new Up bot follows the leaderboard's best Up signal the same way so you can watch the other side: a hollow green badge on the Live chart (with "against the trend" while the last day of markets went Down) and its own record from today. It is a record, not a trade: the bot and the app only buy it if you tick that signal in the app. The Pro chart now marks the bot's and the Up bot's buys (both on by default), and its candles are drawn from CF Benchmarks' index itself, second by second.
- · How a number is made · The Lock is retired. The Lock (a side all but decided in the final minute, at a price that still pays) is off the site: off the chart, the leaderboard, the alerts and the Accuracy page. When a side is that clear, the market has almost always priced it already, so it paid little more than its price. Every Lock that fired stays in the retired signals download on the Accuracy page, wins and losses.
- · Data · The chart shows the official index. The Bitcoin price on the chart and the Live page is now the official index these markets settle on (CF Benchmarks' Bitcoin Real Time Index), read once a second through the exchange's data feed, instead of our estimate of it. In a second it can't be read, the page shows our merged order books' estimate as before. The model, the settlement panel's math and the signals still use our own composite for now; they move over once the recorded index has been checked against the official settlements.
- · New · The signal bot's trades checked against the order book. A reader pointed out that the bot's "real ask prices" assume the ask is still there when the signal fires, which is often exactly when it isn't: quotes get pulled and ask size vanishes when the price moves. Fair point. Every trade of the top-3 signal bot is now also priced against the order book we record every second, from the book one second after the signal fired: only contracts actually resting there at no more than 2¢ above the fired price fill, level by level with the fee, and if the ask was gone there is no trade. Its numbers sit beside the quote-price ones on the bot's page, its picture and its Discord scoreboard, and every trade since the bot's current version started that has a recorded book was priced the same way.
- · How a number is made · The SettleMath model leans only where it beat the price by more than luck. In its first two and a half days the model's side won 31 of the 81 markets where it leaned, 38.3%, where the price said 45.1%: it leaned on edges that were luck. Before, a stretch of the market was used whenever it scored even slightly better than the price on the newest fifth of the markets it was not fitted on, and in the last minute our settlement arithmetic was added without being tested. Now, at each hourly refit, a stretch is used only when a walk-forward test passes: the 14 days in five blocks, each block after the first predicted by a fit on the blocks before it alone, and the model ahead of the price there by two standard errors or more, market by market, over 100 markets or more. The last minute's settlement arithmetic is tested on its own the same way, on the markets since it was added (Oct 28). Everywhere else the model is the market's price: no lean, no line apart, no buy signal. It may lean rarely or not at all; that is the honest answer when the readings say nothing the price doesn't. Its earlier record stays on this page, and its leans from this change on are shown on their own.
- · How a number is made · The SettleMath model is back, built to move before the market. The model returns to the Live page in a new form. It starts from the market's own price and moves off it only where 14 days of recorded markets say the price has been wrong: 33 readings (the 15 indicators, the 12 watched ones and 6 new ones: the market's price velocity, Bitcoin's place in its 5-minute range, the move accelerating, the fastest exchanges ahead of our price, the other coins' markets moving first, and the final-minute arithmetic against the price), weighed by a ridge-penalized logistic fit for each stretch of the market, refitted every hour, its penalty chosen on markets it was not fitted on and reported on markets used for neither. On the chart its line is drawn twice as far from the price as its chance, so a lean is easy to see; its number, its edge and its record use the chance itself. Each market, the first second it sits 3 points or more from the price is logged and scored on this page ("The SettleMath model: is it ahead of the market?"), wins and losses; its buy signal (its chance of a side 3 points or more above that side's asking price after the fee, once a market) is on the chart, on the leaderboard and scored with the signals. The settlement-math card is gone; the Lock still uses that arithmetic. The home page's count of signals tested now starts Oct 23, when the site's own scan began: the scans before Oct 28, 4:31 AM CT were not counted, so the count starts from a low estimate of them (180 million: about 1,210 scan slots, 60% of them run, 250,000 signals a scan; the first counted hour averaged about 378,000).
