The Metric That Moved 82 Percent and the One That Didn't

ClearTrace publishes a median execution slippage for each venue on its leaderboard. On 21 October the cell for ParaSwap on Ethereum read 6.40 basis points, over 17,913 transactions in the trailing seven days. On 29 July the same cell read 36.83. That is a fall of 82.6 percent, and measured in basis points it is the largest move over the quarter of any venue cell ClearTrace publishes.

Two things inflate that figure: the dates it is measured between, and a change in the transactions underneath it. A second ClearTrace instrument holds the comparison steady, pricing the same pairs at the same sizes against the same baseline at the same moment, and across the same quarter it shows ParaSwap's execution barely moving at all.

What the number is

The cell is a median. For every fill routed through the venue on Ethereum in a trailing seven-day window, ClearTrace values both legs at Dune's price for that minute, takes the size of the gap between them, and reports the median in basis points. The fills come from Dune's dex_aggregator.trades, where the venue is labelled velora, the name ParaSwap now trades under. The count published beside the median is distinct transactions, while the median itself runs over fills; on 21 October that was 17,913 transactions and 17,955 fills.

Three details of the query matter later. A transaction is credited to one project only, through a fixed priority order, so a change in how Dune attributes shared transactions moves this count. A fill is dropped unless Dune has a price for both tokens in that minute, so a change in price coverage moves it too. And the query itself changed during the quarter: its file was committed eight times between 10 July and 2 October, some of those to comments, with the measurement changes on 20 August, 9 September and 23 September. The 52 snapshots between 1 July and 21 October are readings from an instrument that changed while it was being read.

The window opened on the series high

36.83 is the highest reading in the whole series, and it is the reading the 90-day window opens on. 6.40 was the lowest reading up to that date, and it is the reading the window closes on. A change measured between those two dates is the full range of the series to that point, by the accident of where the calendar landed.

The opening value is also older than its date. It was first recorded on 24 July and appears again, identical, on 27 July and on 29 July, and so does every one of the other 39 venue cells. When a Dune run fails, ClearTrace's sync serves the last cached result instead, which is consistent with three identical snapshots across all 40 cells. The window opens on a measurement taken five days before the window began. The same series carries two more repeated pairs, on 13 and 15 July and on 26 and 29 August; the second is the late-August peak.

Between the endpoints the series does not trend. It rose to 29.43 in late August, fell to 16.62 by mid-September, climbed to 30.92 by the end of September and dropped 17 points the following week. That is ordinary in this data. Across 35 per-chain cells in the snapshot history, leaving out the ones ClearTrace has withdrawn or does not publish, the median absolute change over the quarter is 2.67 bps and the median range is 10.44. For the typical cell, the range inside the quarter is about three times the net move across it.

Then the population more than doubled

The last step down, from 8.89 on 14 October to 6.43 on 19 October, coincides with the count behind the reading going from 6,838 transactions to 16,940. Two days later it was 17,913. Both readings cover seven days, so the week ending 19 October held about two and a half times as many ParaSwap transactions on Ethereum as the week ending 14 October.

The jump belongs to that venue on that chain. The same venue on Base went from 4,961 transactions to 5,003 on the same sync. Of the 23 other venue cells on individual chains present on both snapshots, the largest rise was 1.37 times, and 16 of them fell. In Dune's public model repository, ParaSwap's own models were last changed in June, and I found no change to the shared aggregator models after 23 September.

The two ParaSwap entrypoints that the snapshot history tracks separately did not grow either: together they show 1,487 transactions on 14 October and 1,396 on 19 October, although those rows refresh less often than the venue cell. The additional ten thousand or so transactions came through other contracts. They could be new flow, or flow that the priority order or the price coverage had been assigning elsewhere or dropping. Telling those apart takes a query I have not run for this piece. Either way, a median over 6,838 transactions and a median over 16,940 are medians over different populations.

The closing reading has a second problem. On 21 October the signed median, which keeps the direction of each gap, was 0.67 bps, and 53.9 percent of scored fills were costs to the taker. The distribution was close to centred on zero. ClearTrace's own methodology note explains the consequence: when the gaps centre near zero, the absolute median reports how noisy the price feed is for those token pairs more than it reports execution. The window closes on a reading close to that regime.

The readings after the window show the surge passing. On 23 October the cell read 6.04 on 15,621 transactions, lower again. On 26 October it read 7.72 on 7,394, with a signed median of 3.14 and 66.4 percent of fills as costs, which is also what ClearTrace's public API showed on 27 October. The count is back near where it sat before 19 October, and the reading has moved up with it.

A steady count is not a steady population

My first reading of this was wrong. One of the two entrypoints, AugustusV6_2, fell from 62.31 bps to 4.31 over the same quarter while its count barely moved, from 871 transactions to 862. I took that as the part of the fall that was real: same population, lower cost.

