Crypto up/down comes in five-minute, fifteen-minute, hourly and four-hourly forms. They look like sizes of one thing. They are not: cost structure, viability, and even data trustworthiness differ.
| Item | Detail |
|---|---|
| The question | does an asset close above or below its open, within a fixed window |
| Resolution | an automatic price oracle — no submission, no dispute window |
| Cadence | a five-minute board runs 288 windows a day, continuously |
| The appeal | sample accumulates very fast — days rather than weeks |
That last row is the real value to a researcher: it is the fastest laboratory on the venue.
Paired inventory locks in; the unpaired part carries all the risk. And pairing is set almost entirely by tenor:
| Holding window | Pair rate | Naked exposure |
|---|---|---|
| 5 minutes | 47.0% | 53.0% |
| 15 minutes | 70.9% | 29.1% |
| 30 minutes | 87.4% | 12.6% |
| 60 minutes | 95.4% | 4.6% |
A naked share costs about −13.5¢ against a +0.25¢ rebate — fifty-four times. Multiplying through:
| Tenor | Adverse selection /share | Rebate /share | Net |
|---|---|---|---|
| 1 hour | −0.475¢ | +0.248¢ | +0.099¢, barely positive |
| 5 minutes | ≈ −7.2¢ | +0.25¢ | an order of magnitude short |
5m and 15m ruled out; 1h barely avoids losing, with almost no tolerance. Full ledger in the up/down market-making kill file.
| Tenor | What the rewards endpoint returns |
|---|---|
| 1 hour / 15 minutes | empty |
| 5 minutes | a placeholder config (daily rate 0.001) |
So this lane's only subsidy is the fill-triggered maker rebate. "Up/down market making has LP rewards" conflates two of the four rebates.
| Percentile | Top-of-book size |
|---|---|
| median | about $94 |
| 25th | $37 |
| 10th | $13 |
Single-order capacity follows: about $200 counting book-walking alone, about $90 once price drift during execution is included. Depth also swings hard by hour and by position in the window.
Research typically judges on the exchange spot close against the open; the market settles on the oracle's final-60-second time-weighted average against the open. Audited across four tenors — whole-sample unconditional mislabel rates:
| Tenor | Mislabel rate | Directional? | Verdict |
|---|---|---|---|
| 5 minutes | 10.1% | yes | findings void |
| 15 minutes | 3.1% | symmetric | unaffected |
| 1 hour | 0% | — | unaffected |
| 4 hours | 0.4% | — | unaffected |
Plain arithmetic: the final 60 seconds is 20% of a five-minute window and 0.4% of a four-hour one.
⚠ 10.1% is the whole-sample unconditional rate, not the rate inside the knife-edge subset. Full account in the broken ruler.
| Tenor | Market making | Label trustworthiness | Sample speed |
|---|---|---|---|
| 5 minutes | excluded | needs a settlement companion read | fastest |
| 15 minutes | excluded | ±1pp of symmetric label noise | fast |
| 1 hour | barely breaks even, minimal tolerance | clean | slow |
| 4 hours | — | clean | slowest |
Note the awkward conclusion: the tenor that accumulates sample fastest is the one whose data is least trustworthy. Not a coincidence — shorter windows sit closer to the boundary, and the boundary is where the rulers part.
Here tenor is not a parameter, it is a category boundary. Change it and pair rate moves from 47% to 95%, market making from deeply negative to barely positive, label error from 10.1% to zero — no conclusion transfers across tenors.