Scheduled fixing and auction flows leave no tradable trace
No rule built on the month-end London fix, gotobi days or Treasury auctions cleared its cost; sterling's +9.33 bp after the fix is 1 of 40 cells and below the 20 bp cost.
Some trades are mandated by the calendar: funds rebalance currency hedges at the London 4pm fix on the last day of the month, Japanese importers buy dollars on fixed days, and dealers absorb new Treasury supply at auctions. Each flow is public and mechanical, so its price effect is widely expected. We fixed a rule for each in advance. Over 81 month-ends the euro rule (trading into the fix in the direction of the month's S&P 500 return) returned -3.3 bp, the dollar-yen rule returned +0.2 bp over 41 gotobi days, and the 10-year yield changed +0.1 bp on 108 auction days (signed so that + is the predicted direction).
Large scheduled flows, at a published fixing time or on a published auction date, move prices in a predictable direction that a trader can collect.
Why people believe it
Three versions of the story are common. At the month-end London fix, foreign holders of hedged US equities must sell dollars forward when US stocks have risen, so the dollar should weaken into the fix and recover after it. On gotobi days (the 5th, 10th, 15th, 20th, 25th and last day of the month, moved earlier when it falls on a weekend or holiday), Japanese importers settle dollar purchases at the Tokyo 09:55 fix, so dollar-yen should rise into it and give the move back. At Treasury auctions, dealers who absorb new supply ask for a concession, so yields should rise into the auction and fall afterwards. The month-end mechanism is described in academic work (2015 and 2018); the test here uses crypto-venue prices rather than interbank quotes.
Why it would persist The argument, as stated before testing
The common belief is that the people on the other side trade at a set time whatever the price, and that an arbitrageur must hold the position through the fix or auction without knowing how large the flow is.
Moves are given in bp (basis points, 1 bp = 0.01%). For each flow we wrote down the window, the direction and the pass bar before computing any return. Month-end test: on the last weekday of each month, take the euro (and in two further pairs sterling and the Australian dollar) from 15:00 to 15:59 London, in the direction the month-to-date S&P 500 return implies, then test the reversal after the fix. Gotobi test: dollar-yen from 09:00 to 10:00 Tokyo on gotobi days, then the reversal. Auction test: the daily change in the 2-, 10- and 30-year Treasury yield from the day before an auction to the auction day, and from the auction day to the next day. Every result is compared with ordinary days at the same clock time.
Net mean above 0 at the realistic cost of the venue where the rule would trade; 95% date-clustered interval excluding 0; median above 0; positive in all three date windows; Benjamini-Hochberg q below 0.05 across all cells of the wider study (these three tests contribute 40, 22 and 12 cells); the placebo mean below half the effect or of opposite sign; same sign on at least one second instrument or venue.
Failing any one of these gates rejects the rule. The auction test was only an existence check, so it could not pass on its own.
Spot pairs: 10 bp per side (20 bp round trip, taker fee at the standard tier). Yen contract: 2.2 bp per side (4.4 bp round trip); 5-minute dollar-yen contract: 6.0 bp per side. Stress case: 3 times the per-side cost. The auction test is an existence check on daily yields with no tradable instrument, so no cost is applied.
Ordinary weekdays at the same clock times (1,671 for the euro spot pair, 98 for the yen contract, about 2,100 per yield series for auctions); the month-end rule at 13:00-13:59 London, an hour with no fix.
No rule cleared its cost: the main gross cells were -3.26 bp (month-end), +0.18 bp (gotobi) and +0.14 bp (auction day); sterling's +9.33 bp is 1 of 40 cells and -10.67 bp net of the 20 bp round trip; the 2-year auction-day move of -1.98 bp went against the prediction.
- Month-end, S&P-conditioned rule on the euro spot pair: -3.26 bp gross over 81 month-ends (interval -8.14 to +1.27, t -1.34), the wrong sign; net of the 20 bp round trip it is -23.26 bp. By date window: -7.22, -2.08 and +6.82 bp.
- The unconditional euro move from 15:00 to 15:59 London on month-ends averaged +1.94 bp (interval -3.08 to +6.40); on 1,671 ordinary days it averaged -0.42 bp, and at 13:00-13:59 with the S&P rule it averaged -1.17 bp over the same 81 month-ends.
- After the fix, fading the pre-fix move earned -2.49 bp from 16:05 to 17:00 London (interval -5.06 to -0.17, t -2.01; windows -2.92, -2.89, -0.89) and -2.88 bp to 20:00: the pre-fix move did not reverse. On 1,671 ordinary days the same rule earned -0.25 bp.
- Sterling, fading to 20:00 London, gave +9.33 bp over 41 month-ends (interval +1.34 to +17.61; windows +6.64, +26.18, +1.51 bp), the only spot-pair cell with an interval above zero, one of 40 cells tested; net of 20 bp it is -10.67 bp, below cost. The Australian dollar over 34 month-ends gave +12.20 bp with an interval from -4.50 to +33.60.
- Gotobi, long dollar-yen from 09:00 to 10:00 Tokyo: +0.18 bp over 41 days (interval -2.36 to +2.69, median +1.90); the same window on 98 ordinary days averaged -0.40 bp. The upper bound of 2.69 bp lies below the 4 to 5 bp size reported historically. Net of the 4.4 bp round trip: -4.22 bp.