It does not hold. That row was refreshed far less often, with 11 distinct readings across 38 snapshots, and its window also opens on its own high and closes on its own low. Its readings span 4.31 to 62.31 on counts that never left the range from 627 to 1,239. The other entrypoint, AugustusFeeVault, went from 67.72 to 63.69 over the same window.

The venue cell shows the same thing more plainly. Between 29 September and 5 October it fell from 30.92 to 13.81 while its count moved by two transactions, from 5,919 to 5,921. A count says how many transactions went into a median and nothing about which ones. The mix can change completely under an unchanged count, and a median follows the mix.

What survives depends on the baseline

Comparing periods instead of endpoints helps, but it does not settle the question. In the first half of October, before the count jumped, the five readings ran from 13.81 down to 8.89, with a median of 11.36. Against August and September, whose readings had medians of 19.44 and 20.78, that is a fall of about 45 percent. Against the first three weeks of July, median 14.5, it is a fall of about 22 percent.

Neither baseline is the obvious one, and the evidence behind each is thinner than the number of readings suggests. The readings are seven-day windows taken every two or three days, so the five October readings cover about two weeks of transactions. The counts are no help either: they held between 5,900 and 6,900 in early October but ran from 4,121 to 7,881 in August and September, and the previous section shows what a steady count is worth. On this cell alone, ParaSwap's October readings were lower than its late-summer ones, by an amount that depends on which period you call normal, for a reason the median cannot identify.

The measurement that holds the comparison still

ClearTrace runs a second instrument that avoids most of this. Its quote sampler asks each aggregator for a route on a set of token pairs at four sizes, from $1,000 to $1 million, and simulates the swap against chain state from a neutral address. It then compares the output with what Uniswap's default route delivers for the same pair and size in the same time window. Each observation carries its own baseline, and the pairs are chosen by ClearTrace rather than by whoever happened to trade that week.

On the seven pairs the sampler priced in every period of the quarter, ParaSwap's median output against the Uniswap default stayed between −0.76 and −1.64 bps at every size, in every period from early July to 26 October, over 8,425 matched windows. The first half of October sits inside the range of the months before it, and the days of the surge differ from the fortnight before them by no more than 0.15 bps at any size. Most of that steady gap is the flat 1 basis point fee ClearTrace's sampler has measured ParaSwap taking on major pairs.

The sampler measures something narrower than the on-chain cell. Its figure is ParaSwap's routing relative to one baseline route, on chosen pairs, with simulated swaps, and the pairs it tracks have shifted at the edges over the quarter, which is why only the seven present throughout are used here. What it shows is that whatever moved the on-chain cell did not show up in how ParaSwap routed these seven pairs at these four sizes.

How the 82 percent got chosen

I found this number by asking which metric had moved most over the last 90 days, ranking every per-chain cell by the change between its first and last reading in the window, and taking the largest among the venue cells ClearTrace publishes.

That ranking leans toward exactly this case. Of the 41 per-chain cells with at least ten snapshots in the window, three open on their own series high and close on their own series low, or the reverse. The ranking put those three first, fourth and seventh, and ParaSwap's cell is the fourth. Asking for the biggest mover between two dates is partly asking which series the calendar caught at its extremes. The unit matters as well: in percentage terms the largest published move went the other way, OKX on Base rising from 7.03 to 16.48 bps, on a count that fell from 5,032 transactions to 1,870.

Three of the six largest moves sit in cells ClearTrace has withdrawn or does not publish, and two of those were withdrawn after we found the number did not mean what the column said.

What would make the comparison mean something

• Compare periods, with their spread, and say which baseline. A month of readings has a median and a range. A change smaller than the range inside each month is weak evidence of anything, and a change whose size depends on the baseline needs every plausible baseline shown.
• Print the count beside every value. A reading whose count jumps far outside its usual range starts a new series, and it should be read as one rather than as the next point on the old line.
• Mark carried-forward readings and query versions. A snapshot that repeats the last value because a run failed is not a new observation, and a series that spans a query revision is two series.
• Pair each observation with its own baseline where the question allows it. The discipline for this is implementation shortfall: the gap between the price at the moment of decision and the price achieved. ClearTrace's sampler is built on that pairing: each row holds a solicited quote and the simulated fill for it, and the comparison above sets each fill beside the baseline route's fill in the same window. Comparing two routes removes the price feed from the comparison, and fixing the pairs removes the change in mix between periods. Neither is available on a median of whatever traded that week.

Why the direction of the move matters

A cost reading that rises 82 percent gives the venue a reason to ask why. One that falls 82 percent gives nobody outside the measurement a reason to look, so the check has to come from whoever publishes it.

ClearTrace has a precedent in the same direction. Its Scorecard Edition 3 records that, in the window then measured, the cell for the 1inch Limit Order Protocol published 13.27 bps where the population left after removing one resolver's maker-side edge measured 29.73. That venue read about twice as good as it was.

ParaSwap's cell was a published number, it held the largest move of the quarter, and on the instrument that holds the comparison still, the quarter barely registers.