Resting deep

1The market and its mechanics

Each market settles one question: whether the price of a coin is higher at the end of a short window than at the start. Two tokens trade, Up and Down. The winning token pays $1 at resolution and the losing one pays nothing, so a matched pair of Up and Down is always worth exactly $1. Prices run from 1¢ to 99¢ in ticks of 1¢. A five-minute series lists 288 markets a day and a fifteen-minute series lists 96. All figures in this study are in USDC, which we treat as US dollars.

Trading runs on a central limit order book. Orders match off-chain and settle on-chain about 2.3 seconds later.

Settlement. Outcomes are decided by an oracle's BTC/USD (or ETH, SOL, XRP) price. From 14 August 2026, five-minute markets settled on a 60-second time-weighted average at the start and at the end of the window. Before that they used a 30-second average. Rebuilding the rule from the oracle's own prints reproduced 2,263 of 2,265 outcomes (99.91%) in the 60-second era, and 93.97% before 14 August, when the 30-second rule applied and the 60-second rule we tested was not expected to match.

Taker delay. The venue holds incoming taker orders briefly before matching them. The delay was 250 ms until 16 August, 50 ms from 17 August, and 150 ms throughout September and October. A resting order can be cancelled during that interval, which matters for anyone quoting while the spot market moves.

Fees. Takers pay a fee of 0.07 × p × (1 − p) per share, where p is the price. That is 1.75¢ at a price of 50¢ and less near the extremes. Makers pay nothing and receive a rebate, which our backtests estimate (section 6).

Size. In a sampled week (markets ending 12 to 18 October 2026, 40 markets drawn per category), BTC five-minute markets carried an estimated $47.3 million of taker notional, more than any other category we sampled, sports included. The median market traded $21,747. BTC fifteen-minute markets carried $11.1 million and ETH five-minute markets $3.7 million, with a median of $1,431 per market. SOL and XRP were thinner again.

Price discovery. The venue follows the spot market with a lag of about one second. The correlation between the one-second change in the venue's mid price and the spot return over the same second was 0.058. Against the spot return one second earlier it was 0.317, and by three seconds it was gone. The settlement oracle trails spot by one to two seconds, and sat a median of 9.4 basis points below the reference exchange's price. That gap matters because markets settle on the oracle, so any fair value built from exchange prices has to be translated into oracle terms.

Over a full window the mid price is well calibrated. Grouped into deciles, a mid of 0.451 resolved Up 43.6% of the time and a mid of 0.851 resolved Up 86.3% of the time (2,225 markets, 14 to 21 August). We tested whether a fast spot-price model knew anything the market did not. Adding it to the market price improved out-of-sample log loss by 0.08% in one week and made it worse by 0.55% in the other. The quoted price is already a good forecast. What remains to be earned is in how orders meet the book.

2Data and method

Order books. We bought historical order-book data from a commercial provider: full book snapshots plus every price change, with millisecond timestamps, and the venue's on-chain fills. The core sample is BTC five-minute markets from 8 to 21 August 2026, 4,032 markets. We added four September days (8, 14, 19 and 25 September, 1,152 markets, of which 1,117 had a live feed and 1,098 a recoverable outcome; the sweep measurement used 1,088) and 26 consecutive days from 26 September to 21 October 2026 for all eight series, 38,937 markets with usable data (for BTC five-minute, 7,112 of a possible 7,488). From 21 October we also ran our own recorder on the venue's market feed.

Reference prices. Spot prices came from the public trade feed of the most liquid BTC, ETH, SOL and XRP spot exchange, at millisecond resolution. Oracle prices came from the oracle's published reports. For the later periods, where oracle prints were not in hand, we rebuilt the oracle from spot and checked that the substitution did not change results (section 7).

Validation. Before any study used the data, we checked it.

• Rebuilding the book from snapshots and price changes for 120 random markets reproduced the reported best price 99.61% of the time across 12.9 million updates.
• Event timestamps never stepped backwards in 13.1 million rows. Median receive delay was 4.9 ms and the 99th percentile 155 ms.
• One wallet's fills in the vendor data matched our own independent capture in 60,290 of 60,293 transaction groups, with every amount equal.
• On 13 and 19 August the vendor's feed froze for long stretches while markets kept resolving. We flagged a market as frozen if it went silent for 10 seconds or more inside its window, or carried fewer than 5,000 updates. That removed 58 of 4,032 markets and left 3,974 live ones. The August studies used between 3,930 and 3,933 of them.
• Our own recorder's parsed book matched the vendor's best price at the same millisecond in every comparison on four of five test markets. The fifth had only five comparable milliseconds.