- Gotobi post-fix short, 10:00 to 12:00 Tokyo: -0.36 bp (interval -3.95 to +3.31) against -0.03 bp on ordinary days; on 5-minute data over 8 gotobi days the 09:00-09:55 window averaged -2.60 bp (interval -6.48 to +1.28).
- Auction findings are signed so that + is the predicted direction (up into the auction; on the day after, + means the yield fell). 10-year auctions (108): +0.14 bp on auction day (interval -0.79 to +1.01) and +0.70 bp the day after (interval -0.44 to +1.97); on 2,107 ordinary days the first window averaged +0.13 bp. 30-year auctions (106): +0.33 bp, then -0.32 bp (the yield rose 0.32 bp the day after); both intervals include 0.
- 2-year auctions (103): the yield fell 1.98 bp on auction day (interval -3.16 to -0.80, t -3.35), opposite to the concession prediction and distinguishable from zero; the day after it rose 0.20 bp (signed -0.20). By date window the auction-day move was +0.03, -5.24 and -4.57 bp.
The month-end fix: wrong sign, or nothing
Average gross move per month-end in bp with date-clustered 95% intervals; counts in brackets. The rules predicted gains. The S&P-conditioned rule averaged -3.26 and the post-fix reversal -2.49. Sterling's +9.33 is the only spot-pair interval above zero and is under half the 20 bp round-trip cost.
Chart data (bp per month-end)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Euro, S&P-conditioned, 15:00-15:59 (81) | -3.26 | -8.14 | 1.27 |
| Euro, unconditional, 15:00-15:59 (81) | 1.94 | -3.08 | 6.4 |
| Euro, after fix, 16:05-17:00 (81) | -2.49 | -5.06 | -0.17 |
| Euro, after fix, 16:05-20:00 (81) | -2.88 | -7.52 | 1.93 |
| Sterling, after fix, 16:05-20:00 (41) | 9.33 | 1.34 | 17.61 |
| Control: S&P rule at 13:00-13:59, no fix (81) | -1.17 | -3.73 | 1.46 |
| Control: S&P rule on ordinary days (1,671) | 0.43 | -0.37 | 1.19 |
Gotobi days look like any other day
Average gross dollar-yen move in bp with date-clustered 95% intervals; counts in brackets. The long window averaged +0.18 on 41 gotobi days and -0.40 on 98 ordinary days. The line at 4.5 marks the middle of the 4 to 5 bp effect reported historically; every gotobi interval lies below it.
Chart data (bp per day)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Long 09:00-10:00 Tokyo, gotobi days (41) | 0.18 | -2.36 | 2.69 |
| Same window, ordinary days (98) | -0.4 | -3.59 | 1.84 |
| Short 10:00-12:00 Tokyo, gotobi days (41) | -0.36 | -3.95 | 3.31 |
| Same window, ordinary days (98) | -0.03 | -2.13 | 2.07 |
| 5-minute data, long 09:00-09:55, gotobi days (8) | -2.6 | -6.48 | 1.28 |
Treasury auctions: no concession, and one reversal
Daily yield change in bp, signed so that + is the predicted direction (up into the auction; on the day after, + means the yield fell); counts in brackets. Both 10-year and 30-year windows include 0. The 2-year auction-day move of -1.98 is distinguishable from zero but opposite to the prediction.
Chart data (bp of yield, predicted direction)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| 10-year, auction day (108) | 0.14 | -0.79 | 1.01 |
| 10-year, day after (108) | 0.7 | -0.44 | 1.97 |
| 30-year, auction day (106) | 0.33 | -0.93 | 1.5 |
| 30-year, day after (106) | -0.32 | -1.24 | 0.64 |
| 2-year, auction day (103) | -1.98 | -3.16 | -0.8 |
| 2-year, day after (103) | -0.2 | -0.97 | 0.56 |
| Control: 10-year, ordinary days (2,107) | 0.13 | -0.11 | 0.36 |
Why it failed Interpretation · not established
Not established by the tests. The flows may be real but small next to the price noise of an hour or a day, which is the resolution used here. Fix orders execute against interbank quotes, while the month-end and gotobi tests ran on crypto-venue prices for the same currencies, so the pre-2015 fix pattern would not necessarily show up in them. The 2-year auction-day move of -1.98 bp fell in the last week of the month, a timing that we suspected, after seeing the result, reflects month-end index buying; that was not tested.
Lesson
A flow that is large, mechanical and public does not guarantee a price trace at the clock and in the instrument a researcher can observe. A pre-registered direction adds a second way to fail: the S&P-conditioned month-end rule (-3.26 bp) and the 2-year auction day (-1.98 bp) came out against the story. Matched ordinary-day placebos and a no-fix hour are what separate a calendar effect from a drift present on every day.
Scope of this result
Covers the euro, sterling and Australian dollar against a dollar stablecoin on a crypto spot venue, a yen perpetual contract with at most 8 months of history, and daily 2-, 10- and 30-year Treasury yields from 2018 to 2026. It does not cover interbank quotes, intraday Treasury or futures data, other fixing times, other currencies, years before these samples, or the gotobi pattern before 2026.
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