One question a day: is Bitcoin about to have a big day. We put a number on it before the day opens, and the next morning the market answers. Right or wrong, the answer stays on the board.
Not "will Bitcoin go up". Nobody calls that reliably, us included, and we print how badly we do at it. The question here is whether today is going to be a big day at all. That one can be called.
Yesterday we said 12%. Bitcoin had to swing 2.43% to count as a big day. It swung 0.70%, so it was not, and the call went on the board as right.
14 indicators read the same day and each one votes. No single one of them matters. What matters is how many point the same way, and today 50% did. Watch that number, because it is usually lower than you would expect: most days they disagree.
The 14 are not a secret: every one of them has an article on this site and the simulator draws them for you. What is not published is the part that took the work: how a reading becomes a score, and what agreement is allowed to be worth. That part is ours, and it stays in the building.
Four readings go in and one number comes out. Every morning they get marked against what happened, and whichever one was furthest off loses some of its say. Keep being wrong and you stop being listened to. Nobody makes that call. The scoring does.
14 indicators on the daily candle, the only one of the four with measured skill of its own.
What the perpetual market is paying to stay long. Crowded longs lean the day slightly the other way.
Sentiment from the curated stream, bounded so it can nudge and never steer.
Trend state against the 200-day. Up or down, nothing in between.
Swipe for act two
Every forecast gets marked the next morning against what Bitcoin actually did. A day that did not resolve cleanly is thrown out rather than guessed. Here is the whole record, misses included.
The number is a pool of forecasters whose weights follow their own graded record, scored every day from 2026-09-26.
The Brier score is how far our number sat from what happened, so lower is better. 0.25 is what you score by saying fifty percent every single day and never thinking about it. Threshold is the middle of the last ninety days: half of them moved more than that, half less. It moved is what the day did, high to low as a share of the open. The record is young; judge it when the count is in the dozens.
The forecast above is about whether today is a big day. This is the other half of the same question: when in the day the movement actually lands. It is two years of hourly Bitcoin prices and one more day every morning, in New York time, because that is where it shows up.
Mean absolute hourly move, New York time · 17,520 hourly prices · one day at a time, no pooling · rank correlation across all 24 hours 0.75 between the two halves · loud against flat on weekday hours t = 20.4 · computed 2026-10-04 · source CoinGecko
It is not the American session. Bitcoin is loudest from 10:00 to 13:00 in New York, every weekday, and then it goes flat. Not when New York goes home, three hours later, but when London does. You are looking at the part of the day when Europe and America are both at their desks, and it ends when Europe leaves. The dead hours are early morning, 05:00 to 09:00 New York, after Asia has gone to bed and before America is up. Friday is the odd one out and empties in the evening instead. Sunday evening is busy, because that is Monday morning in Asia. Saturday has almost no rhythm at all. Seven dials and not one, because a Saturday looks nothing like a Tuesday.
We also went looking for the hours it peaks and bottoms out, across 1512 different windows. There is nothing there. The time of day tells you how much Bitcoin moves, not which way it goes.
A forecast nobody acts on is an opinion. So four paper accounts trade this one, five thousand dollars each, Bitcoin only, and every fill is published.
Reads the Barometer at full resolution: the pool behind the volatility number with each expert's weight, the four direction forecasters, the fourteen indicators behind the composite, and the intraday clock for the day of the week it actually is. Then it decides for itself: open, rest a bracket, close or hold. It is given no rule about what any of it means, and it is told plainly that the direction call is measured near a coin flip. Long and short.
The same reading of the Barometer as StormAI, from the same snapshot at the same instant, with one thing added: it is shown its own past decisions, the thesis it wrote for each one, the falsifier it named, and how that day was graded. Nothing else differs. So the gap between the two is exactly what it is worth to a model to be able to see its own record.
The control. It copies StormAI exactly: the same instant, the same size, the same stop, the same instrument. The only thing it decides for itself is which way, and it decides that at random. It exists to isolate one question. If reading the forecast is worth anything, StormAI has to beat a coin that copies every other choice it makes.
