This line never cost a dollar, because before building anything I read the wallets already doing it. Result: 47 same-shape wallets, median −0.74¢/share. Best return-on-effort kill on this site.
The conditions look good: posting costs no fee, fills pay a rebate (about +0.25¢/share), boards are frequent, opportunities dense. "Quote both sides, earn the spread plus the rebate" is a natural idea.
| Question | Measured |
|---|---|
| Population median | −0.74¢/share (47 same-shape wallets) |
| Best performer | about $95/day |
| Do rebates rescue it? | No. Wallets holding $100–380 of rebates stayed net negative |
| Are the top makers here? | No. None of their recent thousand-odd fills were in this category |
The last row carries the most information: the people best at this business avoid this category. Their money is in sports, where the top of the table earns $545–3,739/day.
| Holding window | Pair rate | Naked exposure |
|---|---|---|
| 5 minutes | 47.0% | 53.0% |
| 15 minutes | 70.9% | 29.1% |
| 60 minutes | 95.4% | 4.6% |
A naked share costs −13.5¢ against a +0.25¢ rebate — 54×:
| Tenor | Naked share | Adverse selection | Rebate | Net |
|---|---|---|---|---|
| 1 hour | 3.5% | −0.475¢ | +0.248¢ | +0.099¢ |
| 5 minutes | 53% | ≈ −7.2¢ | +0.25¢ | deeply negative |
Those figures came from truncated replay. Replaying against recorded real books plus complete tape:
| Tenor | Truncated replay | Full simulation | Net ¢/share |
|---|---|---|---|
| 5 minutes | 47.0% | 36.3% | −3.12 |
| 15 minutes | 70.9% | 33.6% | −9.79 |
| 1 hour | 95.4% | 54.9% | +4.58 * |
70.9% looked high enough to try. Full simulation returned 33.6% and a net of −9.79¢/share — worse than five minutes. Killed outright.
The failure was not inaccuracy but not deciding the bias direction in advance: simulating short tenors from long-tenor flow necessarily overstates pairing. Now a house rule — replays must state their bias.
One board of six contributed +$667; the other five had a median of −$19. Small sample plus fat tail makes the mean meaningless. Static laddering lost money on the typical board at every tenor tested — driven by settlement luck, not spread capture.
That range is this line's most valuable output — not a point estimate but a range with an attributable end on each side: the lower from a conservative-queue full simulation, the upper from a skilled operator's measured result.
Which makes the question concrete: is there evidence I would land at the good end? There was not, so the line was never run.
| Rule | Note |
|---|---|
| Read the population median first | the top performer is not your expectation; the median is |
| Label an approximation's bias direction | fifteen minutes nearly got through on an unlabelled one |
| Report a range, then ask why you deserve the good end | a point estimate cannot pose the question |
This kill cost nothing because the people already doing it had run the experiment for me — their results are public on-chain, and the median is negative. Reading the population median before you start is the highest-return hour on this venue.