- · New · The model number is gone; settlement math in the final minute instead. Tested four ways (which moves first, the fastest public prices, how much of a move is already priced when we see it, and eight days of our best model against the market's price at every time left), our model ties the market's price and never gets ahead of it. So the Live page no longer shows a model number, a model line or what the model would pay; the flip chance and the history still use it underneath, and its test against the market stays on this page. In its place, Pro members see our settlement math in the final minute: the arithmetic of the 60-second average, without the market's price, shown only when it differs from the price by 5 points or more, with the strike official and our price's error measured. The first such second of each window is logged and scored below ("The final minute") from this change on.
- · How a number is made · Lock v6: about twice the fires. Lock v5 fired in about one market in five. v6 was picked from about 110,000 versions replayed on ten days of markets (899), each as a $10 order placed 2 seconds after the signal at no more than 1c over its price, with the fee: with 1:00 to 0:08 left, our settlement math without the market's price at 97.5% or more on a side the market prices at 85-99.5c, or at 99% or more above that, up to 99.8c. It fired in about 39 markets in 100 and lost none, about 20% more in total than v5 (hypothetical). The extra fires are mostly at 99-99.8c, where each win is small; 99.9c is left out because a $10 order there wins 1c and pays a 1c fee. Every version that reached back past 1:00 left, or asked our math for less than 97%, lost markets, and waiting with a bid a cent or two under the price filled mostly when the market turned against the side. Its record restarts now; v5's is in the retired signals' download. Hypothetical results; see the CFTC Rule 4.41 statement.
- · How a number is made · Our settlement estimate is read 2 seconds earlier. A week of our Bitcoin price against the official settlement values (722 markets) found that the official 60-second average lines up best with our price taken 2 seconds earlier: seconds 61 to 2 before the close instead of 59 to 0. Over the last 7 days that brings the typical miss from $2.83 to $2.54 and the 95th percentile from $8.30 to $8.05, closer on 8 of the 9 days. Our error figure, the "too close to call" band and our own strike estimate follow it from this market on; past markets keep the reading of their time. The same check found that adding Gemini and Crypto.com, two more of the index's exchanges, made our price no closer (their prices sit about $10-14 above the official value), and that weighting our three exchanges differently doesn't help either, so the price itself is unchanged.
- · How a number is made · The model trusts the market's price fully; volatility split; Lock v5. An outside review of eight days of our numbers (713 markets) changed three things. (1) The model already read the market's own price for how far our Bitcoin price sits from the official index; a 1c floor on the spread and a 0.5c cap on the price held it near 99% where the market and the results said 99.9%. Both are gone and it now trusts the market's price fully at every time left: scored by log loss it improves at every time left and ties the market's own price. Plainly: the model's chance can't beat the price; what it adds is the settlement arithmetic in the last minute. (2) Volatility: the 1-second estimate ran about 17% low for the minutes before the settlement minute, and the 15-second one about 14% high inside it, so each is now used where it is right. (3) Lock v5: Lock v4 read the model, which in the last minute is the market's price - its fires were the windows where the favourite cost 98-99.9c, nothing more. v5 reads our settlement math without the market's price: with 1:00 to 0:05 left, 98% or more on a side the market prices at 90-99c. On those 713 markets it fired 187 times and lost none, +2.8% a bet (hypothetical), about 25 times a day; its case rests on about 80 fires the market priced at 95-98c. One loss costs about 40-70 wins. Its record restarts now; v4's is in the retired signals' download. The Lab's late sure favorite paper test (v4's rule) keeps running as the comparison, and a 98.5-99c price-only template joins the leaderboard: if they end level with Lock, the math adds nothing. Hypothetical results; see the CFTC Rule 4.41 statement.
- · How a number is made · Indicators judged on a sounder rule; points by tier. The automatic review that keeps, reverses or retires indicators gave a no-information indicator a false permanent verdict about 1 time in 5 within four weeks, and retired most real but small ones. Now each market's four moments count as the one result they share, verdicts are made only when an indicator's readings pass 600, 1,200, 2,400 and so on (doubling), kept needs 3 standard errors, retired needs 2,400 readings within half a standard error of the price, and a kept one is judged again on its readings since. False verdicts fall to about 1-2 in 100. Every indicator was judged again from scratch under it (Streak, retired yesterday, is back on trial). The strength score's points now come in tiers from the records: 10 for each indicator clearly ahead of the price so far, 6 for each a little ahead, none for the rest. A week's record proves no indicator; the score says which way the indicators lean.
- · How a number is made · Lock v4: the last minute's sure favorite; losing signals leave the leaderboard sooner; the Pro chart. Lock v3 fired with 4:00 to 0:10 left at up to 96c, and lost about 25% over its first 46 markets: a minute or more before the close a 93% favorite still loses often enough to cost more than its wins. Lock v4 fires only with 60 to 5 seconds left, the first second our model puts one side at 98% or more while it costs 50c to 99.9c. On the Oct 18-23 markets that rule bought 227 times in 395 markets and lost none (once in 226 with the order book's queue taken into account); each win is a cent or two, so one loss costs many wins. Its record restarts now; v3's is in the retired signals' download. On the leaderboard a signal down 40% or more after 30 new markets now comes off as well as one behind after 100. The Lab's two paper tests are tabs, the late sure favorite first. The Pro chart: one compact line top left, the signals folded into a chip, a clearer Back to now button (Alt+R) that also resets the zoom to the screen, the market's price and our model's chance in one pane, and in the final minute the settlement average so far with what the remaining seconds must average for Up. Hypothetical results; see the CFTC Rule 4.41 statement.
- · New · The Pro chart: Bitcoin candles with every signal. A new page, settlemath.app/chart (Pro): Bitcoin as candlesticks from our own price (15 seconds, 1 or 5 minutes, up to 24 hours back), each window's strike, the market's Yes price in a pane under it, and every signal marked Up or Down at the candle it fired in: the Lock as it really fired and your own signals as the server saw them. On the Live page every signal that fires now gets a box beside the chart saying which way it points and the price then; the settlement range is off the Live page.
- · How a number is made · The signal leaderboard ranks by what a signal made after it was found. The scan tries hundreds of thousands of combinations, so some look great on the markets it searched by luck alone. The leaderboard now ranks by each signal's record on markets that came after it was found or added ($10 a market, from $100): those with 100 or more such fires and still ahead come first as "held up on new markets", the rest are "still being tested" with their count so far; one behind after 100 is not listed and a Lab signal comes off for good. Lock counts from its current version, the templates from when they joined the board. The record over the last 500 markets is still shown under each name. Hypothetical results; see the CFTC Rule 4.41 statement.
- · How a number is made · Indicators are now kept, read the other way or retired automatically. Every 10 minutes each indicator's record on the Accuracy page (how often the side it pointed to won, against what the market's price said, at 10, 5, 2 and 1 minute left) is checked. After about 150 markets, one that wins clearly more often than the price says is kept for good with points in the Live page's score; one that wins clearly less often is kept for good but read the other way round; after about 300 markets, one that does neither is retired: no points, off the Live table and the Lab's scan, its record kept in a download on this page. Every decision is dated in the changes list. On the Lab, the leaderboard can be folded away, the simulator lists only the signals ahead over the leaderboard's markets unless you ask for all of them, and the buttons no longer cover the result on phones.
- · How a number is made · Lock v3: fires in about twice as many markets; the model scored as it works now. The lock now fires with 4:00 to 0:10 left when the market's favorite is at 84c or more but costs at most 96c and our model on its own, without the market's price, puts that side at 93% or more; it no longer also needs the model to be at least as sure as the market. On the Oct 18-22 markets, chosen on the older half and checked on the newer, it fired in 56% and 62% of markets (v2: 31% and 35%) and won 96.4% and 95.9% at about 93c, +3.1% and +3.3% a bet: nearly twice the signals for about 1.5 times the total. Its record restarts now; v2's is in the retired signals' download. The Accuracy page now scores the model on its rows since Oct 22, 10:15 PM CT, when it began reading the market's quote; the first version's rows ignored the market and had been dragging the record. Hypothetical results; see the CFTC Rule 4.41 statement.
- · New · Retired signals off the pages; the leaderboard keeps more Lab signals. Retired signals (the first two Locks and Indicators agree) no longer appear on the Accuracy and Signals pages; their records are a download from the changes list. The leaderboard now keeps up to 60 of the site's own Lab signals (it kept 20, so places 21 and on were always templates), shows how long each signal has been in the top 25 and at its place, and adds any of them to Your signals in one tap. On the Live page, a new card shows which side the trades are taking on this market and across 11 Bitcoin exchanges; the order book and what the model would pay are folded under More detail.