No model at all. It trades only when the Barometer calls a STORM or SEVERE day and stands flat otherwise: a fixed rule, kept as the benchmark it was registered as. Direction comes from the price, never from the forecast; the position closes at the end of the UTC day. It answers the question that was actually asked, whether the forecast is tradeable, without a language model anywhere in the loop.
Trading Trades from the first trigger 51 to go 6.67 graded a week done 2026-11-27
It is not on a schedule. Over the last seven days the watcher looked 672 times and acted 78 of them, standing still for the other 594. Every one of those looks is recorded, because otherwise patience and a broken job look identical.
I woke up because price made a new day high at 85130, but the whole day's range is still only 0.47%, about a fifth of the 2.42% threshold. The forecast is CALM, with pHigh at 0.20 for a Sunday. That number comes mostly from the climatology expert (72.5% of the pool's weight), and the CALM band has only one graded day behind it, so it tells me little. The direction inputs are noise (fit-time AUC 0.52), so the new high is no reason to go long. A bracket on a quiet Sunday morning would probably get whipsawed for small amounts, and the loud hours, 19:00 to 21:00 New York time, are still about 15 hours away. I am holding. If price prints above 86790 before the UTC day ends, today's range has gone past the threshold measured from the 84732 low, and a breakout long would have been the right trade.
Wrong if: a print at or above 86790 before 2026-10-05T00:00:00.000Z
I'm staying flat at 4984.24 equity with nothing pending. What woke me is a new Sunday high at 85130. Including it, the day spans 84732 to 85130, only 0.47%, which is 19% of the 2.42% threshold. It's 04:05 in New York and this hour averages 0.19%, so the break is small and doesn't change the CALM call (0.20 chance of a big day, 72.5% of the weight on the same-weekday base rate). On my record, the last two CALM-day holds were right, and the pool's volatility calls are 6 right out of 8. Chasing this print long would be a direction bet, and the direction score is measured at noise. A bracket wide enough to clear noise and the 15-minute fill slippage would need about 1% of room before the loud 19:00 to 21:00 New York block, which a CALM Sunday is unlikely to give. A long would have been the right call if 86000 (about 1% above spot) prints before 00:00Z.
Wrong if: a print at or above 86000 before 2026-10-05T00:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Band is CALM. This arm trades STORM and SEVERE only.
The band is CALM, with pHighVolatility at 0.198. About 72% of the pool's weight sits on the Sunday base rate (clim, 0.167), and har3 at 0.24 agrees. Yesterday's range was only 0.70%, and today has moved 0.05% so far against a 2.42% threshold. The direction inputs are noise and the news forecaster is saying nothing, so I see no reason to pick a side. A bracket for a breakout would be a bet against a low-vol forecast, and CALM's record is only one graded day, too few to trust or to fade. The loud hour that woke me only means larger hourly moves in both directions, so on a quiet Sunday it mostly adds noise to any stop. A print at 86830, about 2.4% above the day open (roughly a threshold-sized move), before the day ends would show a breakout bracket was the right call and this hold was wrong.
Wrong if: a print at or above 86830 before 2026-10-05T00:00:00.000Z
I'm staying flat at 4984.24 equity with nothing pending. This wake-up is Saturday's loudest block (21:00 to 23:00 New York), but 'loud' here means only about 0.20 to 0.22% a typical hour, about 0.6% over the three hours. A bracket that clears noise and covers the spread would need about 1% of room, and this block isn't likely to provide it. Sunday is called CALM with a 0.20 chance of passing the 2.42% threshold. 72.5% of that number comes from the same-weekday base rate, and the pool's volatility calls are 6 right out of 8. The day so far spans 84759 to 84804 (0.05%), and the direction inputs (5.6, pCloseAboveOpen 0.516) are noise. If spot moves 1% to 85630 before the block ends at 04:00Z, a breakout bracket would have paid and this hold was wrong.
Wrong if: a print at or above 85630 before 2026-10-04T04:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Band is CALM. This arm trades STORM and SEVERE only.
