Archive / Crypto & DeFi / E090
No detectable pre-CPI/FOMC crypto drift that survives costs
Bitcoin rose 47.3 bp in the 4 hours before 57 US CPI releases in 2022-2026, but that vanished out of sample: 12.8 bp before 52 releases in 2017-2021, below the 26.3 bp of ordinary weekdays.
A common belief says crypto prices drift up before big US data releases. In 2022-2026, Bitcoin gained 47.3 bp on average in the 4 hours before 57 CPI releases (t 4.8), one of 4 CPI direction cells that passed a multiple-testing check among 1,029 event cells. It did not hold out of sample: the same rule on 2017-2021 gave 12.8 bp over 52 releases, below the 26.3 bp of ordinary weekdays. The 1-hour CPI cell kept +16.0 bp gross but only +2.0 bp after the 14 bp cost. A Bitcoin pre-FOMC cell fell from 28.2 bp to 1.2 bp. None of 11 shorter re-tests beat 14 bp.
Crypto prices drift upward in the hours before major US macro releases such as CPI and FOMC, so buying ahead of the release earns a premium.
Why people believe it
People argue that holding a position through an event whose outcome is unknown carries risk that traders must be paid for, and that leveraged traders cut exposure beforehand and rebuy afterwards. The pre-FOMC cell was motivated by the drift documented in US equities before FOMC announcements (an academic study); this test applies that idea to crypto prices, not to equities.
Why it would persist The argument, as stated before testing
Common belief, not tested here: CPI comes out 12 times a year and FOMC decisions 8 times, so there are few repetitions for anyone to learn the pattern and compete it away.
We listed the release times of US inflation (CPI), jobs (NFP) and Federal Reserve (FOMC) announcements and measured what Bitcoin, Ether and a 24-coin basket did in the hours before each one. We compared that with the same clock time on ordinary days. Searching 1,029 combinations of event, window and asset flagged Bitcoin drifts before CPI; a Bitcoin and a basket cell before FOMC were weaker (q 0.21 and 0.06). We then wrote down the direction of each drift and re-tested it on years the search never saw (2017-2021), with no change to the rule.
No dated pre-registration. In-sample filter coded in the grid: Benjamini-Hochberg q below 0.05 over all cells, sign positive in the selection split, net above 0 after 14 bp in validation and holdout. Reading after the fact for the re-tests: one-sided p below 0.05, Bonferroni pass over the 23 re-tests, and a signed mean above 14 bp.
No dated pre-registration. Reading after the fact: a drift is rejected when its fixed-sign out-of-sample mean misses one-sided 5% significance or the Bonferroni bar, or does not exceed the 14 bp cost.
14 bp round trip (7 bp per side).
Same clock time on non-event weekdays (08:30 New York time for CPI and NFP, 14:00 for FOMC), with the weekday mix matched in the in-sample grid; 148 placebo days at 08:30 and 99 at 14:00 in the 2017-2021 test. NFP, PCE, PPI and FOMC minutes serve as in-sample comparison events.
The 4-hour pre-CPI cell was not significant out of sample (t 0.6) and below placebo. The 1-hour cell (52 releases) was +16.0 bp gross (one-sided p 0.04, placebo +2.6) but +2.0 bp net of 14 bp, -24.0 bp at a 40 bp stress cost, and failed the Bonferroni bar; the 11 shorter re-tests all missed the 14 bp bar.
- In 2022-2026, a Bitcoin long from 4 hours before each CPI release to the release earned +47.3 bp on average over 57 releases (median +42.1 bp, 77% positive, 95% interval +28.7 to +66.4 bp, t 4.8). A positive number means the price rose. The same hours on non-event days averaged +6.0 bp.
- The grid had 1,029 cells. 86 passed Bonferroni and all 86 were absolute-move cells; none of the 573 return or continuation cells passed Bonferroni. The pre-CPI Bitcoin cell had a whole-grid false-discovery q of 0.032.
- Inside the sample the drift shrank: 60.4 bp in the selection split, 38.3 bp in validation, 18.4 bp in the holdout. By year it was +84, +34, +61, +33 and +13 bp for 2022 to 2026.
- Out of sample (2017-09 to 2021-12) the same Bitcoin rule earned +12.8 bp over 52 releases (median -3.3 bp, 48% positive, 95% interval -29.1 to +53.7 bp, t 0.6). Ordinary weekdays at 08:30 New York time earned +26.3 bp over 148 days (interval +1.4 to +53.8 bp), so the CPI mean was 13.5 bp below the placebo (derived: 12.8 minus 26.3). After 14 bp the mean was -1.2 bp.