Backtests. Strategies were replayed event by event through the recorded books. Our orders arrived after a modelled delay to see the book and a further delay to reach the venue. A resting order counted as filled only when a recorded trade consumed its whole price level ahead of it. We call this the sweep fill model. An earlier, simpler rule filled an order whenever a trade printed through its price; we replaced it with the sweep model, which follows the book level by level. We also ran a variant that gives our order a proportional share of the queue. Positions were held to resolution. Profit is reported gross, then net of fees and the estimated rebate. Standard errors are clustered by market, and "t" is net profit divided by its standard error.

Regimes in August. The August sample spans two rule changes: the settlement average moved from 30 to 60 seconds on 14 August, and the taker delay from 250 ms to 50 ms on 17 August. Strategy results are reported by week (week 1: 8 to 13 August, 30-second settlement; week 2: 14 to 21 August, 60-second settlement). The census and the depth measurement in section 4 pool the fortnight.

Completeness check. In the census, the net result of all 37,312 wallets, before rebates, was −⁠$2,079,296, and the taker fees charged were $2,079,296. Trading between wallets is zero-sum, so the two must match. They match to the dollar, which confirms that every fill and fee in the period was captured.

Discipline. Each test's rules were written down before its data was read, and results were logged whether they helped or not. Every reported number came from committed code and hashed raw data. Where a design was chosen after seeing a period, we treat that period as in-sample and say so.

Fill rule

A resting order counts as filled only when a recorded trade consumes its whole price level, with our order at the back of the queue.

Edge

The outcome (1 or 0) minus the price paid, in cents per share, held to resolution.

Standard errors

Clustered by market. "t" is net profit divided by its standard error.

Rebate

The maker rebate is estimated and included in net figures. The main out-of-sample result stays positive without it.

Latency model

Our orders see the book after a set delay and reach the venue after a second one. Measured values were 8 ms and 12 ms.

Data checks

Books rebuilt and compared with the vendor's reported best price, 58 frozen markets removed, settlement outcomes reproduced from oracle prints.

3Where liquidity sits and who provides it

The census. We attributed every on-chain fill in BTC five-minute markets from 8 to 21 August 2026 to its wallet: 21,084,556 fills across 37,312 wallets. We called 2,539 wallets active (at least $5,000 of volume, five days and 200 fills).

In aggregate the wallets lost exactly the taker fees they paid, as section 2 notes. How that loss was shared is the result.

Of the active wallets, 727 were profitable before rebates and 985 after an estimated rebate. Only 67 had a t-statistic above 3 across markets. The ten largest winners took a third of all positive profit.

Source: census of on-chain fills. Style is assigned from each wallet's share of maker fills. Taker fees deducted, rebates excluded.

Taking liquidity lost money in aggregate. Takers who did well in the first week did not keep doing well. Of 293 takers with at least $5,000 of taker volume in each week who were profitable in week one, 59.4% lost money in week two, and as a group they finished week two at −⁠0.08% of volume. Takers who lost more than 3% in week one kept losing (−⁠3.06% in week two, 80.2% of them down). Losing persisted. Winning did not.

The best makers. Three wallets stood out. Each made between $45,600 and $52,400 in the fortnight on $747,000 to $853,000 of volume, about 6% of what they traded, with 77% to 86% of it as maker. They filled on the same taker order in 5.9% of their transactions, which strongly suggests one operator splitting its quotes. We studied them as one group.

We measured each maker fill's depth: how many ticks behind the best price the order stood when the taker reached it. For the leading group, 38.5% of volume filled one tick or more behind the best price and 42.3% at it. For a 10% sample of all other makers the split was 12.3% and 55.6%. The rest of each group's volume (19.2% and 32.2%) filled at levels the book feed had not yet shown, so its depth cannot be assigned. The leading group earned 2.37¢ a share at the best price, against 0.20¢ for other makers. Deeper fills earned more again, and the larger the order that reached them, the more they earned.

The group also made money from whom it traded with. It earned 5.11¢ a share against takers who lost more than $5,000 over the fortnight, where other makers earned 0.94¢, and lost 3.26¢ against those who won more than $5,000, about the same as other makers. Its deep orders caught the impatient. When the spot price had moved more than 2 basis points against the maker in the second before a fill, other makers lost 1.93¢ a share. The leading group lost 0.85¢, and only 1.3% of its volume filled in that state.