The pool calls Sunday CALM, with pHighVolatility at 0.198. Most of that weight (72.5%) sits on the same-weekday base rate, and har3 (Sunday-aware) agrees. Yesterday's range was only 0.70%, and the day so far has used about 2% of the 2.42% threshold. A breakout bracket would be betting against the forecast. The direction inputs are noise by their own measure, and the news forecaster is saying nothing, so there is no basis for a directional open. The volpool's own record is short (6 of 8 right, and only one CALM day graded), so I'm not trading against it or around it. I'm standing aside. If price prints 86100 (about +1.5% from 84804, most of the way to a threshold-sized day) before the UTC day ends, a long breakout would have been the right call and this hold was wrong.
Wrong if: a print at or above 86100 before 2026-10-05T00:00:00.000Z
I'm staying flat at 4984.24 equity. The pool calls Sunday CALM with a 0.20 chance of the range clearing the 2.42% threshold, and 72.5% of its weight is on the same-weekday base rate (2 of 12 recent Sundays). Its volatility calls are 6 right out of 8, and Saturday's CALM call held with a 0.70% range. The day so far spans 84759 to 84804 (0.05%), and the direction inputs (5.6, pCloseAboveOpen 0.516) are noise, so they give no reason to open a position. The clock does show Sunday's loudest block (19:00 to 21:00 New York, about 0.47% an hour) inside this UTC day. But a bracket placed far enough out to clear ordinary noise and the 15-minute fill slippage I saw on 2 October would need a move of about 1.5% each way, which a calm Sunday rarely gives. If 86030, about 1.5% above the open, prints before the day closes, a long bracket would have paid and this hold was wrong.
Wrong if: a print at or above 86030 before 2026-10-05T00:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Band is CALM. This arm trades STORM and SEVERE only.
The pool puts the chance of a high-volatility Saturday at 0.12. Most of its weight sits on the same-weekday base rate (0.71) and the Saturday-aware har3 (0.22). Only har1, har2 and the persistence expert, about 5% of the weight together, carry over yesterday's 3.97% range. A calm read gives a breakout bracket nothing to work with, and the direction inputs are noise (5.6 neutral, news missing), so there is no trade worth paying a spread for. The calm-band record has no graded days yet, and the overall Brier is no better than a coin, so this is a weak read and not an edge. If 86,250 prints (about +2% from spot) before the day ends, the calm call was wrong and a bracket above would have paid.
Wrong if: a print at or above 86250 before 2026-10-04T00:00:00.000Z
I'm staying flat with 4984.24 in equity and nothing pending. Saturday is forecast CALM, with a 0.12 chance that the range clears the 2.43% threshold. Most of that number comes from the same-weekday base rate (weight 0.71) and the weekday-aware har3 model (weight 0.22). The pool's own volatility calls are 5 right out of 7. The CALM band itself has no graded days yet, so I'm taking it with that gap in mind. So far the day spans 84492 to 84530 (0.04%), and the clock shows quiet hours now and a weekend with little spread between its loudest and quietest blocks. That leaves little room for a breakout bracket to run, and yesterday a bracket filled about 1% worse than its level because the watcher only looks every 15 minutes. Direction is noise (AUC 0.52), so I have no reason to open long or short. This hold is wrong if price prints above 85770, about 1.5% over the open, before midnight UTC. A long bracket would have paid in that case.
Wrong if: a print at or above 85770 before 2026-10-04T00:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Band is CALM. This arm trades STORM and SEVERE only.
The STORM call has already come true: the day has run 3.50% from 86983 down to 84041, or 1.45 times the 2.42% threshold. So the band has nothing left to say about how big the rest of the session will be. The price-break wake-up is a fresh day low, but the only thing that would make shorting that break pay is direction, which is noise at an AUC of 0.52. STORM calls have also been right on only 2 of 5 graded days, and the pool is leaning 62% on a same-weekday base rate drawn from 12 days. The clock is getting quieter from here: average hourly moves fall from 0.41% now to about 0.18% by NY hours 21 to 23. That leaves little room for a breakout to run before a stop is hit by ordinary noise, so I stay flat. I am wrong if Bitcoin prints below 82,800 (about 1.5% below spot) before 04:00 UTC, because that would mean a short taken with the break would have paid.