- · New · The Lab follows penny resting bids, market by market. A new paper test on the Lab: a bid resting at 2c on each side of every market from 14 to 3 minutes before the close, then cancelled, worked out on the market's recorded order book and trades - each bid waits behind the orders already at 2c and fills only when real trades reach it, and is held to the result. It shows the record since Oct 24, 6 AM (markets after the idea was chosen) and over every market with its order book recorded. No real orders. Hypothetical results; see the CFTC Rule 4.41 statement.
- · New · Three templates from other traders' strategies: pennies early, the favorite in the last minute. We tested strategies that accounts on the other 15-minute venue use on this market's own quotes (Oct 18-24, 494 markets). Sides costing 2c or less with 3 to 14 minutes left won more often than their price (7 of 204, positive in every stretch of time, but few wins so far); in the last minute the opposite held. The market's favorite at 90-99c with under a minute left won a little more often than its price. They join the Lab and the leaderboard as templates to be judged on their live record. Hypothetical results; see the CFTC Rule 4.41 statement.
- · New · Two "Overreach" templates: against strong indicator agreement once its side is already priced. We tested Indicators agree on 134 markets of stored indicator data (Oct 22-24). The indicators were right about as often as the market's own price already said; adding them to the price made its forecasts slightly worse on newer markets; and following strong agreement (80 points or more) lost money in every stretch of time. Going against strong agreement once its side already costs 60-85c made money overall in the same data, but not steadily, so it joins the Lab and the leaderboard as two templates to be judged on their live record, not as a built-in signal. Hypothetical results; see the CFTC Rule 4.41 statement.
- · How a number is made · Indicators agree retired; the leaderboard shows when it last updated and scanned. Indicators agree no longer fires and is off the chart, the leaderboard, alerts and the daily and weekly records. When the indicators line up strongly, the market has almost always already priced that side at 87c or more, so following them paid no better than the price (v1: 72% won at 72c; v2: 1 of 4 in its first two hours). Both versions' records stay on the Accuracy page as history; it comes back only if a test on the stored indicator data shows a version that holds up. The leaderboard now says when it last updated, when the site's own scan last ran and how many Lab signals were luck-checked again, and a signal that held its place says so.
- · New · Leaderboard: scans every 5 minutes, the luck check re-run at every update, and arrows for each move. The site's own scan now runs every 5 minutes (was 10). A Lab signal's luck check is no longer only the one it passed when it was found: after every update it runs again on the newest third of the markets, and a signal that no longer reads "worth watching" or "unlikely to be luck" comes off for good. Each signal shows how many places it moved since the previous update (green up, red down), and a newly found one is tagged New.
- · How a number is made · Indicators agree v2: the indicators' direction, only at a price that pays. The first version fired whenever the indicators' strength passed 70 points and won about as often as its price said (72% at 72c), so after the fee it lost. v2 fires once a window, between 14 and 1 minutes left, when the strength is 50 points or more one way, that side costs 40c to 85c, and our model on its own puts it at least 8 points above that price. Tested on the Oct 18-22 markets in three stretches of time, it made money in each; it was picked from a grid of rules, so expect the live record to come in lower. Its record restarts today; the first version's stays on the Accuracy page as history. Hypothetical results; see the CFTC Rule 4.41 statement.
- · How a number is made · Lock v2: two separate reads of the side, at a price that still pays; the lock's record restarts. The lock now fires with 3:00 to 0:20 left when the market's favorite is at 84c or more but costs at most 95c, our model is at least as sure as the market, and our model on its own, without the market's price, puts that side at 95% or more. On the stored windows of Oct 18-22, chosen on the older half and checked on the newer, it won 96.7% and 95.6% at about 92c, against the previous rule's 97.7% at 96c, about three times the return a fire, in about a third of the windows. Picked from many rules, so expect the live record to come in lower. The lock's record restarts at Lock v2 (10:48 PM CT Oct 23); the first version's stays below as history.
- · New · Signals can fire only once a window; scanner starting points. Each of your signals has an "only once per window" switch (on for new ones): the first fire in a window, not again each time it comes back. The Lab's scanner starts from Recommended settings, with Safe and Risky one tap away, and looks at every stored window unless you change it.