The edge shrank after August. In four September days, under the 150 ms taker delay, the three wallets made 2.7% of their volume, down from 6.1%. A larger share of their fills sat at the best price (61.2%) and earned 0.52¢. The deep fills still paid: 15.1¢ a share at three to five ticks and 23.1¢ at six to ten. They were simply rarer.

4Sweeps

A sweep is a taker order large enough to consume the best price level and keep going. The last shares of a sweep are bought or sold at prices the rest of the market would not have offered a moment later. A resting order deep in the book is filled only by such orders.

We measured this directly. For every recorded trade, we asked what a small resting order k ticks behind the best price would have earned if the trade consumed its entire level. Edge is the outcome minus the price paid, held to resolution, with no rebate.

Three things follow. The edge rises steadily with depth and has the same shape in both periods. Fills fall away just as fast, from 55 per market at one tick to under 4 at eight, so depth trades frequency for edge. And one tick back there is nothing to earn at all.

Timing matters as much as depth. Deep fills paid most in the first minute of a window and in the last, and least in the middle. At eight ticks in September, fills in the first 60 seconds earned 18.3¢ a share and fills between 60 and 240 seconds earned 0.55¢. The pattern held at three, five and eight ticks in both periods.

5Undertow: the mechanism and its backtest

How it works

Undertow is a maker strategy built on the two findings above. It has four parts.

There is no fair-value filter. The strategy does not try to forecast the outcome. It earns only when a taker overshoots, pays no fee, and steps aside when the book or the spot market says the next fill is likely to be a bad one. The 150 ms taker delay is what makes the lookout possible: when spot moves, there is a short interval in which a resting order can be withdrawn before a fast taker's order is matched.

How we got there, including what failed

Undertow is the third design. Both earlier designs failed, the second in a pre-registered test.

1. Design 1: a ladder priced off fair value. Bids at fixed distances below a model fair value lost money in the first August week (−⁠$4,825 at 10 shares, t −⁠2.31) and roughly broke even in the second. Simply joining the best price did worse: −⁠1.70¢ a share in week one, t −⁠5.88.
2. Design 2: a ladder anchored behind the best bid. Resting orders at fixed depths behind the best bid, filtered by a fair-value edge, worked in August. Under the stricter sweep fill model, the 3/6/10-tick version made 1.17¢ and 0.78¢ a share in the two weeks (t 3.44 and 2.68).
3. Design 2 fails its pre-registered September test. We fixed four versions of the anchored ladder in advance and ran them on the four September days. None worked: results ran from −⁠0.40¢ to +0.29¢ a share, and no t-statistic exceeded 0.41. A pre-open maker variant lost 1.19¢ a share.
4. Design 3, Undertow, then a fresh test. Exploratory runs after September pointed to deeper levels (5/8/12), no edge filter, and two guards. Adding the tripwire raised profit per share from about 2.5¢ to 4.5¢ in both August weeks, and almost fivefold in the September days, from 1.45¢ (t 1.84) to 6.85¢ (t 4.39), at 10 shares. Because this design was chosen after seeing September, the August and September results for it are in-sample. We locked it and tested it on 26 days that no design decision had touched.

The out-of-sample test

The locked design ran on every BTC five-minute market from 26 September to 21 October 2026, using rules written down in advance. At the 5 ms delays we had assumed, it passed: +$14,904 at 30 shares, t 6.93, positive on 24 of 26 days. Gross profit before the rebate was $13,868.

We then measured the delays from a server close to the venue: about 8 ms for book updates to arrive and 12 ms for an order to reach the venue. (An earlier measurement had found 16 ms and 18 ms; the second one replaced it.) At those delays the same 26 days returned $8,855 at 30 shares, t 3.70, with 17 of 26 days positive. The measured-latency figures are the ones we rely on below.

By series

We then ran version 1 on all eight series, with BTC's parameters unchanged, at the measured delays.

Every series was positive. BTC's two series produced 57% of the combined total at 30 shares and 69% at 100 shares. SOL fifteen-minute markets, at t 1.82, were the weakest and would not pass a conventional significance test on their own. Running all eight at once meant about 41 order messages a second across the live markets.

By size

Larger orders earned more per share in the backtest without adding many fills.

The pattern is consistent with section 4: a large resting order takes more of the overshoot from the biggest sweeps. It is also the least reliable part of the backtest. The model assumes that takers would have sent the same orders with our extra size in the book. At 1,000 shares a level we would have been visible, and we have no evidence of how the market would respond. We treat everything above 100 shares as an upper bound.