Wrong if: a print at or below 82800 before 2026-10-03T04:00:00.000Z
I'm staying flat at 4984.24 equity with nothing pending. Spot is 84041, a new day low below the earlier 84560, so the day now spans 84041 to 86983, a 3.50% range and 145% of the 2.42% threshold. My last hold named 'below 84400 by 21:00Z' as the level that would make a position right. It has printed, so on its own terms that hold was wrong, and a short bracket under the morning low would have paid. That does not make chasing the low now a good trade. The size call is long since used up, direction is noise (AUC 0.52), and after this 14:00 New York hour (0.41%) the clock fades through 15:00 (0.33%) into the quietest hours (0.18 to 0.28%), so a fresh short has little room to run. My record also shows the 15-minute watcher filling about 1% late on a fast break. If price prints 83200 (another 1% lower) before 21:00Z, staying flat was the wrong call and a short with the break would have been right.
Wrong if: a print at or below 83200 before 2026-10-02T21:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Already holding. This arm carries one position at a time.
The STORM call has already come true. The day's range is 2.87% against a 2.42% threshold, so whatever the volatility forecast knew is already in the price. The loud 10–12 block is over, and from 15:00 to 23:00 New York time the clock shows average hourly moves falling from about 0.33% to 0.18%. That leaves little room for a fresh break of the day's high or low to run after the close. The bracket at 87050/84480 is still live until 18:00 UTC and already covers a late break, so a second position or bracket would only add exposure to a range that has mostly played out. I don't trade the direction score: it is noise, and the news forecaster has gone silent.
Wrong if: a print at or below 83900 before 2026-10-03T00:00:00.000Z
I'm staying flat at 4984.24 equity with nothing pending. Today has already run 84560 to 86983, a 2.87% range that is 118% of the 2.42% threshold. That settles the STORM call, and it says nothing about the size of what is left. Only two loud hours remain (13:00 and 14:00 New York, about 0.41 to 0.42% each), then the clock goes quiet. A new bracket would have little room to run. My own record also shows a bracket filling about 1% worse than its level because the watcher only looks every 15 minutes, and that trade then stopped out for -37.46. Direction is noise (AUC 0.52), so the 5.7 score gives no reason to open a long or a short. If price breaks below the day low and prints 84400 before 21:00Z, a short bracket under the low would have been the right call and this hold was wrong.
Wrong if: a print at or below 84400 before 2026-10-02T21:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Already holding. This arm carries one position at a time.
The day has already moved 2.87%, more than the 2.42% threshold, so the STORM call has paid off whatever its weak record says (2 of 5 right in STORM). Nothing is open. The existing bracket at 87050 above and 84480 below sits just past today's high (86983) and low (84560), expires at 18:00Z and covers the loud 11:00 to 13:00 New York hours. A 1.2% stop on $1500 risks about $18, so it already expresses the one view I have: that the day may stretch past its extremes. Opening now from the middle of the range at 85683 would be a bet on direction, and the snapshot says the direction call is noise. I'm leaving the bracket as it is. The read is wrong if the day high prints again before 18:00Z, because a long opened now would then have been the better call.
Wrong if: a print at or above 86983 before 2026-10-02T18:00:00.000Z
The long I bought at 86809 hit its stop and is closed (record shows -37.46), so the account is flat at 4984.24 equity with nothing pending. The day has run 84560 to 86983, a 2.87% range that is already 118% of the 2.42% threshold. That settles the STORM call, which says nothing about the size of whatever move comes next. Direction is measured at noise (AUC 0.52), and this 11:00 New York hour is the loudest of the day, which makes a move against a new position as large as one for it. My record today shows a bracket that filled about 1% worse than its level because the watcher only looks every 15 minutes, so I won't pay that cost again for a range call that is already spent. I'll be wrong if price prints above 87740 before 21:00Z, where a long opened now would have reached a 2.4% target.
Wrong if: a print at or above 87740 before 2026-10-02T21:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Already holding. This arm carries one position at a time.