- · New · Lab signals on the leaderboard; nothing losing listed. The site now scans for signals on its own every few minutes (never while a member's scan runs). A result joins the leaderboard as a "Lab:" signal, named for what it does, only if its luck check on windows the scan never saw reads "worth watching" or "unlikely to be luck", it fires in at least 40% of markets and its side costs 20c or more on average (no lottery signals). At every update a Lab signal that is losing, firing in under 40% of markets or lottery-priced comes off for good, and no signal with a loss over the last 500 markets (all of them while there are fewer) is listed. The Lab's scanner switch now keeps signals that are worth watching too.
- · New · The signal leaderboard; Indicators agree off the chart by default. New for Pro on the Live page and in the Lab: every signal - the lock and indicators agree as they really fired, each template, and the models the site's own scanner finds (only those whose luck check on windows it never saw held up) - run as the strategy simulator runs it, $10 the first time it fires in each of the last 500 markets (all of them while there are fewer), from $100, and ranked; a scanner find also shows its record since it was found. Hypothetical results with the CFTC Rule 4.41 statement. Indicators agree no longer starts on the chart (its record stays here and on the leaderboard).
- · New · The lock signal: 84c and 90% in the last 3 minutes. The lock now fires with 3:00 to 0:20 left when the market's favorite is at 84c or more and the model puts the same side at 90% or more, at least as high as the market (before: both at 90, any time with 20 s or more left). On the stored windows of Oct 18-22 the old rule's side won 92.8% of the time at 91.7c on average, the new one's 97.4% at 95.9c: both about what the price said, so this makes the signal clearer, not a promise of profit. Locks before Oct 23, 3:46 AM CT used the old rule and stay in the record.
- · How a number is made · Where the window should settle: the range stays near our price when the market is less sure. When the market's price is less sure of a side than our price's distance from the strike suggests (a quiet hour, say), the range used to read the whole difference as our price being off from the official value, and could sit $40-60 under or over our price with minutes left. Our price has been within about two of its typical errors of the official value 95% of the time, so the range's centre now moves at most that far and the rest of the difference widens the range instead. Its share above the strike is still exactly the model's chance of Yes, so the model's number doesn't change. Ranges from Oct 23, 3:18 AM CT on are scored this way on this page.
- · New · The Lab: a signal scanner and a signal calculator. New in the Lab for Pro: the scanner tries signals made of one to three conditions on the page's numbers for as long as you choose (10 to 60 seconds) and lists the ones that did best when taken at the market's own asking price, chosen on the older two thirds of the windows and checked on the newest third, which the scan never saw; any result goes on the chart in one tap. The calculator runs any signal over the last windows second by second and shows what an amount a window, taken at the asking price the first time it fired, would have come to, the fee included. Both are hypothetical results and carry the CFTC Rule 4.41 statement. On the Live page, Your signals moved up beside the chart, with the built-in signals' switches and a list of recently deleted signals to restore.
- · How a number is made · The ETH page reads the market's price near the close too; the Against crowd line. The ETH fair value now reads the ETH market's own quote for how far our ETH price is from the official index, as the Bitcoin page has since earlier tonight, so it no longer sits far from that market in the last seconds. On the chart, "Against crowd" draws the crowd line turned over (the share of the market's trades taking No) and is on by default; the crowd line itself is a button away. The built-in signals can each be taken off the chart, and members' own signals can use "against the crowd" directly.
- · New · Big money against small trades; zoom on the chart; easier tables on phones. A new watched reading (no points yet): of the market's trades in the last 2 minutes, the share of contracts on Yes in trades of $100 or more minus the same share in trades under $5. In the trades measured so far, the side the small trades piled into won less often than its price said, and the big trades' side more often; with templates "Big money against the crowd". The chart zooms (pinch on a phone; the + and - buttons; Ctrl or Cmd with the wheel; drag to move, double-click for the whole window), and on phones the tables here open one row at a time.