6Risks and failure modes

Adverse selection. A resting order is filled when someone wants to trade with it, and the person most eager to trade is often the one who knows the price is about to move. Undertow's guards exist for this. Without either guard, the 3/6/10-tick ladder lost 1.81¢ a share in September, t −⁠5.85. Without the tripwire, the 5/8/12 ladder's September edge fell from 6.85¢ to 1.45¢. The mid-window problem remains even with both guards: in the out-of-sample test at 30 shares and 5 ms, fills between 60 and 240 seconds earned 3.12¢ a share against 5.52¢ in the first minute.

Fast markets. The worst fills come just after the spot price jumps and before the venue catches up. Among other makers, fills after a move of more than 2 basis points against them in the preceding second lost 1.93¢ a share. The lookout pulls quotes in exactly this state. It can only do so if our view of spot is fresh and our cancel arrives within the taker delay. Our spot feed arrived about 115 ms after the reference exchange's trade time (median of 2,536 measurements, 99th percentile 122 ms), which uses most of a 150 ms delay.

Latency. Speed was the largest single risk. On the 26 BTC days at 30 shares, net profit fell as both delays rose together:

Our measured setup, 8 ms to see the book and 12 ms to reach the venue, returned $8,855 (t 3.70), close to where the curve would put it.

The curve has a standard error of about $2,100 at each point (net divided by t), so neighbouring points are not separable on their own. That is why the 5 ms result in section 5 ($14,904, with the same 120 ms spot delay) sits nearer the 3 ms point than the 6 ms one. The slope is clear over several milliseconds and should not be read one millisecond at a time.

The two legs mattered about equally: 5 ms to see the book and 11 ms to send returned $12,144, and the reverse returned $11,640. Slower setups also filled more often, from 16,754 fills at 3 ms to 62,776 at 21 ms. The extra fills came from orders that a faster setup would have cancelled in time.

Distance from the venue mattered far more than distance from the spot market. At planning estimates for different sites, a 25 ms site lost $997 over the 26 days. A site 1 ms from the reference exchange and 110 ms from the venue lost $23,874, t −⁠5.79, with only 2 positive days of 26. These site figures rest on planning estimates of latency. None of them was measured.

We also measured slower tails on the quieter coins' book feeds: a 99th percentile of 40 to 108 ms, against about 30 ms for BTC. We did not find the cause.

Fees and rebates. Undertow pays no taker fee, so fees are a risk only through the rebate, which we estimated and did not observe. In the 5 ms out-of-sample run the rebate added $1,036 to a gross profit of $13,868, about 7% of net. If the rebate were zero, the strategy would still have been positive. Every taker strategy we tested, by contrast, was decided by the fee (section 8).

Capacity. Capacity is set by the flow of sweeps. At 100 shares a level, Undertow would have taken between 0.34% and 1.0% of the shares traded in each series, and combined backtest profit was $2,714 a day across all eight. Capital is tied up until each market resolves: at 100 shares the most held in a single BTC market was $2,027. Splitting the same total size across more levels did not help. Three levels of 30 shares made $35,552 over 18 design days, and twelve levels of 7.5 shares made $31,173 while sending four times the messages (23.9 a second per market against 6.1).

Regime change. The market's rules changed three times in our sample. The taker delay went from 250 ms to 50 ms to 150 ms, and the settlement average from 30 to 60 seconds. Edges moved with them. A simple latency rule that bought when spot jumped and the venue lagged earned 2.58¢ a share in the first August week and nothing in the second (0.11¢ ± 0.96). The leading makers' margin fell from 6.1% to 2.7% of volume between August and September. A future rule change could remove the taker delay that the lookout depends on.

7Sensitivity checks

Fill model. Giving our orders a proportional place in the queue, in place of the default assumption, left results almost unchanged. On the design days the 30-share ladder made 4.59¢, 5.18¢ and 7.37¢ a share in the three periods against 4.60¢, 5.14¢ and 7.57¢ under the default.

Latency on the design days. On the August and September design days, slowing one leg at a time to 20 ms, with the other left at 5 ms, kept results strong. Slowing one leg to 50 ms weakened them (September: 1.53¢ a share, t 1.86, with the book 50 ms late). The out-of-sample curve in section 6 slows both legs together and is steeper. We rely on it.

In-sample: the design was chosen after these periods were seen. Net edge per share, with t.