Today's range is already 2.87%, past the 2.42% threshold, so the STORM call has paid out for today whatever the band's weak record (2 of 5 right in STORM). That record is the reason I am sizing small, not trusting the call. The next three New York hours, 10 to 12, are the loudest of the day, at roughly 0.5 to 0.6% per hour. That is enough room to push past the day's high (86983, 1.2% above spot) or its low (84560, 1.6% below), so I am arming a bracket just outside both and taking whichever side breaks, with no view on direction because the direction score is noise. My read is that the range keeps growing. If price instead slides back to 85200, in the lower half of the range, without breaking either end, then the loud hours are being spent inside the range rather than past it, and the read is wrong. A stop-out at 1.2% of a $1,500 position costs about $18, well inside the equity limit.
Wrong if: a print at or below 85200 before 2026-10-02T18:00:00.000Z
I'm keeping the long I bought at 86809. Spot is 85933, so it is down $25.26. Its 1.2% stop near 85767 is $166 (0.19%) below spot, about a third of this 10:00 New York hour's average move of 0.61%. A stop-out would cost about $30 in total, only about $5 more than closing now. Direction is measured at noise (AUC 0.52), so that gap does not justify paying a spread to close early, and I'm not treating this repeated stop-within-reach wake-up as new information. The day's 2.87% range has already passed the 2.42% threshold, so the STORM size call is settled and says nothing about the next move. With the loudest block of the day running now, I think the stop will probably print before 21:00Z, and I'm rating my chance of being right low.
Wrong if: a print at or below 85767 before 2026-10-02T21:00:00.000Z
Mirror of the treatment's bracket: same levels, size, stop and expiry. The side is drawn at random when a level prints, with or against the break.
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Already holding. This arm carries one position at a time.
The Barometer said this would be a big day, and that has already happened. The day's range is 2.87%, which is 118% of the 2.42% threshold, so the forecast has nothing left to tell me about how much further the day will move. Its record also gives me no reason to trade more on it: 2 right out of 5 STORM days, and an overall Brier score about the same as always saying 50/50. The 10:00–12:00 New York hours are loud, but loud means big moves either way, not a direction, and the direction signals are noise. I have nothing in the snapshot that says a break of the day's high (86983) or low (84560) will keep running. If price prints 87900 before 20:00 UTC, a bracket set above the day high would have caught the move and holding was the wrong call.
Wrong if: a print at or above 87900 before 2026-10-02T20:00:00.000Z
I'm keeping the long I bought at 86809. Spot is 86104, so it is down $20.34. Its 1.2% stop near 85767 is $337 (0.39%) below spot, which is less than one average 10:00 New York hour (0.61%), so this stop-within-reach wake-up is the same one I have held through all day. A stop-out would cost about $30 in total, only about $10 more than closing now. Direction is measured at noise (AUC 0.52), so that gap is not worth paying a spread to close. The day's 2.87% range is already past the 2.42% threshold, so the STORM call is settled and says nothing about the next move. The loud 10:00 to 13:00 block still has hours to run before 21:00Z, so I think the stop is a bit more likely than not to print, and my confidence is below even.
Wrong if: a print at or below 85767 before 2026-10-02T21:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Already holding. This arm carries one position at a time.
The Barometer's call for today has already played out: the range so far is 2.87% against a 2.42% threshold, 118% of it, so the STORM label has nothing left to say about the rest of the day. The record behind the label is also thin. STORM days are 2 right out of 5, and the overall Brier score is slightly worse than a coin. A breakout bracket above the 86983 high, timed for the loud 10:00 to 12:00 New York hours, would be a bet that a move already this large keeps going, and nothing in the snapshot measures that. The direction inputs are noise by their own measurement, so they cannot pick a side either. I am holding. A print above 88000 before the end of the day, about 1.2% above the current high, would mean a breakout long armed now was the right call and this hold was wrong.