- · How a number is made · The fair value reads the market's price near the close. The model now also reads the market's own quote for how far our BTC price is from the official index, which the market's traders can see and we can't. Early in a window this changes the fair value by a point or two; in the last minute it does most of the work, so the model is no longer 30-40 points away from the market in the final seconds when our price is a few dollars off (on Oct 20-22's windows, with 10 seconds or less left, the two used to be 30 points or more apart 42% of the time and the market's side won every time). The predicted settlement range moves the same way. Rows written before this keep the formula they were shown with. Also: a second crowd reading, by contracts instead of trades (watched, no points), and the crowd templates renamed "Against the crowd".
- · New · Two more watched indicators, the Lab, and today's card. Two more indicators are computed every second and scored here with no points yet: orders pulled from the market's own book, and the ETH 15-minute market moving before this one. New for Pro: the Lab, every indicator ranked by how much more often its side won than the market's price said (from the same records as this page), and a tester that runs any signal on the stored seconds of the last two weeks. The Live page adds today's windows and signals so far. The strength score, the signals' levels and the model's fair value are unchanged.
- · New · Six more watched indicators, the week's record, signals by conditions. Six more indicators are computed every second and scored on this page with no points yet: new positions behind the move (the futures' open interest), the options' skew, whether spot or the futures led the move, how often history says the market's price ended Yes, the market's own big trades, and streaks of windows. Each signal's record is also split by the hour it fired and by how busy Bitcoin was, and the last 7 days' record can be saved as a picture. Members' own signals can start from a template. The strength score, the signals' levels and the model's fair value are unchanged.
- · New · Live model thoughts, signal notifications, the last windows recapped, liquidations watched. New for Pro: what the model's fair value would pay on the order book right now (the price, the contracts and the expected profit by its own numbers, after the fee), and every signal as a notification on a phone or computer or in a Discord channel. The Live page recaps each settled window from the stored numbers, every signal with whether its side won, next to the signals' whole record, and the Pricing page replays the last finished window as Pro saw it. Forced liquidations on the futures exchanges are recorded and scored as a watched indicator with no points yet. The model's fair value is computed as before.
- · New · 15 indicators, a 100-point strength score, and the crowd line. Eight more indicators (a short trend, the exchanges' order books, the futures premium, the Coinbase premium, whale orders, the market's own price move and its trades, and the other 15-minute Bitcoin market) join the seven, each carrying points, 100 between them; the indicators-agree signal now fires when the strength score reaches its level one way. The crowd line shows which side the market's own trades are taking. A new section of this page scores every indicator on its own. The model's fair value is unchanged.
- · New · Predicted settlement range, indicators, signals, whale orders. New on the Live page (Pro): where the settlement is expected to land (the model's own distribution, read as a range of prices: nothing about how the fair value is computed changed), seven indicators, two default signals logged with their record below, members' own test signals, and whale orders from 11 exchanges. The 15-minute ETH market gets its own Live page with the same model.
- · How a number is made · σe added to the model. The fair value now includes how far our price has been from the official settlement value (σe, measured over the last 7 days), so near the strike it no longer claims more certainty than our price can give. Rows before it were scored as they were computed then; the difference shows only in a window's final minute, mostly its last seconds.
- · New · "Too close to call" and the closed window. The settlement panel says too close to call when the average is within our measured error (95th percentile, last 7 days) of the strike, and a window stays on the page through its last second.
- · Data · The record begins. Our collector starts recording; every number scored on this page comes from it, live, from this moment on.
Past accuracy does not guarantee future accuracy. These numbers are information, not advice; see the Disclaimer.
Indicators turned against the favorite
Taken off the chart on Nov 17 2026; its record stays here. On the Pro chart, a dashed line marked the first second a market's favorite (60¢ or more, with 1 to 14 minutes left) had all three turn against it: the indicators (three quarters of the last 30 seconds' readings), the crowd (at least 10 points of the last 2 minutes' trades on the other side) and Bitcoin's last minute. It is information, not a call: tested on 1,500 markets (Nov 10 2026), the favorite went on to lose about as often as its price already said, with this rule and with 324 variants of it, so it shows what the indicators see, not where the price goes next. Its record, every one since it started (renamed from "Possible reversal" on Nov 10):
| Checks scored | 475 |
|---|---|
| The favorite then lost | 127 of 475 (26.7%) |
| Its price said it would lose | 28.3% |
If the favorite loses more often after a check than its price said, the check found something the price had not. With few checks scored the two can differ by luck alone.