Rebuilt settlement prices. For the later periods we rebuilt the oracle price from spot. On August data, where real oracle prints were available, the rebuilt price correlated with the real one at 0.96 to 0.99 across window segments. Strategy results barely moved: one test version made 3.62¢ a share with real prints and 3.81¢ with rebuilt ones.

Tuning. We tried 15 alternative depths and tripwire settings on the design days, at a then-current latency estimate of 16 ms and 18 ms. None clearly beat version 1. Deeper nets earned less in total: the 12/18/25-tick versions made $19,160 to $19,534 on the design days, against $34,547 for version 1's settings at the same latency. The strongest alternative (6/10/15 ticks, 2-tick tripwire) was then run on the 26 out-of-sample days. It made $737 less than version 1, inside two standard errors (2 SE = $1,902). Version 1 stayed.

Days as the unit. At 5 ms per leg, the out-of-sample run was positive on 24 of 26 days (sign test p < 0.0001). Daily net averaged $573 with a day-level t of 4.5. The worst day lost $210, which was also the largest drawdown. At measured latency, 17 of 26 days were positive (sign test p = 0.08), so the day-level evidence at that speed is suggestive, and the market-level t of 3.70 carries the result. Daily figures at measured latency were not recorded.

From design days to fresh days. On the 18 in-sample design days, at 5 ms and 30 shares, every day was positive, with a mean of $1,936 a day. On the 26 fresh days at the same speed the mean was $573. Most of the fall was already visible inside the design period: August days averaged $2,225 and the four September days $925. Deep fills thinned over the same months. Fills per market at eight ticks fell from 3.5 in August to 1.4 in September, and the leading makers' margin fell from 6.1% to 2.7% of volume. In-sample selection also inflates the design-day figure. The notes cannot separate the two effects.

Markets. On the design days, 37.6% of individual markets lost money and the worst single market lost $357. The edge is earned across thousands of markets.

8What did not work

Undertow is the one design that survived. The table lists the others we tested on the same recorded data. Several of them were ideas we expected to work.

The pair-arbitrage idea was the project's starting thesis, so its failure deserves a sentence. Buying one side cheaply and waiting to buy the other is an unhedged bet until the second leg fills. In a simulation where prices were fair by construction, the rule locked a pair in most markets and still earned nothing before costs. On real paths it lost, because the side that has just fallen tends to keep falling over the next few seconds: after a 15-second fall, Up won 2.2 points less often than its calibrated price implied.

The taker strategies share a pattern. Several had positive gross edge, and the fee took it away. The machine-learned predictor's week-two taker version with a 0.5¢ threshold made $5,454 gross and paid $12,764 in fees.

9Limitations

1. Replay of recorded data. Every result here is a replay of recorded market data. Real fills may differ from the fill model.
2. Short out-of-sample record. The only data untouched by design choices is 26 days, from 26 September to 21 October 2026. The other seven series used BTC's parameters unchanged over those same days: new markets, but not a new period.
3. Fill model. The sweep fill model puts our order behind everyone already resting at its level in the recorded book, including the leading makers, so a sweep must clear them first. It assumes takers behave the same with our orders in the book. Results above 100 shares a level are optimistic for that reason.
4. One data provider. The order books come from a single commercial source with known gaps. We removed 58 frozen markets, and the provider disclosed missing fills for 32 markets. Our own recorder agreed with it where we could compare, but only from 21 October, so the out-of-sample days rest on the provider's books alone.
5. Correlated days and series. t-statistics are clustered by market. Days are not independent, and the eight series overlap in time and move together. We did not compute a t-statistic for the eight-series total, so it should not be read as eight independent results.
6. Per-cell uncertainty. Standard errors were recorded by depth but not by time in the window. The 8-tick September edge has 2 SE of 3.42¢ overall, and each time cell in Figure 4 is wider.
7. One latency measurement. The 8 ms and 12 ms delays were measured from one server, and the site comparison uses planning estimates.

10Status and further work

The study is complete and covers recorded data to 21 October 2026. On recorded order books from August to October 2026, Undertow was profitable in all eight series at measured latency: about $1,135 a day at 30 shares per level and $2,714 at 100, with the result set by reaction time.

Three extensions would add the most.

1. Fresh out-of-sample days. The study covers recorded data to 21 October 2026. An all-series test on later days, with rules fixed in advance, has not been run.
2. Real fills against the model. Comparing real fills with the sweep fill model over the same hours would replace the main modelling assumption with a measurement.
3. Latency at or below the measured setup. The curve rewards every millisecond: 8 ms to see the book and 12 ms to reach the venue returned $8,855, and 6 ms on each leg returned $13,399.