Wrong if: a print at or above 88000 before 2026-10-02T23:59:00.000Z
I'm keeping the long I bought at 86809. Spot is 86312, so it is down $14.34, and its 1.2% stop near 85767 is $545 (0.63%) below spot. That is about one average 10:00 or 11:00 New York hour (0.61 to 0.63%), so this wake-up is the same stop-within-reach signal I have already held through many times, now during the loudest block of the day. A stop-out would cost about $30, about $16 more than closing now. Direction is measured at noise (AUC 0.52), so that gap does not justify paying a spread to close. The day's 2.87% range has already passed the 2.42% threshold, so the STORM size call is settled. Adding up roughly 1.2% of typical movement in the hours left before 21:00Z, I put the chance that the stop prints before then a little above even.
Wrong if: a print at or below 85767 before 2026-10-02T21:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Already holding. This arm carries one position at a time.
The STORM call is already settled for today: the range so far is 2.87% against a 2.42% threshold (118% used). So the vol forecast says nothing about what is left of the day, and the band's own record is weak anyway (STORM 2 right of 5, Brier 0.32, overall slightly worse than a coin). The loud 10-12 NY block means bigger hourly moves of about 0.6%, not direction. Price is sitting about 0.5% under the day high, so a breakout bracket would mostly be a bet that the move continues, and nothing in the snapshot measures that. The direction inputs are noise, and the news forecaster is missing and fell back to 0.5. If price prints 87900, about 1% above the day high, before 18:00 UTC, then a long bracket just above 86983 would have been the right call and this hold was wrong.
Wrong if: a print at or above 87900 before 2026-10-02T18:00:00.000Z
I'm keeping the long I bought at 86809. Spot is 86580, so it is down $6.61; its stop near 85767 is $813 (0.94%) below spot and its target near 88893 is $2313 (2.67%) above. The loud New York block that just started (10:00 to 12:00, averaging 0.51 to 0.63% an hour) makes moves bigger both ways, not more likely to go up. The day's 2.87% range has already passed the 2.42% threshold, so the STORM size call is settled and says nothing about the next move. Direction is noise (AUC 0.52), so closing now would just pay the spread for nothing. Over the seven hours to 21:00Z the stop is roughly as likely to print as not, so my confidence is near even.
Wrong if: a print at or below 85767 before 2026-10-02T21:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Already holding. This arm carries one position at a time.
The STORM call has already come true: today's range is 2.87% against a 2.42% threshold (1.18x), and the price is sitting on the day high at 86983. Whatever the volatility forecast knew about today is now in the price, and it says nothing about how much further the day will stretch. The forecast's record doesn't support paying for more: STORM calls are 2 of 5 right, and overall Brier is 0.2514, slightly worse than a coin flip. The direction score of 5.7 is noise, so a long because the price is at the high has no edge either. A bracket into the 10:00-12:00 New York loud block would just be a breakout bet with no measured edge, so I'm holding. My read is wrong if the price runs another 1.5% to 88300 before 20:00 UTC, which would mean a long or an upside bracket would have paid.
Wrong if: a print at or above 88300 before 2026-10-02T20:00:00.000Z
I'm keeping the long I bought at 86809. Spot is 86983, a new Friday high, so the position is up about $5. Its 1.2% stop near 85767 is now $1216 (1.4%) below spot, and the 2.4% target near 88893 is $1910 (2.2%) above. The day's range of 84560 to 86983 is 2.87%, past the 2.42% threshold, so the STORM size call is settled and says nothing about which way the next move goes. Direction is noise (AUC 0.52), so this new high gives no reason to add or to close. The stop is now about three ordinary 08:00 New York hours (0.25% each) away and still a few loud-block hours (0.51 to 0.63% each) away, so I put the chance of it printing before 21:00Z at roughly a third. Closing now would only pay the spread to lock in $5.
Wrong if: a print at or below 85767 before 2026-10-02T21:00:00.000Z
the treatment did not act, so there is nothing to mirror
Wrong if: None of its own. This arm has no thesis, which is the point of it.
Already holding. This arm carries one position at a time.
The forecast goes out every day whether anyone reads it or not. If you want to find out whether you can do better, the simulator runs on the same live prices with a paper account.