- The 1-hour version fell from +24.9 bp in sample to +16.0 bp out of sample (52 releases, interval -1.0 to +33.4 bp, t 1.8, one-sided p 0.04) against +2.6 bp on placebo days. Net of 14 bp it is +2.0 bp, and -24.0 bp net of a 40 bp stress cost; one-sided p 0.04 is above the Bonferroni bar of 0.05 divided by 23 re-tests. By year it was +28.0, +12.5, +5.3 and -4.8 bp for 2018 to 2021. On the shorter 2019-09 to 2021-12 re-test (28 releases) it was -1.5 bp.
- Before FOMC announcements, Bitcoin in the hour before the 14:00 release earned +28.2 bp in sample (38 meetings) and +1.2 bp out of sample (34 meetings, interval -30.4 to +29.0 bp). The 24-coin basket cell (+47.2 bp in sample, 38 meetings) returned -8.6 bp when re-tested on Bitcoin (18 meetings).
- All 11 macro-release re-tests (8 cells; basket cells counted on both Bitcoin and Ether) had a signed out-of-sample mean from -13.2 to +2.7 bp, a smallest one-sided p of 0.41, and a net of 14 bp from -11.3 to -27.2 bp. The cost is not what decided these: the gross mean was already at or below +2.7 bp.
- Before NFP, which is released at the same 08:30 time as CPI, the in-sample Bitcoin cells showed +8.0 bp at 4 hours and +7.2 bp at 1 hour (57 releases, whole-grid q 0.90 and 0.81).
The CPI drift is gone in 2017-2021
Gross mean Bitcoin return before CPI releases with 95% intervals (cost line shown separately). In 2022-2026 the 4-hour interval (+28.7 to +66.4 bp) is far above zero; in 2017-2021 it is +12.8 bp, interval -29.1 to +53.7 bp, and ordinary weekdays at the same hour average more (+26.3 bp). The 1-hour version is +16.0 bp gross, interval -1.0 to +33.4 bp; after the 14 bp cost line it is +2.0 bp net.
Chart data (bp)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| 4 h before CPI, 2022-2026 (57) | 47.3 | 28.7 | 66.4 |
| 4 h before CPI, 2017-2021 (52) | 12.8 | -29.1 | 53.7 |
| 4 h, ordinary weekdays 2017-2021 (148) | 26.3 | 1.4 | 53.8 |
| 1 h before CPI, 2022-2026 (57) | 24.9 | 15.7 | 33.8 |
| 1 h before CPI, 2017-2021 (52) | 16 | -1 | 33.4 |
| 1 h, ordinary weekdays 2017-2021 (148) | 2.6 | -11.3 | 16.6 |
Re-tested cells fall to about zero
In sample 2022-2026Out of sample 2019-2021
Mean return of each macro cell in sample (2022-2026) and in the out-of-sample re-test on Bitcoin (2019-09 to 2021-12), signed as in the grid. No out-of-sample bar reaches the 14 bp cost; two of four are below zero (-1.5 and -8.6 bp). Basket cells are re-tested on Bitcoin.
Chart data (bp)
| In sample 2022-2026 | Out of sample 2019-2021 | |
|---|---|---|
| CPI, Bitcoin, 4 h before | 47.3 | 1.1 |
| CPI, Bitcoin, 1 h before | 24.9 | -1.5 |
| CPI, basket, +15 min to +1 h, in direction of first 15 min | 26.4 | 2.7 |
| FOMC, basket, 1 h before | 47.2 | -8.6 |
Why it failed Interpretation · not established
The test cannot say whether the 2022-2026 drift was noise from picking one cell out of 1,029 or a feature of the years after crypto began to trade on US inflation data. Two facts fit either reading: the drift shrank within the sample (60.4, 38.3, then 18.4 bp), and the earlier years show a drift no bigger than ordinary weekdays at the same clock hour (+26.3 bp).
Lesson
A drift picked from a large grid can shrink inside its own sample (60.4 to 18.4 bp across splits) and still carry a t-statistic of 4.8. Comparing with the same clock time on ordinary days, and with years the search never saw, separates a release effect from an hour-of-day pattern or a regime.
Scope of this result
Covers Bitcoin, Ether and a 24-coin basket, long positions in the 4 hours and 1 hour before CPI (08:30 New York time) and FOMC (14:00) releases, 15-minute prices, 2017-09 to 2026-10, and a 14 bp round-trip cost. Out-of-sample prices are spot, in-sample prices are perpetual futures. With 18 to 57 events per cell, effects below roughly 30 to 60 bp could not be reliably detected (about 59 bp for the 4-hour out-of-sample CPI cell).
A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.
1. Test design
Pre-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 5Finding 6Finding 7Finding 83. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade